Search AI: a Machine Learning algorithm for chronic kidney disease risk detection using eight readily available clinical features

Search AI: a Machine Learning algorithm for chronic kidney disease risk detection using eight readily available clinical features | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Search AI: a Machine Learning algorithm for chronic kidney disease risk detection using eight readily available clinical features Julian Martinez, Natalia Castano-Villegas, Alejandra Perez, Daniel Jimenez, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7888843/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 6 You are reading this latest preprint version Abstract Background: Chronic kidney disease (CKD) is a leading global cause of morbidity and mortality, particularly in low- and middle-income countries (LMIC) where access to specialized laboratory tests is limited. Early detection is essential but often delayed due to reliance on serum creatinine-based estimated glomerular filtration rate (eGFR). Artificial intelligence (AI) offers opportunities for simple, sensitive screening models using routinely available variables. Methods: We trained and tested a low-cost machine learning algorithm in a multicenter Latin American dataset of 203,067 anonymized records to identify patients at risk of CKD, defined as an eGFR <60 mL/min/1.73m² (CKD-EPI 2021). Eight routinely available, non-invasive variables were used: age, sex, systolic and diastolic blood pressure, body mass index, hypertension, presence of type 2 diabetes (T2D), and diabetes duration (T2DD). To address the imbalance between CKD-positive and CKD-negative cases, oversampling techniques were applied before splitting the dataset into training (70%), validation (12%), and testing (18%). Using the Arkangel AutoML platform, 424 candidate models were generated, including decision trees, random forests, support vector machines, XGBoost, and deep neural networks. Models were prioritized based on predefined criteria: sensitivity >90%, followed by AUC, precision, specificity, and F1 score. Results: The final model was a decision tree trained in a non-stratified sample with the SMOTE augmentation technique. Sensitivity was 90.2%, specificity 92.7%, precision (PPV) 89%, and AUC 91.4%. Binary regression demonstrated the statistical relevance of all the model’s features in predicting CKD risk in our sample. SHAP analysis identified age and diabetes duration as the most influential features in the final ML model. Conclusions : A decision tree model trained with eight routine clinical variables accurately identified individuals at risk of CKD, achieving high sensitivity and balanced performance without requiring specialized tests. This approach is feasible for large-scale screening in low-resource settings and can be integrated into electronic health records to prioritize confirmatory diagnostics and timely care. It also represents one of the first approximations to CKD diagnosis using ML models trained exclusively on Latin American data. Renal Insufficiency Chronic Machine Learning Artificial Intelligence Risk Assessment Early diagnosis Latin America Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 INTRODUCTION Chronic kidney disease (CKD) is one of the leading causes of worldwide mortality ( 1 , 2 ). It affects over 10% of the adult population, especially people with Type 2 Diabetes (T2D) and hypertension ( 3 ), which contributes to its high morbidity and mortality, ultimately resulting in the need for dialysis or transplantation due to loss of kidney function. Early identification is vital for appropriate treatment and to slow organ deterioration ( 4 ), as pharmacologic and non-pharmacologic interventions to preserve kidney function have been demonstrated to improve prognosis ( 5 ). Diagnosis relies on the creatinine-based estimated glomerular filtration rate (eGFRcr)( 6 , 7 ). If available, the KDIGO Clinical Practice Guideline ( 8 ) recommends cystatin C in the diagnostic calculation of eGFR (eGFRcr-cys). Considering cystatin C’s limited availability in low- and middle-income countries (LMIC) and the relative novelty of the update, the abbreviation eGFR in this manuscript will refer exclusively to eGFRcr. Diagnosis is confirmed when an eGFR < 60mL/min/1.73m2 persists for over three months or if a urinary albumin-creatinine ratio (UACR) measurement is ≥ 30mg/g. However, most affected individuals are asymptomatic until eGFR is under ≤ 30mL/min/1.73m2 ( 4 , 9 ). The risk factors for the presentation and progression of CKD have been extensively described elsewhere ( 10 ), and several diagnostic and prognostic algorithms have been developed. However, the lack of integration of validated models into electronic health records (EHRs) curtails their generalization. Their use is often limited to individual physician criteria in online CKD risk calculators ( 11 ). Furthermore, the complexity and numerosity of attributes used in most clinical algorithms make it unfeasible to apply them in low-resource settings due to their unavailability in EHRs and healthcare systems. Artificial Intelligence (AI) has emerged as a valuable tool for the early detection of CKD. Machine learning (ML) algorithms have shown promise in identifying at-risk patients from large volumes of data, facilitating timely intervention, and improving patient outcomes ( 12 ). Although some clinicians are still sceptical about AI models' genesis, performance, and reliability, these models have proven to be as good as traditional biostatistical methods, with the advantage of processing complex and numerous data accounts faster and more precisely ( 12 , 13 ). Consequently, they are being used to increase the efficiency of the screening and diagnosis processes and provide accessible risk stratification for individuals who would remain undetected until advanced stages of CKD otherwise ( 14 , 15 ). In 2011, Tangri et al. created the clinical algorithm Kidney Failure Risk Equation (KFRE) for CKD progression probability after stage 3. It has demonstrated high discrimination and good calibration and is a reference in the field ( 15 – 17 ). Most research using AI algorithms has followed this general structure, training on public clinical databases such as the UCI ML repository ( 18 ) with anywhere from 24 to 100 variables. These complex models have achieved AUCs between 60–70% ( 19 , 20 ) Although some of these models demonstrated excellent performances (Table 1 ) ( 10 , 21 – 23 ), they rely upon a battery of specialised laboratory tests from secondary information sources, often costly, and usually incomplete. Considering the need for a more straightforward screening process and the resource limitations in LMIC ( 24 ), we aimed at developing a machine learning algorithm capable of mass identification of patients at risk of CKD using simpler, less invasive variables, based on eGFR early alterations. Table 1 Performance metrics for some of the CKD ML algorithms revised. Model # of variables Model type (architecture) Data Source AUC Sensitivity Specificity Precision (PPV) Accuracy F1 score Arkangel AI (Ours) 8 Decision Tree Multicentric dataset from latinamerican health institutions 0.914 0.902 0.927 0.890 0.917 0.896 Ravizza et al. ( 20 ) 7 Logistic Regression IBM Explorys database 0.794 0.682 0.726 0.217 0.721 0.329 KidneyIntelX ( 19 ) 12 Random Forest BioMe Biobank at the Icahn School of Medicine at Mount Sinai and the Penn Medicine Biobank (PMBB) 0.770 - - - - - Qin et al. ( 64 ) 24 Random Forest UCI machine learning repository - - - - 0.998 0.997 Islam et al. ( 13 ) 24 XGBoost UCI machine learning repository - 0.980 - 0.980 0.983 0.980 Jeong et al. ( 65 ) 19 Autoencoder Health examination cohort database provided by the National Health Insurance Service (NHIS) in Korea - 0.956 - 0.959 0.996 0.956 Kumar et al. (66) 24 Genetic Programming UCI machine learning repository 1.000 0.990 1.000 1.000 0.998 1.000 Weber et al. (67) 11–19 Artificial Neural Network (ANN) Retrospective dataset derived from the Jena Part of the 3000 PA text corpus of the Smart Medical Information Technology for Healthcare (SMITH) consortium 0.910 1.000 0.820 - - - Table 1 . Performance metrics, number of features, and sources used for training of some of the CKD ML algorithms revised. MATERIALS AND METHODS Population and variables: Over 600,000 records were collected from three Latin American health institutions: Databases 1 and 2 (D1, D2) from Colombia and Database 3 (D3) from Peru (Fig. 1 ). The independent variables in the pooled data set included eight common features readily available in face-to-face patient consultation and used for model training. They were Age, Sex, Hypertension, Systolic and Diastolic Blood Pressure (SBP-DBP), Body mass Index (BMI), Diabetes (T2D), and Diabetes Duration (T2DD) (Table 2 ). These variables were selected based on expert opinion, due to their availability and strong associations with CKD, summarized in Additional File 1a. They also constitute basic clinical information, which is commonly recorded in routine medical checkups, especially for follow-up in chronic disease. They are also non-invasive and more likely to be registered in clinical records than more sophisticated tests or laboratory results. Data Preprocessing: Missing values were imputed before merging as follows (Fig. 1 , part a): ● Database 1: Hypertension was missing in 172 records, or 0.89% of the Database. It was imputed using the mode. ● Database 2: the Duration of Diabetes was missing in 13,513 records, corresponding to 21,6% of the dataset. We used a Machine Learning (ML) method, the k-Nearest Neighbour imputer. This method completes missing values using k-Nearest Neighbour ( 25 ). Each sample’s missing values are imputed using the mean value from the n nearest neighbour found. In this case, we used the default value for n = 5. ● Database 3: The duration of diabetes was missing from 133 records, corresponding to 0.11% of the data. Here, we used a traditional imputation method, the mean. The eight independent variables were used in their original form. For quantitative variables, values outside physiologically plausible ranges were excluded (Table 2 ). The ninth variable in the Table, SCr, was only used during the training phase. Its use will be discussed in detail in the following sections. Sample Size The sample size was selected from the pooled dataset and based on convenience, according to available records, and the following selection criteria: Inclusion: 1. Records from individuals older than 18 years. 2. Records with at least one registered serum creatinine (SCr) result, in mg/dl. 3. Records from patients with or without chronic conditions (CKD, Hypertension, T2D). Exclusion: 1. No established exclusion criteria were applied after the initial exclusion of unfeasible values for quantitative variables. Variable Distribution Quantitative variables were described using the mean and standard deviation (SD), and qualitative variables were described using absolute and relative frequencies. Table 3 summarizes the distribution of characteristics in each database and in total. Table 2 Variable descriptions. Feature Possible Ranges Unit Nature Order Age 0-100 Number of years Quantitative Continuous Sex 1 − 0 Man/Woman Qualitative Nominal Hypertension 1 − 0 Yes/No Qualitative Nominal Systolic Blood Pressure (SBP) 40–200 mmHg Quantitative Continuous Diastolic Blood Pressure (DBP) 40–120 mmHg Quantitative Continuous Body Mass Index (BMI) 12–50 Kg/m Quantitative Continuous Diabetes (T2D) 1 − 0 Yes/No Qualitative Nominal Diabetes Duration (T2DD) 0–60 Number of years Qualitative Nominal Serum Creatinine (SCr)* 0–20 mg/dl Quantitative Continuous *SCr was used ONLY during training to calculate the eGFR primary outcome, with the 2021 CKD-EPI equation. Table 3 Basal Distribution of Clinical and Demographic Characteristics (before augmentation) Number of registries Total Database 1 Database 2 Database 3 N = 203,067 n = 19,276 n = 62,553 n = 121,238 Quantitative Variables mean std min, max mean std min, max mean std min, max mean std min, max Age 47.55 17.96 18, 100 65.23 11.86 21, 99 64.64 13.69 18, 100 36.08 13.69 18, 100 SBP 115.74 14.26 40, 180 129.69 15.58 40, 180 122.61 15.57 40, 180 109.99 9.4 60, 180 DBP 73.27 9.72 40,12 77.45 8.89 40, 120 78.09 10.85 40, 120 70.11 7.77 50, 100 BMI 26.98 4.33 12, 50 27.89 5.26 12.42, 49.95 26.82 5.16 12, 50 26.92 3.62 13.92, 49.67 Diabetes Duration (T2DD) 1.55 4.16 0, 56.87 9.43 7.22 0, 56.87 2.07 3.95 0, 9.6 0.03 0.53 0, 31 Qualitative Variables Absolute Frequency (n) Relative Frequency (%) Absolute Frequency (n) Relative Frequency (%) Absolute Frequency (n) Relative Frequency (%) Absolute Frequency (n) Relative Frequency (%) Sex (F) 66,863 32.93% 14,32 74.29% 40,043 64.01% 12,5 10.31 Hypertension (Yes) 53,864 26.53% 15,855 82.25% 36,995 59.14% 1,014 0.84 Diabetes (Yes) 32,909 16.21% 18,922 98.16% 13,513 21.60% 474 0.39 CKD (Yes) 27,502 13.54% 8,622 44.73% 18,453 29.50% 427 0.35 Outcome The dependent variable or primary outcome was the estimated Glomerular Filtration Rate (eGFR). Table 3 shows that CKD was a variable in the original database. We decided not to use this specific classification as the standard of reference for model training because we could not determine which criteria or formula was used to define it. Instead, we followed the 2024 KDIGO Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease (KDIGO 24) ( 8 ). We took the SCr measurement in the records to calculate eGFR using the 2021 updated version of the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula to standardize the measurement (Additional File 1b). As per the guideline, we defined a threshold of eGFR equal to 60mL/min/1.73m2 to separate individuals into two categories: on the one hand, those with risk of presenting with CKD when eGFR was < 60mL/min/1.73m2 (defined in Chap. 2 of the guideline as G3a - Mildly to moderately increased CKD risk). These individuals would undergo medical follow-up according to guidelines, including confirmatory tests (a second eGFR, albuminuria), and specialized kidney function evaluation. On the other hand, an eGFR ≥ 60mL/min/1.73m2 would classify the record as belonging to someone who did not need to be prioritized for immediate further testing (although it does not mean zero risk, and the guideline recommends UACR). Table 3 . Basal Distribution of Clinical and Demographic Characteristics (before augmentation) Figure 1 . Population Flowchart *CKD Positive is defined as an eGFR /=60mL/min/1.73m2. Both were calculated using the 2021 updated version of the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula. These values were used as the reference standard for comparing the AI model and were only used during the training phase of the algorithm development. None of them were used during the testing phase. Designed using Canva https://www.canva.com/design Legend Part a. The corresponding lines and boxes depict each step's exclusion and inclusion criteria and the imputation method. Part b. Absolute and relative frequencies of excluded and imputed population, Part c. Definite sample size and outcome distribution. AI Methods Data Augmentation The imbalance between positive and negative classes is common in ML algorithms. To better represent all classes (positive and negative CKD risk), we used an augmentation called Synthetic Minority Over-Sampling Technique (SMOTE) ( 26 ). When using SMOTE, the researcher determines a desired ratio of the samples in the minority class to the number of samples in the majority class after resampling to make the classes more equal in total numbers. Other authors have described the SMOTE augmentation technique as a valid strategy for unbalanced sample sizes in ML model training, significantly improving the classification's balance and accuracy ( 27 – 29 ). The augmentation process was performed as follows: 1. Total sample size, before augmentation = N 203067 (Fig. 1 , part a). 2. CKD positive class before augmentation = n 27502 (see Table 2 ). 3. CKD negative class before augmentation = n 175565 4. Using SMOTE 1 (desired ratio of 1 between classes), the CKD positive class was augmented until reaching the absolute number of the negative class, resulting in a positive class of n = 175565 after augmentation (Fig. 1 , part c) 5. Conversely, the CKD negative class was not modified after augmentation using SMOTE 1 (n 175565). 6. The absolute number of augmented data was = n 148063. 7. Therefore, the new total augmented sample was = N 351130 8. After augmentation techniques using SMOTE 1, 148063 synthetic records were created. These represent 42.2% of the total sample. Then, the training and following steps were performed using the augmented sample size (feature distribution in the augmented sample is exhibited in Additional File 1c). Training and Testing Phases The construction of an AI algorithm can be separated into two main parts: the training and testing phases. The total sample after augmentation was 351130. We split it into 70% for training, 12% for validation, and 18% for testing (Table 4 ). During training, the model learns to recognize patterns. The validation set evaluates performance as the model is trained; it is considered part of the training set (consequently, the total split for the training phase was 82%). In the testing phase, the model is used to perform outcome predictions on a portion of the sample size that it has never been in contact with, which allows for the internal validation of the model. The training phase was performed using the supervised learning approach, in which the human programmer trains the algorithms by showing them examples marked as “positive” and “negative” or “healthy” and “ill,” etc., according to the outcome’s nature. The variables used to train the algorithms included the eight non-invasive, readily available features listed in Table 2 . The ninth, SCr, calculates baseline eGFR using the standardized 2021 CKD EPI formula. Serum creatinine was exclusively present during the model's supervised learning training. The ML's automatic learning capacity uses the information of the estimated GFR to assimilate the positive and negative case criteria. Then, the model can acknowledge patterns and recognize associations among the other eight variables that are not evident to the naked eye. Consequently, it determines how different combinations or values of these features are associated with an altered eGFR when no SCr or eGFR are provided. This capability is assessed during the testing phase. Here, the model is evaluated using a portion of the database previously unknown to the algorithm. In this stage, the model is given records without any data on eGFR or any other information directly related to renal function measurement. It then performs the predictions or inferences, using the other eight features. As the outcome, it yields the probability of an eGFR being higher, equal, or lower than 60mL/min/1.73m2. However, its benefit is not simply estimating an eGFR. Recognizing these association patterns among variables determines how different combinations or values of the features are associated with an altered eGFR (< 60mL/min/1.73m 2 ) as an outcome, and how much each of the variables in the model influences it (see SHAP values ahead). Table 4 Distribution of data splits Phase Absolute number (n) Percentage (%) Train 245791 70 Validate 42136 12 Test 63203 18 Total 351130 100 Arkangel AI platform Arkangel AI is an Auto ML tool designed to train AI models based on clinical data. It was built to simplify the human programmer's training and experimentation process by automatically training and evaluating multiple candidate algorithms based on the variables provided ( 30 ). The models are trained exclusively by ML engineers with specific access credentials. Based on the datasets, they perform automatic but exhaustive training experiments using many combinations of architectures and hyperparameters to find the best-performing models. The app trains around 25 models per experiment. It is also set to automatically use cross-validation to assess the model and hold out a portion of the dataset to estimate performance and compute all relevant metrics. According to these cross-validated metrics, models can be organized from highest to lowest performance. The results stem from multiple automated iterations executed by the app. Each iteration presents a confidence score, which estimates how confident the model is of its result. Finally, the platform automatically organises the algorithms and recommends the best-performing models based on their performance metrics (e.g., sensitivity, specificity, or AUC) (Fig. 3 ). The architecture search includes KNeighborsClassifier (KNC), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Gradient Boosting (GB), Random Forest Classifier (RFC), and deep learning (DL) models like Deep Fully Connected Neural Networks (DNN). Finally, these models are available so users can observe and assess their composition and performance through the user-friendly interface. AI Model Training Our training was conducted on Arkangel’s platform. We performed several experiments. We trained models of several architectures, like DNNs, RFs, and KNCs, and explored different hyperparameters. Additional File 2 displays the extensive list of algorithms trained for this study, organized in seven sheets (using Google Sheets) by architecture, demonstrating their technical characteristics and hyperparameters. Figure 2 exemplifies the Sheet file with a screenshot. These architectures were also trained by exploring different sample size distributions in the search for the best-performing models. For example, we examined training with the Colombian registries (Col) to test them in the Peruvian dataset. We also stratified the sample into Diabetic (T2D) and non-diabetic patient records (non-T2D), on top of the usual approach of using the sample size (AllData or Merged). Additionally, we experimented with SMOTE using ratios of 0.5 and 1. All the resulting models constitute Additional File 2. The projects and models were named using the terms in parentheses above, according to the sample distribution they were trained in, the hyperparameters, and the augmentation technique used (Fig. 2 , columns A and B). The file also contains each algorithm's performance: accuracy, sensitivity (or recall), specificity, positive predictive value (PPV) (or precision), Area Under the Curve (AUC), and F1. Table 5 summarizes the algorithms by architecture. Table 6 displays the number of experiments or projects explored and the number of models trained on each experiment. Table 6 Machine Learning Model Distribution by Experiment/Project Name Experiment/ Project Name Number of models AllDataAugmentedDownsampledErc 20 AllDataSmoteEqual0Point5Erc 20 AllDataSmoteEqual1Erc 20 COLAugmentedpositivesx2 20 ColDataAugmentedDownsampledErc 22 ColDataImputedSmoteEqual0Point5Erc 22 ColDataSmoteEqual1Erc 22 DbSmoteEqual0Point5Erc 22 DbSmoteEqual1Erc 66 DiabeticosImputadoMerged 23 MergedColombiaImputed 22 mergedColombiaImputedUpsampled 22 MergedColombiaImputedUpsampledDownsampled 22 MergedImputed 20 MergedImputedUpsampled 20 MergedImputedUpsampledDownsampled 20 NdbSmoteEqual1Erc 20 nodiabeticosMerged 21 Total 424 Figure 2 . Screenshot of Additional File 2 displaying the sheet corresponding to the Decision Tree architecture*. *Row 15, highlighted in yellow, presents some details of the final selected model. Developed with Google Sheets https://docs.google.com/spreadsheets/d/16OAaJ5FbgINUhguW_DtNIjJlhFNltXExDcp6dXCGaXw/edit?gid=1799842340#gid=1799842340 AI Model Prioritization The trained models are stored in the App and grouped by project name (Fig. 3 ). Because this is a screening use case, sensitivity (true-positive rate) was the primary selection criterion for prioritizing the models. At the same time, other metrics were required to remain within predefined acceptable ranges. They are stated below. Figure 3 illustrates how models can be organized from higher to lower sensitivity or the user’s preferred metric. Figure 4 shows how the results of each algorithm are presented in the App. Figure 3 . Screenshot of the app’s interface when accessing a Project*. Taken from https://hippocrates.arkangel.a . All rights reserved. *Project’s name in the figure: alldataSMOTEequal1erc, at the top left of the screen. Legend The metrics can be organized from highest to lowest performance, choosing any metric from the dropdown menu. Figure 4 . Screenshot of the information displayed on the App for every model trained or tested Taken from https://hippocrates.arkangel.a . All rights reserved. Legend Performance metrics are summarized in the screen's upper left, while the model’s technical characteristics are displayed in the upper right. The confusion matrix is represented in the lower right quadrant. The Shap values, which quantify the influence of each factor on the model's outcome, are also depicted (not shown in the Figure). AI algorithm selection Once all projects were finalized, we evaluated and organized them from highest to lowest sensitivity. We grouped them by project, as can be replicated in the second sheet inside Additional File 2. We did not use the grouping by architecture because not all of them produced algorithms that were competitive in performance. Then, we selected the top-performing models following predefined rules, agreed upon by the research and medical team, as follows: 1. Select the models with a sensitivity/recall higher than 90% 2. From those, the ones with an AUC higher than 90% 3. Then, those with a PPV/precision higher than 90% 4. From those, the ones with specificity higher than 85% 5. F1 higher than 90% 6. If the best models belonged to those trained in the stratified sample size, from both T2D and non-T2D, then an ensemble model of the two strata would be explored. 7. If a model trained in the Colombian population were classified as top-performing, its further assessment would depend on how it performed using the Peruvian sample. 8. Models trained using the entire merged sample size would be preferred if the other best-performing models were those trained separately in the T2D or the non-T2D stratified samples. Criteria six to eight were agreed upon in case several algorithms fulfilled the first five criteria. By applying the criterion, we ensured a highly sensitive model with optimal discrimination ability (AUC) and capacity to predict a positive case (PPV) despite lower specificity. An algorithm with satisfactory, balanced global metrics (F1). The final step of this process is depicted in Table 7 . The detailed steps are displayed on each sheet of Additional File 3. The chosen algorithm is presented using simple tables and figures: technical characteristics (Fig. 3 ), performance metrics, and a confusion matrix to calculate them (Table 8 ), and the Shapley Additive Values (Fig. 6 ), which quantify the weight of each variable in the algorithm’s inference, supporting model explainability. Figures 3 to 5 display screenshots of how results are presented in the Arkangel AI web app. Model Validation The Arkangel app uses iterative cross-validation to automatically evaluate the model's results; it is a functionality embedded in its algorithm. Sensitivity analysis is further assessed by appraising the model's behaviour on the records of diabetic and non-diabetic individuals in the test set (Additional File 1m-1o). Additionally, we performed stratified analyses by gender and age group, framing it into three segments: younger than forty, forty to sixty, and older than sixty years old (Additional File 1p-1r). We also present the 95% Confidence Intervals (CI 95%) for performance metrics. In addition to AI and ML analyses, we undertook traditional statistical analyses. First, we conducted bivariate hypothesis tests to obtain p-values to explore statistical differences (p < 0.05). Then we used regression models to evaluate associations, model variation explainability, and to control for confusion. The prevalence ratio was also determined to explore associations not evident in the AI model. Overfitting Control Possible overfitting was assessed on two fronts. The first was comparing the training and testing performance metrics and their respective confusion matrices (Tables 8 to 10 ). The second was the evaluation of the maximum depth graph, which depicts the model's learning behaviour when training and testing, in terms of accuracy ( 31 ) (Fig. 7 ). As approaches to control for this factor we used automatic cross validation for all results and confidence intervals, de-duplication in the row dataset ( 32 ), regularization (using L1, L2 to prevent too large coefficients), early stopping (by monitoring the validation set during training for when it starts deteriorating) ( 31 ) and data augmentation (to make the model more robust and less likely to overfit) ( 33 ). Methods for Traditional Statistical Analyses We performed traditional statistical data analyses to support and strengthen our methodology and conclusions. We performed the Shapiro-Wilk test and conducted a bivariate analysis accordingly. We used Pearson/Spearman’s correlations for independent quantitative variables and the Chi-squared test for qualitative variables (Additional File AF 1dt o 1f). Student´s T/Mann-Whitney’s U test (Additional File 1g) was used to explore differences in means or medians in quantitative-qualitative variable associations. Then we confirmed the association between each independent variable and the outcome (Additional File 1h). Finally, multivariate analysis (Additional File 1j) was performed to evaluate associations in the presence of all model variables to control for confusion; the assumptions of independence of errors and homoscedasticity were previously evaluated (Additional File 1i). Only variables with p-values of > 0.05 were maintained in the model. Variance due to the regression models was assessed using Nagelkerke’s R-squared and AIC (Additional File 1k). Additional File 1l shows the explainability for different models, exploring the model’s variability when one variable was included or dropped at a time. ETHICAL CONSIDERATIONS This paper was developed in accordance with the stipulations and rules cited in the Declaration of Helsinki and the CIOMS International Ethical Guidelines for Health-related Research Involving Humans in their last version from 2016. It describes the algorithm training, testing, and selection process as thoroughly as possible to ensure transparency. Databases were voluntarily provided and completely anonymised. The study was approved by an external ethics committee in Peru and by the ethics committees at each of the health institutions in Colombia. Additional File 4 compiles each committee's approved research protocols and acceptance letters. No exclusions were made based on gender, race, sexual orientation, or nationality. Figure 5 . User interface for performing single predictions based on the algorithm selected and the patient’s variables Taken from https://hippocrates.arkangel.a . All rights reserved. Legend Using patient information as input (right upper quadrant), the algorithm predicts his/her CKD risk. CKD risk is defined as an eGFR < 60mL/min/1.73m2. Prediction (1 = risk, 0 = not at risk) is displayed on the left as the primary outcome, accompanied by the SHAP values for each variable. RESULTS Study population Three databases from three healthcare institutions were obtained and merged, giving 645412 initial registries (Fig. 1 , part a). After applying quality control and selection criteria to the merged database (part b), we obtained a sample size of 203067 registries. In the baseline sample, CKD prevalence was 13.54%, since 27502 records presented with an eGFR < 60 ml/min/1.73m2, as defined for this study’s outcome, and calculated using the CKD EPI 2021 updated formula. Diabetic individuals accounted for 16.2% of the sample (32909). The mean age was 47.55 years, and the mean time since T2D diagnosis was 1.55 years. Thirty-three percent of registries were female. Twenty-seven were hypertensive. The average record belonged to a normotensive, overweight patient (Table 3 ). When evaluated separately, Diabetes was almost five times more prevalent in D1 than in D2. This difference was even more marked in D3, where virtually no records reported T2D. CKD and hypertension were also minimal in D3, while more than half of the records in D1 presented with these chronic diseases. Blood pressure and BMI were similar among the three databases. However, D3 was predominantly composed of young male adults, while D1 and D2 were twice the age and more balanced in the sex distribution. Algorithm Testing and Selection The total number of models trained was 424 (Additional File 2). Table 5 summarizes them according to their distribution by architecture. Table 6 demonstrates the number of experiments performed ( 18 ) and the number of models trained on each (22 for most). After grouping them by experiment, the algorithms with sensitivities > 90% were chosen, resulting in 120 preselected high-performing models (Additional File 3). Then we selected the ones with AUC > 90% (24 algorithms), and then those with precision > 90% (2 algorithms remaining) (Table 7 ). Specificity > 85% was the following criterion, resulting in the selection of the final model. Criteria 5 to 8 were not needed. We observed general poor performance in the models trained using only the Colombian (Col) sample; they never obtained sensitivities higher than 86% and presented a low average AUC (Additional File 3). Therefore, they were never tested in the Peruvian population. The performances derived from models trained on stratified data were higher for non-T2D. On the contrary, models trained in T2D records presented with lower general metrics, and sensitivities did not surpass 81% for any of them (Additional File 3, sheet 2). Consequently, none of the diabetic-strata trained models made the top performing, according to the criteria stated in the Methodology section (Tables 11 and 12 ). The final selected model was trained on the undivided sample size, using an augmentation of SMOTE 1. It had a Decision Tree architecture with a depth of 33 layers (Fig. 3 ). The augmentation technique was explained in detail in the Methods section under “Data Augmentation.” The performance metrics and confusion matrix are displayed in Table 8 . Algorithm Performance and Validation Compared to the standard of reference for risk of CKD, which was defined as an eGFR < 60 ml/min/1.73m2 and calculated using the standardized 2021 CKD EPI equation, our model achieved a sensitivity of 91.4%, specificity of 90.1%, precision of 90.2%, AUC of 90.9%, f1 score of 90.8%, and accuracy of 90.7%. SHAP values were calculated and are shown in Fig. 6 . The most influential variable was age, followed by T2DD and BMI. The SHAP value for the binary T2D was consistently zero for most of the best-performing models. Performance in the diabetes-status stratified sample shows an equilibrium between metrics in T2D and non-T2D (Additional files 1m-1o). Sensitivity and precision were higher for diabetics, and specificity and accuracy were higher for non-diabetics. It is important not to confuse this stratified by diabetes status assessment of the chosen model with evaluating the performance of algorithms trained in those divided populations, which yielded poor-performing models, before final model selection. The Z test p-value demonstrates no statistical differences between the performance metrics of both groups. Additionally, we performed stratified analyses by sex and age (Additional Files 1p to 1r). The prevalence rates for males and females were 53% and 47%, respectively. The analysis of age stratified as 60 years old provided evidence that CKD risk in patients younger than sixty years was similar (24% prevalence). For those older than sixty, it was twice as likely to be present (53%) in our population. Table 10 presents absolute differences in performance metrics in training and testing datasets. The absolute difference in accuracy is lower than 10 (9.4), which may or may not indicate overfitting, since there is no established threshold (see Discussion) ( 33 , 34 ). The maximum depth curve (Fig. 7 ) also presents some indicators of overfitting, such as a validation curve surpassed by the training curve. Despite this, it evidences a plateau from depth = 29, which means it reaches stability and non-degradation with increases in depth, which is an argument in favour of the non-overfitted model. Traditional Statistics Results Additional Files 1d to 1l provide evidence of the tests described in Methods. We found a statistical association among all the independent variables and the outcome. We confirmed the assumption of independence of errors but found a high probability of heteroscedasticity (Additional File 1i). Consequently, we performed a multivariate analysis with a binomial regression (Additional File 1j). All the features tested had a p-value < 0.05. We evaluated the model, including the eight variables in Table 2 , and dropping one at a time, to assess performance and p-values (Additional File 1l). These were compared based on McFadden and Nagelkerke’s R-squared and AIC (Additional File 1k). The model with all features presented the highest values. Confusion was controlled in the presence of all variables, and statistical associations were sustained. Consequently, all eight variables were statistically and clinically relevant and were kept in the model. Additional File 1a summarizes the clinical associations described and supported in the literature. We also calculated and interpreted the Prevalence Ratio (PR). Its value was closer to the unit (1 = null association) for SBP and DBP, followed by BMI, Diabetes Duration, and age, in that order. The highest PR in our sample was for sex (male) with 2.6, followed by the presence of Diabetes (PR 2) and Hypertension (PR 1.9). Nagelkerke’s R-squared indicates that this regression explains 53% of the outcome variability under the conditions of our study. Table 5 Machine Learning Model Distribution by Architecture Arquitecture Absolute Frequency Relative Frequency DecisionTree 38 8.96% NN 360 84.91% KNC 21 4.95% LogRegression 2 0.47% RFC 2 0.47% SVC 1 0.24% Total 424 100.00% Table 6 . Machine Learning Model Distribution by Experiment/Project Name Table 7 The top two performing models* Project name Sample Distribution Model Name Architecture Sensitivity/ Recall AUC PPV/ Precision Specificity F1Score Accuracy AllDataSmoteEqual1Erc All data DecisionTreeC_max_depth-33_min_samples_split-2_RandomSearch DecisionTree 0.914 0.909 0.902 0.901 0.908 0.907 DbSmoteEqual1Erc Stratified data - db RFC_n_estimators-388_max_depth-35_RandomSearch RFC 0.966 0.9378 0.906 0.6423 0.935 0.895 *The defining step was to apply criterion 4: “specificity higher than 85%”, which led to the model in bold. Table 8 Testing phase dataset’s confusion matrix and performance metrics for the selected best-performing model Arkangel app Classification Testing dataset CKD EPI = 60mL/min/1.73m 2 (false case) Total Metric Value IC 0.95% - Lower IC 0.95% - Upper Sensitivity 0.914 0.911 0.917 Specificity 0.901 0.898 0.904 Precision 0.902 0.898 0.905 AI Algorithm true case 28797 3142 31939 Accuracy 0.907 0.905 0.910 AI Algorithm false case 2718 28546 31264 F1 score 0.908 0.905 0.911 Total 31515 31688 63203 AUC 0.909 0.910 0.910 Table 9 Training phase dataset’s confusion matrix and performance metrics for the chosen model Arkangel app Classification Training dataset CKD EPI = 60mL/min/1.73m 2 (false case) Total Metric Value IC 0.95% - Lower IC 0.95% - Upper Sensitivity 0.997 0.997 0.997 Specificity 0.996 0.996 0.997 Precision 0.996 0.996 0.997 AI Algorithm true case 143619 505 144124 Accuracy 0.997 0.997 0.997 AI Algorithm false case 431 143372 143803 F1 score 0.997 0.996 0.997 Total 144050 143877 287927 AUC 0.997 0.996 0.997 Table 10 Comparison between training and testing performances Model Sensitivity Specificity Precision Accuracy F1 score AUC Training 0.997 0.996 0.997 0.996 0.997 0.997 Testing 0.914 0.901 0.909 0.902 0.97 0.908 Difference 0.083 0.095 0.088 0.094 0.027 0.089 Z test p-value 0.078 0.077 0.079 0.08 0.065 0.600 Figure 6 . Shapley Additive Values for each attribute on the model Legend The Shapley Additive Values are a measurement used in ML models to quantify each variable's weight in the model, on top of performance metrics. Despite its documented strong association, T2D appeared as the least influential variable for the algorithm's outcome. See the Discussion section for in-depth analysis. Figure 7 . Maximum Depth of Individual Estimators Taken from https://hippocrates.arkangel.a . All rights reserved. Legend The chosen model, DecisionTreeC_max_depth-33_min_samples_split-2_RandomSearch, reveals clinically optimal performance despite showing traditional indicators of overfitting, such as the training curve surpassing the validation curve. The model achieved 99.7% accuracy in the training set and 91% in validation, presenting a gap of 9 percentage points between the two sets. The stability of validation performance, evidenced by the plateau reached from depth = 29, and the absence of significant model degradation, suggests that the observed difference does not compromise clinical generalization. The conclusion is that the model does not present substantial overfitting and can be considered viable for implementation in medical screening settings, where practical utility and case identification capability outweigh strict criteria for statistical perfection. Table 11 Best performing model for those trained in the Colombian sample exclusively Project name Training Sample Distribution Model Name Architecture Sensitivity/ Recall AUC PPV/ Precision Specificity F1 Score Accuracy COLAugmentedpositivesx2 Colombian Sample DeepFullyConnectedNeuralNetwork_Adagrad_l1_dropoutFalse_numHiddenLayers5 Deep Fully Connected Neural Network 0.999 0.828* 0.672 0.001 0.804 0.672 *The metric in bold signifies the selection criteria for which the respective algorithm was disqualified (see Methods). In this case, AUC should be > 90% to qualify. Table 12 Best performing model for those trained in the diabetes/non diabetes stratified sample Project name Training Sample Distribution Model Name Architecture Sensitivity/ Recall AUC PPV/ Precision Specificity F1 Score Accuracy DbSmoteEqual1Erc Diabetic Sample RFC_n_estimators-388_max_depth-35_RandomSearch Random Forest Classifier 0.967 0.938 0.906 0.643* 0.935 0.895 NdbSmoteEqual1Erc Non-diabetic sample DecisionTreeC_max_depth-26_min_samples_split-12_RandomSearch DecisionTreeClassifier 0.901 0.951 0.885* 0.923 0.893 0.914 *The metric bold signifies the selection criteria for which the respective algorithm was disqualified (see Methods). In this case, PPV should be > 90% (non-diabetics), and Specificity > 85% (diabetics) to qualify. DISCUSSION We trained AI algorithms using anonymised data from over 200.000 Colombian and Peruvian EHRs, attempting to reliably classify their CKD risk by predicting the probability of having a eGFR <60mL/min/1.73m2, based on eight readily available, non-invasive, and measurable at the point-of-care attributes: age, sex, SBP, DBP, BMI, hypertension, presence of diabetes (T2D), and diabetes duration (T2DD). The study responds to the need for an extremely sensitive screening strategy for CKD, especially for LMIC, where human and technical resources for finding and prioritizing patients are scarce. Research in LMIC has emphasised producing models using fewer variables than other good-performing ML models, without compromising their quality. Rashed-Al-Mahfuz et al. trained AI algorithms for CKD prediction in Bangladesh, with an accuracy of 99.50%, sensitivity of 98.75%, specificity of 100%, precision 100%, and AUC of 99.38%. Their best-performing model was a Random Forest, with 13 features from those in the UCI-ML repository. Using SHAP values, they determined that the most influential variables for their model were haemoglobin levels, hypertension, and blood sugar levels (21). In Colombia, Isaza-Ruget et al. (2024) developed an AI model to predict the risk of CKD progression in patients in stages 3-5, using Gradient Boosting and ten features: sex, residence, diabetes, hypertension, haemoglobin, creatinine, HDL, LDL, and eGFR (14). However, the population with early alterations in eGFR would still go unnoticed. We tested and internally validated our algorithm, evidencing its optimal performance, using even fewer attributes. The chosen model was a Decision Tree with 33 layers, trained on an augmented sample size using SMOTE=1 (351130 EHRs). Within the study’s conditions, our model detects individuals at risk of presenting altered eGFR defined as <60mL/min/1.73m2 (mildly to moderate increased CKD risk, according to 2024 KIDGO guidelines), with sensitivity of 91.4%, specificity of 90.1%, a precision 90.2%, AUC of 90.9% and f1 score of 90.8%, only using eight variables. Automatic cross-validation was carried out on the performance metrics in the Arkangel App. The narrow 95% CIs in Table 8 support the validity of the results, on top of the hypothesis test comparisons (no statistical differences, p-value >0.05) of the proportions of performance metrics in the training and testing datasets. Subgroup analyses, implemented by running the model in the sample stratified by diabetes status, evidenced a balanced, consistent, high-performing model in both populations (AF 1l to 1n). This stratified analysis is also interesting given the difference in absolute numbers of diabetic patients compared to non-diabetic patients in the base population. Only 16% of EHRs belonged to diabetic patients, predominantly from the Colombian population (D1-D2). They were older and more chronically ill than the Peruvian population (D3) (mean ages 65 and 36, respectively). There are two important things to emphasise regarding this issue: the first is that the nature of D3 most likely explains the baseline population differences; the Peruvian healthcare company was mainly an insurer for vehicle-related accidents, which are more common in the ages under 35 (35,36). Conversely, Colombian health institutions, although they also acted as insurance companies, had a broader spectrum of health-related events, especially chronic disease care. The second fact is that this imbalance is common in retrospective, secondary source research (37), especially in the face of AI capabilities (38–41), where enormous quantities of data are expected to be processed at once. This bottleneck is worsened when trying to find specific literature about hetero/homogeneity of sample sizes in ML-trained models; there is virtually none. For the results to be valid, the sample size should be as similar as possible in all characteristics, except the one expected as the outcome. This traditional epidemiological view is supported by statistical theorems and methods amply described elsewhere (42–47). Nevertheless, one could also state that to achieve maximum efficiency with AI models, it is not a bad practice to include different base populations in the studies and show the model’s possible variations of relevant characteristics (48,49). The latter was the rationale behind exploring the effect of sample size’s hetero/homogeneity by training using different sample distributions. We trained 88 models using only the Colombian and evaluated their metrics before testing them on the Peruvian sample as an external validation. We never got to that point because, as mentioned in Results, the model's performances in this “more homogeneous” sample resulted in poor metrics that did not comply with the selection criteria stated in Methods. An example is the DNN model shown in Table 11, which was the top performer in this group of experiments. The same goes for the models trained on the diabetic/non-diabetic population (152 algorithms). Table 12 shows the best-performing model for these experiments, in which a model was trained on each group. The algorithms trained in diabetics had even poorer performance, which could be explained by the class imbalance, in favour of the non-diabetics, even after SMOTE augmentation of the diabetic strata. This evidence meant that training with a stratified sample was not satisfactory for addressing the imbalance. The performance of the non-diabetic trained model was better, but still not up to established standards. In addition, these two models needed an extra step: an ensemble learning model approach, where each algorithm is weighted and computed to produce a final, single result. Although we explored this avenue for the present study, the lack of a strong rationale for dividing the sample into diabetes strata, the evidence of better results by controlling for class imbalance using ML augmentation techniques alone, and the complexity of the method itself, determined that these experiments were not included in the tables displayed. Their results would be available upon reasonable request. Another essential topic to discuss is the SHAP values and their meaning compared to traditional statistical analyses. According to SHAP, the ranking of features by their importance or weight for the final model’s decision is Age, T2DD, BMI, SBP, sex, DBP, hypertension, and T2D, in that order. At first glance, this is remarkable because there is a strong association between diabetes and CKD, as documented in countless scientific articles (Additional File 1a). After a closer look, this association might be observed with T2DD in second place, representing the burden of the association with CKD. Nevertheless, it ranks first since it provides more information about the condition (years of disease progression vs. presence or absence of T2D). To test this hypothesis, we evaluated collinearity using the Variance Inflation Factor (VIF) (Additional File 1s). General recommendation is that a VIF value lower than five is acceptable (50–52), according to the study's nature and the variable's clinical relevance. T2DD and T2D have a VIF of three, so we believe it is prudent to declare potential collinearity. However, due to the clinical value of the variables, this is not enough to warrant exclusion. Besides, literature sustains that ML architectures, particularly Random Forest and Decision Trees, are unaffected by collinearity (53–55). They are uniquely qualified to deal with multicollinearity through the conditional subgroups technique, a strategy used in Explainable Artificial Intelligence (XAI) methods. This method manages multicollinearity using conditional subgroups with permutation feature importance and partial dependency plot. In Decision Trees, data splits are made based on individual variables, rather than coefficients, into groups that secure a more homogeneous distribution in one group and a more heterogeneous distribution in the rest. The rationale behind this method is to create groups so that the feature of interest is less dependent on all other features in all the subgroups (56). To further evaluate the behaviour of variables in the AI model, we performed bi- and multivariate analyses with traditional statistical methods. All features were confirmed as statistically associated with CKD in the presence of the others. The best model included the eight variables, which explained 53.1% of our sample's model variability. The magnitude of the positive association was evidenced by the Prevalence Ratios (PR): T2DD was among the weaker associations, while T2D (PR 2) and sex (PR 2.6) displayed the highest associations. The latter ranking is not equivalent in meaning to the ranking established by SHAP. These values are a machine learning interpretability tool that quantifies the contribution of each feature to predicting an individual instance in a model (57–59). They provide a local explanation of model output, showing how much each feature increases or decreases the predicted value relative to a baseline. Conversely, PRs are epidemiological/statistical measures that compare the prevalence of an outcome in two groups (60). They quantify the relative risk or association between exposure and outcome in population-level data (61). Consequently, all results point to the maintenance of both variables in the model: their documented association with the outcome, the relevance of T2DD in the ML model, the strong positive association of T2D proven by its PR in statistical analysis, and the better statistical explainability of the variations in outcome when including all eight variables in the binary regression. The fourth significant issue is the possibility of overfitting. According to accuracy, Figure 7 shows the model's learning behaviour over time. Overfitting is generally identified when a model performs significantly better on the training data than on the testing data, indicating poor generalization (31,33,34). The ideal graph would display an increasing curve that stabilizes at the threshold. The training curve should not surpass the validation curve. However, there is no threshold to define this difference in performance. It has been described that this difference in accuracy should not be over 10%. Although our model differs in accuracy by 9.4 percentage points, its performance in both datasets and the subgroup assessment is optimal. Additionally, we employed multiple approaches to control for it (see Methods). Specifically, for the augmentation strategy, we used SMOTE, which is an oversampling technique that is less prone to introducing overfitting than others because synthetic data is created by an algorithm (K-nearest neighbour) that seeks to balance the dataset (33). This motivates us to share our algorithm's development and internal validation through this manuscript. This algorithm needs external validation to appraise whether its performance metrics are maintained in populations with entirely different characteristics. The study is being developed thanks to our access to the iCaReMe (CardioRenal and Metabolic) global registry (NCT03549754). The dataset contains healthcare data of individuals across 11 countries (62,63). Conclusions Our study demonstrates that a decision tree model trained with eight non-invasive clinical variables can accurately identify individuals at risk of CKD without requiring specialized tests. This approach is feasible for large-scale screening in low-resource settings and supports integration into electronic health records to prioritize confirmatory testing and timely care. Declarations Clinical Trial Number : not applicable ETHICS APPROVAL AND CONSENT TO PARTICIPATE This study is classified as minimal risk according to Resolution 8430 of 1993 (Colombia), which stipulates that minimal risk-observational studies, based on secondary data, exclusively from clinical sources, do not need individual consent (64). In addition, the Declaration of Helsinki (2022) and the CIOMS Guidelines (2016) state that individual informed consent may be waived when the conditions of minimal risk, impracticability of obtaining consent, and high social and scientific value are met, while always ensuring strict confidentiality of the information (65). Furthermore, the study was reviewed and approved by an external ethics committee in Peru and the ethics committees at each participating health institution in Colombia (see Additional File 4, which includes the letters of ethics approval, the research protocols, and the letter we sent to the external committee in Perú, in PDF format into one compressed file) CONSENT FOR PUBLICATION Not applicable. In accordance with the Resolution 8430 of 1993 from Colombia, the Declaration of Helsinki (2022) and the CIOMS Guidelines (2016), individual consent for publication was waived, given that the research was based on anonymized secondary data, was classified as minimal risk, it was not operationally feasible to obtain individual consent, and strict confidentiality of the information was guaranteed at all times. There are no images or any other sensitive information that could compromise anonymity. PUBLICATION AVAILABILITY OF DATA AND MATERIALS The datasets used and/or analysed during the current study are not publicly available due to restrictions related to patient confidentiality and institutional agreements. However, anonymised data may be made available from the corresponding author on reasonable request and with permission of the participating health institutions. COMPETING INTERESTS All authors completed the International Committee of Medical Journal Editors (ICMJE) disclosure form from https://www.icmje.org/disclosure-of-interest/ JM, AP, DJ, JZ, IL, and NCV work at Arkangel AI. IL received support from AstraZeneca to attend and present a poster at the Colombian Nephrology Conference from ASOCOLNEF on August 1, 2, and 3, 2024. DC works as Medical Therapeutic Area Lead at AstraZeneca, responsible for medical strategy for company products, including Dapagliflozin. He has received support from AstraZeneca to attend ERA 2023; EASD 2022, 2023, 2024; ESC 2025 and ESCMID 2022 as part of his job and participated as an organizer of AdBoards for different products in AstraZeneca JJA, WB, and WC don’t report any conflict of interest. VE is an expert advisor and speaker for Novo Nordisk and an expert advisor for some pharmaceutical companies namely Novo Nordisk and Astrazeneca and received support from Boehringer Ingelheim to attend ERA 2024. AML is the brand ambassador and is a Corporate medical manager for Pulso Salud Corporación Médica, FUNDING The project was funded by AstraZeneca. AUTHORS' CONTRIBUTIONS JM: Methodology design and supervision of information of data processing, AI design and training and model output AP: Methodology execution, including information processing, AI design, experimentation and training, model output, results, and interpretation DJ: Methodology execution, including information processing, AI design, experimentation and training, model output, results, and interpretation JZ: Project management and supervision IL: Scientific paper writing and editing, including abstract, introduction, methodology, results, and discussion NCV: In charge of scientific paper writing and editing, including abstract, introduction, methodology, results, and discussion DC: Medical expert, revisor, and editor JJA: Expert reviewer of final manuscript WB: Expert reviewer of the final manuscript and provided anonymized data for the study VE: Expert reviewer of the final manuscript and provided anonymized data for the study AML: Expert reviewer of the final manuscript and provided anonymized data for the study WC: Expert reviewer of the final manuscript and provided anonymized data for the study ACKNOWLEDGEMENTS We thank Somedyt, Medisinu, and Pulso Salud for providing the data necessary for the development of this project. We also thank AstraZeneca for their medical insight and collaboration. References GBD 2019 Mental Disorders Collaborators. Global, regional, and national burden of 12 mental disorders in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Psychiatry. 2022 Feb;9(2):137–50. Forecasting life expectancy, years of life lost, and all-cause and cause-specific mortality for 250 causes of death: reference and alternative scenarios for 2016–40 for 195 countries and territories - The Lancet [Internet]. [cited 2025 May 30]. Available from: https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(18)31694-5/fulltext Levey AS, Astor BC, Stevens LA, Coresh J. Chronic kidney disease, diabetes, and hypertension: what’s in a name? Kidney Int. 2010 Jul;78(1):19–22. Kovesdy CP. Epidemiology of chronic kidney disease: an update 2022. Kidney Int Suppl. 2022 Apr 1;12(1):7–11. Kalantar-Zadeh K, Jafar TH, Nitsch D, Neuen BL, Perkovic V. Chronic kidney disease. The Lancet. 2021 Aug;398(10302):786–802. INTERNATIONAL SOCIETY OF NEPHROLOGY. KDIGO Clinical Practice Guideline for Acute Kidney Injury [Internet]. INTERNATIONAL SOCIETY OF NEPHROLOGY; 2012. Available from: https://kdigo.org/wp-content/uploads/2016/10/KDIGO-2012-AKI-Guideline-English.pdf Cardiovascular Implications of the 2021 KDIGO Blood Pressure Guideline for Adults With Chronic Kidney Disease [Internet]. [cited 2025 Jul 7]. Available from: https://www.jacc.org/doi/epdf/10.1016/j.jacc.2022.02.040 Internationl Society of Nephrology. KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease [Internet]. 2024. Available from: https://www.researchgate.net/profile/Ifeoma-Ulasi/publication/379470748_KDIGO_2024_Clinical_Practice_Guideline_for_the_Evaluation_and_Management_of_Chronic_Kidney_Disease/links/660e35f0b839e05a20bd32b7/KDIGO-2024-Clinical-Practice-Guideline-for-the-Evaluation-and-Management-of-Chronic-Kidney-Disease.pdf Evans M, Lewis RD, Morgan AR, Whyte MB, Hanif W, Bain SC, et al. A Narrative Review of Chronic Kidney Disease in Clinical Practice: Current Challenges and Future Perspectives. Adv Ther. 2022 Jan 1;39(1):33–43. Echouffo-Tcheugui JB, Kengne AP. Risk Models to Predict Chronic Kidney Disease and Its Progression: A Systematic Review. PLOS Med. 2012 Nov 20;9(11):e1001344. Padmanabhan AA, Balczewski EA, Singh K. Artificial Intelligence Systems in CKD: Where Do We Stand and What Will the Future Bring? Adv Chronic Kidney Dis. 2022 Sep 1;29(5):461–4. Wu CC, Islam MM, Poly TN, Weng YC. Artificial Intelligence in Kidney Disease: A Comprehensive Study and Directions for Future Research. Diagnostics. 2024 Jan;14(4):397. Islam MdA, Majumder MdZH, Hussein MdA. Chronic kidney disease prediction based on machine learning algorithms. J Pathol Inform. 2023 Jan 12;14:100189. Isaza-Ruget MA, Yomayusa N, González CA, H CA, de Oro V FA, Cely A, et al. Predicting chronic kidney disease progression with artificial intelligence. BMC Nephrol. 2024 Apr 26;25(1):148. Tangri N, Peach EJ, Franzén S, Barone S, Kushner PR. Patient Management and Clinical Outcomes Associated with a Recorded Diagnosis of Stage 3 Chronic Kidney Disease: The REVEAL-CKD Study. Adv Ther. 2023 Jun;40(6):2869–85. Tangri N, Ferguson TW. Artificial Intelligence in the Identification, Management, and Follow-Up of CKD. Kidney360. 2022 Mar 31;3(3):554–6. Tangri N, Grams ME, Levey AS, Coresh J, Appel LJ, Astor BC, et al. Multinational Assessment of Accuracy of Equations for Predicting Risk of Kidney Failure: A Meta-analysis. JAMA. 2016 Jan 12;315(2):164–74. L. Rubini PS. Chronic Kidney Disease [Internet]. UCI Machine Learning Repository; 2015 [cited 2024 May 22]. Available from: https://archive.ics.uci.edu/dataset/336 Chan L, Nadkarni GN, Fleming F, McCullough JR, Connolly P, Mosoyan G, et al. Derivation and validation of a machine learning risk score using biomarker and electronic patient data to predict progression of diabetic kidney disease. Diabetologia. 2021 Jul;64(7):1504–15. Ravizza S, Huschto T, Adamov A, Böhm L, Büsser A, Flöther FF, et al. Predicting the early risk of chronic kidney disease in patients with diabetes using real-world data. Nat Med. 2019 Jan;25(1):57–9. Rashed-Al-Mahfuz Md, Haque A, Azad A, Alyami SA, Quinn JMW, Moni MA. Clinically Applicable Machine Learning Approaches to Identify Attributes of Chronic Kidney Disease (CKD) for Use in Low-Cost Diagnostic Screening. IEEE J Transl Eng Health Med. 2021;9:1–11. Swain D, Mehta U, Bhatt A, Patel H, Patel K, Mehta D, et al. A Robust Chronic Kidney Disease Classifier Using Machine Learning. Electronics. 2023 Jan;12(1):212. Chittora P, Chaurasia S, Chakrabarti P, Kumawat G, Chakrabarti T, Leonowicz Z, et al. Prediction of Chronic Kidney Disease - A Machine Learning Perspective. IEEE Access. 2021;9:17312–34. Stevens P, Ahmed S, Carrero J, Foster B, Francis A, Hall R, et al. KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease. Kidney Int. 2024 Apr 1;105:S117–314. scikit-learn.org. KNNImputer [Internet]. Available from: https://scikit-learn.org/1.5/modules/generated/sklearn.impute.KNNImputer.html imbalanced-learn.org. SMOTE [Internet]. [cited 2025 Feb 20]. Available from: https://imbalanced-learn.org/stable/references/generated/imblearn.over_sampling.SMOTE.html Kummer A, Ruppert T, Medvegy T, Abonyi J. Machine learning-based software sensors for machine state monitoring - The role of SMOTE-based data augmentation. Results Eng. 2022 Dec 1;16:100778. Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res. 2002 Jun 1;16:321–57. Udu AG, Salman MT, Ghalati MK, Lecchini-Visintini A, Siddle DR, Dong H. Emerging SMOTE and GAN Variants for Data Augmentation in Imbalance Machine Learning Tasks: A Review. IEEE Access. 2025;13:113838–53. Arkangel AI. Hippocrates [Internet]. Arkangel AI; [cited 2024 Dec 20]. Available from: https://arkprod.us.auth0.com/u/login/identifier?state=hKFo2SBTelNBSXBFaWJtLVhIcHU3R29ZRERWeHAyUkJSUDRQOaFur3VuaXZlcnNhbC1sb2dpbqN0a WTZIEtUMjRnSF9JTGNzTkc5eTNQY3hOZFFKX2w1eUUwRm5Qo2NpZNkgNzNXMFZVQWdzdzZLbGx0SXpsbUd3RUcwN0Nza0lRYjU Ying X. An Overview of Overfitting and its Solutions. J Phys Conf Ser. 2019 Feb 1;1168:022022. Mohammed S, Budach L, Feuerpfeil M, Ihde N, Nathansen A, Noack N, et al. The effects of data quality on machine learning performance on tabular data. Inf Syst. 2025 Jul 1;132:102549. Aburass S, Rumman MA. Quantifying Overfitting: Introducing the Overfitting Index. In: 2024 International Conference on Electrical, Computer and Energy Technologies (ICECET [Internet]. Sydney, Australia: IEEE; 2024 [cited 2025 Aug 29]. p. 1–7. Available from: https://ieeexplore.ieee.org/document/10698575/ Mitigating Overfitting in Decision Trees [Internet]. [cited 2025 May 24]. Available from: https://tilburgsciencehub.com/topics/analyze/machine-learning/supervised/decision_trees/ Calle-Toro J, Ordoñez C, Sánchez A, Sanjuan J, Badiel M, Pino L, et al. Epidemiologia de Lesiones Relacionadas con Colisiones de Vehiculos Motorizados en dos Centros de Referencia del Suroccidente Colombiano. Reporte del Registro Internacional de Trauma de la Sociedad P. Panam J Trauma Crit Care Emerg Surg. 2014 Apr 1;3:16–22. Alberdi F, García I, Atutxa L, Zabarte M. Epidemiología del trauma grave. Med Intensiva. 2014 Dec 1;38(9):580–8. Abdulsadig RS, Rodriguez-Villegas E. A comparative study in class imbalance mitigation when working with physiological signals. Front Digit Health. 2024 Mar 26;6:1377165. Chen W, Yang K, Yu Z, Shi Y, Chen CLP. A survey on imbalanced learning: latest research, applications and future directions. Artif Intell Rev. 2024 May 9;57(6):137. Ferreiro, Gonzalo. Medium. [cited 2024 Apr 17]. Class Imbalance: a classification headache | by Gonzalo Ferreiro Volpi | Towards Data Science. Available from: https://towardsdatascience.com/class-imbalance-a-classification-headache-1939297ff4a4 Thölke P, Mantilla-Ramos YJ, Abdelhedi H, Maschke C, Dehgan A, Harel Y, et al. Class imbalance should not throw you off balance: Choosing the right classifiers and performance metrics for brain decoding with imbalanced data. NeuroImage. 2023 Aug 15;277:120253. Research on expansion and classification of imbalanced data based on SMOTE algorithm | Scientific Reports [Internet]. [cited 2025 Aug 29]. Available from: https://www.nature.com/articles/s41598-021-03430-5 Palinkas LA, Horwitz SM, Green CA, Wisdom JP, Duan N, Hoagwood K. Purposeful sampling for qualitative data collection and analysis in mixed method implementation research. Adm Policy Ment Health. 2015 Sep;42(5):533–44. Jager J, Putnick DL, Bornstein MH. More than Just Convenient: The Scientific Merits of Homogeneous Convenience Samples. Monogr Soc Res Child Dev. 2017 Jun;82(2):13–30. Guest G, Namey EE, Mitchell ML. Sampling in Qualitative Research. In: Collecting Qualitative Data: A Field Manual for Applied Research [Internet]. SAGE Publications, Ltd; 2013 [cited 2025 Aug 29]. p. 41–74. Available from: https://methods.sagepub.com/book/mono/collecting-qualitative-data/chpt/2-sampling-qualitative-research Hennink M, Kaiser BN. Sample sizes for saturation in qualitative research: A systematic review of empirical tests. Soc Sci Med. 2022 Jan 1;292:114523. Perez N. n -Sizes, Attributes, and A Priori Sampling: A Qualitative Sampling Model for Large, Heterogeneous Populations. Am J Qual Res. 2024 Jul 26;8(3):193–207. Lewis DD, Catlett J. Heterogeneous Uncertainty Sampling for Supervised Learning. In: Cohen WW, Hirsh H, editors. Machine Learning Proceedings 1994 [Internet]. San Francisco (CA): Morgan Kaufmann; 1994 [cited 2025 Aug 29]. p. 148–56. Available from: https://www.sciencedirect.com/science/article/pii/B978155860335650026X Schinkel M, Bennis FC, Boerman AW, Wiersinga WJ, Nanayakkara PWB. Embracing cohort heterogeneity in clinical machine learning development: a step toward generalizable models. Sci Rep. 2023 May 24;13(1):8363. Zhao Q, Nooner KB, Tapert SF, Adeli E, Pohl KM, Kuceyeski A, et al. The Transition From Homogeneous to Heterogeneous Machine Learning in Neuropsychiatric Research. Biol Psychiatry Glob Open Sci. 2025 Jan;5(1):100397. Benito BM. Blas M. Benito, PhD. 2023 [cited 2025 Jul 21]. Multicollinearity Hinders Model Interpretability. Available from: https://blasbenito.com/post/multicollinearity-model-interpretability/ Detecting Multicollinearity Using Variance Inflation Factors | STAT 462 [Internet]. [cited 2025 Jul 21]. Available from: https://online.stat.psu.edu/stat462/node/180/ Akinwande MO, Dikko HG, Samson A. Variance Inflation Factor: As a Condition for the Inclusion of Suppressor Variable(s) in Regression Analysis. Open J Stat. 2015 Dec 11;5(7):754–67. Saslow E. Collinearity - What it means, Why its bad, and How does it affect other models? [Internet]. Future Vision. 2018 [cited 2025 Jul 21]. Available from: https://medium.com/future-vision/collinearity-what-it-means-why-its-bad-and-how-does-it-affect-other-models-94e1db984168 Multi Collinearity in the Tree Based Models | Kaggle [Internet]. [cited 2025 Jul 21]. Available from: https://www.kaggle.com/questions-and-answers/444752 Yıldırım H. The Multicollinearity Effect on the Performance of Machine Learning Algorithms: Case Examples in Healthcare Modelling. Acad Platf J Eng Smart Syst. 2024 Sep 25;12(3):68–80. Salih AM. Explainable Artificial Intelligence and Multicollinearity : A Mini Review of Current Approaches [Internet]. arXiv; 2024 [cited 2025 Jul 21]. Available from: http://arxiv.org/abs/2406.11524 Hamilton RI, Papadopoulos PN. Using SHAP Values and Machine Learning to Understand Trends in the Transient Stability Limit. IEEE Trans Power Syst. 2024 Jan;39(1):1384–97. Nohara Y, Matsumoto K, Soejima H, Nakashima N. Explanation of machine learning models using shapley additive explanation and application for real data in hospital. Comput Methods Programs Biomed. 2022 Feb 1;214:106584. Ponce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024;17(11):e70056. Deddens J, Petersen MR. Approaches for estimating prevalence ratios. Occup Environ Med. 2008 Aug 1;65:481, 501–6. Fajardo-Gutiérrez A. [Measurement in epidemiology: prevalence, incidence, risk, impact measures]. Rev Alerg Mex Tecamachalco Puebla Mex 1993. 2017;64(1):109–20. ClinConnect [Internet]. 2020 [cited 2025 Sep 17]. ClinConnect | iCaReMe Global Registry (NCT03549754). Available from: https://clinconnect.io/trials/NCT03549754?queryID=bb5320cb7c05993d60da&apos=20 iCaReMe Global Registry [Internet]. [cited 2025 Apr 4]. Available from: https://www.astrazenecaclinicaltrials.com/study/D1690R00044/ Ministerio de Salud de Colombia. Resolución número 8430 de 1993 (4 de octubre), por la cual se establecen las normas científicas, técnicas y administrativas para la investigación en salud [Internet]. Bogotá, Colombia: Diario Oficial No. 41.215; 1993 [cited 2025 Feb 19]. Available from: https://www.minsalud.gov.co/sites/rid/lists/bibliotecadigital/ride/de/dij/resolucion-8430-de-1993.pdf Asociación Médica Mundial. Declaración de Helsinki – Principios éticos para las investigaciones médicas en seres humanos [Internet]. Fortaleza, Brasil: Asociación Médica Mundial; 2013 [cited 2025 Feb 24]. Available from: https://www.wma.net/policies-post/wma-declaration-of-helsinki-ethical-principles-for-medical-research-involving-human-subjects/ Additional Declarations Competing interest reported. All affiliations have been disclosed. The authors are current employees of Arkangel AI (J.M., N.C.). J.Z. is one of the founders of Arkangel AI, and D.C. is an employee of AstraZeneca. The remaining authors have affiliations with the health institutions that provided the data. Supplementary Files AdditionalFile1StatisticalTestingandTables.docx ● Additional file 1. Statistical analyses and Tables referenced in the body of text. (.txt) AdditonalFile2ModelsbyArchitecture.xlsx ● Additional file 2. Compilation of all models trained (n:424), organized by architecture. (.xls) AdditionalFile3Modelselectionprocess.xlsx ● Additional file 3. Step-by-step (per sheet) model selection process. (.xls) AdditionalFile4IRBEthicsCommittee.zip ● Additional File 4. Compilation of each committee's approved research protocols and acceptance. (.zip) Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 07 Jan, 2026 Reviewers invited by journal 03 Nov, 2025 Editor assigned by journal 03 Nov, 2025 Editor invited by journal 23 Oct, 2025 Submission checks completed at journal 22 Oct, 2025 First submitted to journal 22 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7888843","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":539419435,"identity":"94e65508-f399-4415-83f4-29eafbe3e889","order_by":0,"name":"Julian Martinez","email":"","orcid":"","institution":"Arkangel AI","correspondingAuthor":false,"prefix":"","firstName":"Julian","middleName":"","lastName":"Martinez","suffix":""},{"id":539419438,"identity":"9b41685b-d47c-4888-836f-7c799e161447","order_by":1,"name":"Natalia Castano-Villegas","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6ElEQVRIiWNgGAWjYLCCBwYMcgwMbGC2DAMDYwMB9cwMDAkGDMYwLTxEamFgSGxAaCEAzNnPH/yQUGCX3j/7WOLngl82PAzsh/HbYtmTzCyRYJCcO+Nc2mHpmX1pPAw8ifi1GBxIZgBqYc5tOMPeIM3bc5gH7Ei8Ws4/Zv6RYFCfLn+Gvfk3b89/Hgb+hwS03EhmA9pyOMHgDNsxaZ4fB3gYJAjZcuOxmUWCwXHDjWfY0qx5G5J52CQI2XI+8fGND3+q5eXOsBnf5vljJ8fPn/4ArxZUwNgGi1HiwR8S1Y+CUTAKRsGIAAALo0OZ9dvkYQAAAABJRU5ErkJggg==","orcid":"","institution":"Arkangel AI","correspondingAuthor":true,"prefix":"","firstName":"Natalia","middleName":"","lastName":"Castano-Villegas","suffix":""},{"id":539419440,"identity":"c7292a73-36ad-425f-b0ef-08ba5df396c5","order_by":2,"name":"Alejandra Perez","email":"","orcid":"","institution":"Arkangel AI","correspondingAuthor":false,"prefix":"","firstName":"Alejandra","middleName":"","lastName":"Perez","suffix":""},{"id":539419441,"identity":"4d07a81c-0853-44ca-873c-6b31bccbcf5e","order_by":3,"name":"Daniel Jimenez","email":"","orcid":"","institution":"Arkangel AI","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"Jimenez","suffix":""},{"id":539419442,"identity":"8d73e5bd-4317-4e9e-b743-415a9f6bde0f","order_by":4,"name":"Jose Zea","email":"","orcid":"","institution":"Arkangel AI","correspondingAuthor":false,"prefix":"","firstName":"Jose","middleName":"","lastName":"Zea","suffix":""},{"id":539419443,"identity":"0122b8ff-77ee-44c8-8932-849f57de8b7b","order_by":5,"name":"Isabella Llano","email":"","orcid":"","institution":"Arkangel AI","correspondingAuthor":false,"prefix":"","firstName":"Isabella","middleName":"","lastName":"Llano","suffix":""},{"id":539419444,"identity":"a197344d-8b71-4db0-b48c-c63010d726e0","order_by":6,"name":"Diego Caro","email":"","orcid":"","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Diego","middleName":"","lastName":"Caro","suffix":""},{"id":539419446,"identity":"b692654d-d312-4594-b96f-8b4908d7d3b5","order_by":7,"name":"Jose Javier Arango","email":"","orcid":"","institution":"Universidad de Caldas","correspondingAuthor":false,"prefix":"","firstName":"Jose","middleName":"Javier","lastName":"Arango","suffix":""},{"id":539419448,"identity":"89fcf6e1-a115-41cd-afcf-4ae041619768","order_by":8,"name":"Walberto Buelvas","email":"","orcid":"","institution":"Medisinú","correspondingAuthor":false,"prefix":"","firstName":"Walberto","middleName":"","lastName":"Buelvas","suffix":""},{"id":539419451,"identity":"d9560fd5-1706-40a0-a16c-85040dd80b9e","order_by":9,"name":"Victor Espriella","email":"","orcid":"","institution":"Medisinú","correspondingAuthor":false,"prefix":"","firstName":"Victor","middleName":"","lastName":"Espriella","suffix":""},{"id":539419453,"identity":"c35ab3d4-7026-430f-97ff-e539f8697ef8","order_by":10,"name":"Ana María Llerena","email":"","orcid":"","institution":"Pulso Salud","correspondingAuthor":false,"prefix":"","firstName":"Ana","middleName":"María","lastName":"Llerena","suffix":""},{"id":539419456,"identity":"fb9a33a3-3745-4b6e-942d-c4c71905fa3a","order_by":11,"name":"William Castro","email":"","orcid":"","institution":"Somedyt","correspondingAuthor":false,"prefix":"","firstName":"William","middleName":"","lastName":"Castro","suffix":""}],"badges":[],"createdAt":"2025-10-17 17:38:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7888843/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7888843/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95844808,"identity":"6fd2159f-a029-4dcc-8d9a-1c53c44eb256","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":87264,"visible":true,"origin":"","legend":"","description":"","filename":"R2ManuscriptBMCNephro22.10.25.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/3a784f549a4f8734fe4e42c2.docx"},{"id":96240231,"identity":"18a36da3-34b4-4fb6-91d2-3fffa9a2889d","added_by":"auto","created_at":"2025-11-19 07:08:37","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":370985,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/ed2ce1397411e4b4cad09b77.docx"},{"id":96240320,"identity":"6d34964e-1cce-4f21-840b-3e0121e164b2","added_by":"auto","created_at":"2025-11-19 07:08:48","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":477823,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/2f607d784f8e10af94eb5c87.docx"},{"id":96240435,"identity":"0b1895b1-af7e-4f93-9085-a1d39d2077fa","added_by":"auto","created_at":"2025-11-19 07:08:54","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":273517,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/35c5fcb69c54dc082601c5eb.docx"},{"id":96240561,"identity":"53e3959a-31cf-4297-ba7c-b9fbf7f2f404","added_by":"auto","created_at":"2025-11-19 07:09:07","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":135863,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/42657b4710d44446b59a9a1b.docx"},{"id":95844820,"identity":"c5d42ea0-3b08-4ae7-973a-0795c3da1b51","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":133127,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/52620dee97f0aab86a0ad993.docx"},{"id":95844809,"identity":"fa7b34ac-b48c-45e3-8673-17f7ef11da35","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40846,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/6b7a613571f907402f1072bc.docx"},{"id":95844818,"identity":"ef7feff4-f70e-4c34-8c4e-3af1e5357974","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"docx","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":110597,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/32cb7fbc2bdd18cc78ec4d99.docx"},{"id":95844824,"identity":"1415c76d-6538-4f59-bbdf-eb69b45c9eca","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"json","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14249,"visible":true,"origin":"","legend":"","description":"","filename":"67633244d29e4fd8bfaf5f64ec6cf757.json","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/fee5043b418b734d887f225a.json"},{"id":96241688,"identity":"a2345374-3fda-47b2-b3d8-cd2e0d4e735f","added_by":"auto","created_at":"2025-11-19 07:11:17","extension":"docx","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3974620,"visible":true,"origin":"","legend":"","description":"","filename":"AdditionalFile1StatisticalTestingandTables.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/15d62657e3d63b56389454de.docx"},{"id":96240356,"identity":"4c337b39-ffa1-4d57-b5a1-529068887676","added_by":"auto","created_at":"2025-11-19 07:08:51","extension":"xlsx","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":76014,"visible":true,"origin":"","legend":"","description":"","filename":"AdditionalFile3Modelselectionprocess.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/09a0a864dd4fc8739958592c.xlsx"},{"id":95844838,"identity":"d362b6e0-e897-4ae1-a95c-f9d002559601","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"zip","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5102967,"visible":true,"origin":"","legend":"","description":"","filename":"AdditionalFile4IRBEthicsCommittee.zip","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/8ad7ddc78c58a6230017ba94.zip"},{"id":95844835,"identity":"d5218c38-10d5-44d5-984a-1c7aab06f8dc","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"xlsx","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":164257,"visible":true,"origin":"","legend":"","description":"","filename":"AdditonalFile2ModelsbyArchitecture.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/bdb08ab8ade67b0485e81461.xlsx"},{"id":96240817,"identity":"bbd86d5e-9b98-4f73-a5ee-a9a57736cdc7","added_by":"auto","created_at":"2025-11-19 07:09:36","extension":"xml","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":217310,"visible":true,"origin":"","legend":"","description":"","filename":"67633244d29e4fd8bfaf5f64ec6cf7571enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/3fb4b22a6689c8b44a937ca4.xml"},{"id":96240576,"identity":"49401eb7-e966-4159-bd87-dd5b4612aae5","added_by":"auto","created_at":"2025-11-19 07:09:08","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":356664,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/0ceb4698fc1a63ead76e2515.png"},{"id":95844821,"identity":"44a9fb32-38fb-4eec-9b31-6445d3345706","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":871291,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/ea7dc2a40166de7a18823da5.jpeg"},{"id":95844822,"identity":"f820a5ab-429b-4eb8-a728-bf589e275666","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"jpeg","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":335023,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/4dff36e45493866fce344748.jpeg"},{"id":95844825,"identity":"ee9ca963-941a-4a98-9087-760110c82e7f","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"jpeg","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":200846,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/1d4b18b261fd20a3de3a68f5.jpeg"},{"id":96240037,"identity":"4d698106-f28c-4f0f-8621-5dee9b7afdf0","added_by":"auto","created_at":"2025-11-19 07:08:17","extension":"jpeg","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":181318,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/04f69c8c3260da2ee6719ebc.jpeg"},{"id":95844829,"identity":"5b2bc765-28eb-411b-bc74-9418c43aed82","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"jpeg","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":120623,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/49112945d00e21f562815405.jpeg"},{"id":96241747,"identity":"f6f4f456-b3ef-4109-939b-61ddfa41ad90","added_by":"auto","created_at":"2025-11-19 07:11:18","extension":"jpeg","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":203951,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/1a89b35592dd1e0e76d51108.jpeg"},{"id":95844836,"identity":"9e18389b-5713-4a06-9e62-1e545a195c97","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":77379,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/d87d1393a31dcafddfaf13d1.png"},{"id":95844830,"identity":"08808eed-42ff-4d01-826f-6fb5603c9e0b","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":151040,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/cee55b4f4f42164103a33dd6.png"},{"id":95844831,"identity":"6daace70-d379-42a9-8efe-ccc69f4075fc","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64252,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/146d1b1bce81a6070bbeb5f2.png"},{"id":95844828,"identity":"a3a2aeff-3ccf-424c-8d98-d4717972d4b0","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":30753,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/98a603039490d520f2e00af1.png"},{"id":95844839,"identity":"cd8feb6c-a97b-43bd-8049-00cf91209312","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24364,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/e08340f2909e5895e1b7dfc2.png"},{"id":96239951,"identity":"57076899-a5c8-42e7-ae24-aee1b1932e19","added_by":"auto","created_at":"2025-11-19 07:08:03","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28414,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/2fcd0b504359de47a171412d.png"},{"id":96240312,"identity":"2f5dc5e4-1e6f-4e45-b970-01e24e17df45","added_by":"auto","created_at":"2025-11-19 07:08:47","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":31249,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/6a7feb6b416af4484f6a680b.png"},{"id":95844834,"identity":"e308d6e8-8d76-45f1-8667-bc7ae23629be","added_by":"auto","created_at":"2025-11-13 14:46:44","extension":"xml","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":213355,"visible":true,"origin":"","legend":"","description":"","filename":"67633244d29e4fd8bfaf5f64ec6cf7571structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/be271e933936646089ade933.xml"},{"id":96240492,"identity":"3a140557-36c3-4a9d-b87f-229b5da4e1e2","added_by":"auto","created_at":"2025-11-19 07:08:58","extension":"html","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":233469,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/0b614de476266c0029ee760a.html"},{"id":95844802,"identity":"d52d94b4-82cb-4287-808a-586a5ed51593","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":280996,"visible":true,"origin":"","legend":"\u003cp\u003ePopulation Flowchart\u003c/p\u003e\n\u003cp\u003e*CKD Positive is defined as an eGFR \u0026lt;60mL/min/1.73m2. CKD Negative is an eGFR \u0026gt;/=60mL/min/1.73m2. Both were calculated using the 2021 updated version of the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula. These values were used as the reference standard for comparing the AI model and were only used during the training phase of the algorithm development. None of them were used during the testing phase.\u003c/p\u003e\n\u003cp\u003eDesigned using Canva https://www.canva.com/design\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e Part a. The corresponding lines and boxes depict each step's exclusion and inclusion criteria and the imputation method. Part b. Absolute and relative frequencies of excluded and imputed population, Part c. Definite sample size and outcome distribution.\u003c/p\u003e","description":"","filename":"Binder11.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/7fffd3385c3128f430ead943.png"},{"id":95844803,"identity":"4bb39cea-7f0f-4ae6-a541-5e4afbf8629b","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":315961,"visible":true,"origin":"","legend":"\u003cp\u003eScreenshot of Additional File 2 displaying the sheet corresponding to the Decision Tree architecture*.\u003c/p\u003e\n\u003cp\u003e*Row 15, highlighted in yellow, presents some details of the final selected model.\u003c/p\u003e\n\u003cp\u003eDeveloped with Google Sheets\u003ca href=\"https://workspace.google.com/products/sheets/\"\u003e \u003c/a\u003e\u003ca href=\"https://docs.google.com/spreadsheets/d/16OAaJ5FbgINUhguW_DtNIjJlhFNltXExDcp6dXCGaXw/edit?gid=1799842340#gid=1799842340\"\u003ehttps://docs.google.com/spreadsheets/d/16OAaJ5FbgINUhguW_DtNIjJlhFNltXExDcp6dXCGaXw/edit?gid=1799842340#gid=1799842340\u003c/a\u003e\u003c/p\u003e","description":"","filename":"Binder12.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/449f5adccd94c15939729aa0.png"},{"id":96240700,"identity":"04433f90-e80b-4569-9d2f-c81f5bc4af86","added_by":"auto","created_at":"2025-11-19 07:09:24","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":116630,"visible":true,"origin":"","legend":"\u003cp\u003eScreenshot of the app’s interface when accessing a Project*.\u003c/p\u003e\n\u003cp\u003eTaken from \u003ca href=\"https://hippocrates.arkangel.ai/\"\u003ehttps://hippocrates.arkangel.a\u003c/a\u003e. All rights reserved.\u003c/p\u003e\n\u003cp\u003e*Project’s name in the figure: alldataSMOTEequal1erc, at the top left of the screen.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e The metrics can be organized from highest to lowest performance, choosing any metric from the dropdown menu.\u003c/p\u003e","description":"","filename":"Binder13.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/15495126d5ba6ae5deb385d8.png"},{"id":95844805,"identity":"1b8f4133-2d60-40d8-8dd4-0c8d6eb5283d","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":59715,"visible":true,"origin":"","legend":"\u003cp\u003eScreenshot of the information displayed on the App for every model trained or tested\u003c/p\u003e\n\u003cp\u003eTaken from \u003ca href=\"https://hippocrates.arkangel.ai/\"\u003ehttps://hippocrates.arkangel.a\u003c/a\u003e. All rights reserved.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend: \u003c/strong\u003ePerformance metrics are summarized in the screen's upper left, while the model’s technical characteristics are displayed in the upper right. The confusion matrix is represented in the lower right quadrant. The Shap values, which quantify the influence of each factor on the model's outcome, are also depicted (not shown in the Figure).\u003c/p\u003e","description":"","filename":"Binder14.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/3498db6f4f99d86b6f5f813d.png"},{"id":96241613,"identity":"4f25fabd-dedb-42bb-93c0-9f5ff785847b","added_by":"auto","created_at":"2025-11-19 07:11:08","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":50067,"visible":true,"origin":"","legend":"\u003cp\u003eUser interface for performing single predictions based on the algorithm selected and the patient’s variables\u003c/p\u003e\n\u003cp\u003eTaken from \u003ca href=\"https://hippocrates.arkangel.ai/\"\u003ehttps://hippocrates.arkangel.a\u003c/a\u003e. All rights reserved.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend:\u003c/strong\u003e Using patient information as input (right upper quadrant), the algorithm predicts his/her CKD risk. CKD risk is defined as an eGFR \u0026lt;60mL/min/1.73m2. Prediction (1= risk, 0= not at risk) is displayed on the left as the primary outcome, accompanied by the SHAP values for each variable.\u003c/p\u003e","description":"","filename":"Binder15.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/e3ac37c885ba3be04502df41.png"},{"id":95844806,"identity":"1722b667-2d3f-48ce-a889-f82e7918f529","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":32621,"visible":true,"origin":"","legend":"\u003cp\u003eShapley Additive Values for each attribute on the model\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend: \u003c/strong\u003eThe Shapley Additive Values are a measurement used in ML models to quantify each variable's weight in the model, on top of performance metrics. Despite its documented strong association, T2D appeared as the least influential variable for the algorithm's outcome. See the Discussion section for in-depth analysis.\u003c/p\u003e","description":"","filename":"Binder16.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/ae514f22f696114050143f0b.png"},{"id":95844816,"identity":"6340a716-9cdf-49fa-8d04-3763d1d75653","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":51098,"visible":true,"origin":"","legend":"\u003cp\u003eMaximum Depth of Individual Estimators\u003c/p\u003e\n\u003cp\u003eTaken from \u003ca href=\"https://hippocrates.arkangel.ai/\"\u003ehttps://hippocrates.arkangel.a\u003c/a\u003e. All rights reserved.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend: \u003c/strong\u003eThe chosen model, DecisionTreeC_max_depth-33_min_samples_split-2_RandomSearch, reveals clinically optimal performance despite showing traditional indicators of overfitting, such as the training curve surpassing the validation curve. The model achieved 99.7% accuracy in the training set and 91% in validation, presenting a gap of 9 percentage points between the two sets. The stability of validation performance, evidenced by the plateau reached from depth=29, and the absence of significant model degradation, suggests that the observed difference does not compromise clinical generalization. The conclusion is that the model does not present substantial overfitting and can be considered viable for implementation in medical screening settings, where practical utility and case identification capability outweigh strict criteria for statistical perfection.\u003c/p\u003e","description":"","filename":"Binder17.png","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/0f25a36259200fc3414680c6.png"},{"id":96362727,"identity":"481f3794-78dd-4e3f-8124-0dee01577e04","added_by":"auto","created_at":"2025-11-20 09:46:54","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2924593,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/9e49c978-a2a4-4b3c-be86-79062c76cd21.pdf"},{"id":96240343,"identity":"b70e60fb-e562-46c4-9d85-51e29e80d3a3","added_by":"auto","created_at":"2025-11-19 07:08:50","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":3974620,"visible":true,"origin":"","legend":"\u003cp\u003e● Additional file 1. Statistical analyses and Tables referenced in the body of text. (.txt)\u003c/p\u003e","description":"","filename":"AdditionalFile1StatisticalTestingandTables.docx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/06e48193e44925b1c8833464.docx"},{"id":96240051,"identity":"b4078bef-d38d-4512-a41f-e9f0eca5f884","added_by":"auto","created_at":"2025-11-19 07:08:18","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":164257,"visible":true,"origin":"","legend":"\u003cp\u003e● Additional file 2. Compilation of all models trained (n:424), organized by architecture. (.xls)\u003c/p\u003e","description":"","filename":"AdditonalFile2ModelsbyArchitecture.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/8342d5899cf2d434f5cabefd.xlsx"},{"id":96239863,"identity":"9a8fa071-8990-421a-b457-0cfb2576d146","added_by":"auto","created_at":"2025-11-19 07:07:50","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":76014,"visible":true,"origin":"","legend":"\u003cp\u003e● Additional file 3. Step-by-step (per sheet) model selection process. (.xls)\u003c/p\u003e","description":"","filename":"AdditionalFile3Modelselectionprocess.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/557abf913590609d5b947eb1.xlsx"},{"id":95844814,"identity":"15c8d0f3-91ac-40a7-a748-266a2422495f","added_by":"auto","created_at":"2025-11-13 14:46:43","extension":"zip","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":5102967,"visible":true,"origin":"","legend":"\u003cp\u003e● Additional File 4. Compilation of each committee's approved research protocols and acceptance. (.zip)\u003c/p\u003e","description":"","filename":"AdditionalFile4IRBEthicsCommittee.zip","url":"https://assets-eu.researchsquare.com/files/rs-7888843/v1/06113dcdf2e9472c71e0efa7.zip"}],"financialInterests":"Competing interest reported. All affiliations have been disclosed. The authors are current employees of Arkangel AI (J.M., N.C.). J.Z. is one of the founders of Arkangel AI, and D.C. is an employee of AstraZeneca. The remaining authors have affiliations with the health institutions that provided the data.","formattedTitle":"Search AI: a Machine Learning algorithm for chronic kidney disease risk detection using eight readily available clinical features","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eChronic kidney disease (CKD) is one of the leading causes of worldwide mortality (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). It affects over 10% of the adult population, especially people with Type 2 Diabetes (T2D) and hypertension (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e), which contributes to its high morbidity and mortality, ultimately resulting in the need for dialysis or transplantation due to loss of kidney function. Early identification is vital for appropriate treatment and to slow organ deterioration (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e), as pharmacologic and non-pharmacologic interventions to preserve kidney function have been demonstrated to improve prognosis (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eDiagnosis relies on the creatinine-based estimated glomerular filtration rate (eGFRcr)(\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). If available, the KDIGO Clinical Practice Guideline (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e) recommends cystatin C in the diagnostic calculation of eGFR (eGFRcr-cys). Considering cystatin C\u0026rsquo;s limited availability in low- and middle-income countries (LMIC) and the relative novelty of the update, the abbreviation eGFR in this manuscript will refer exclusively to eGFRcr. Diagnosis is confirmed when an eGFR\u0026thinsp;\u0026lt;\u0026thinsp;60mL/min/1.73m2 persists for over three months or if a urinary albumin-creatinine ratio (UACR) measurement is \u0026ge;\u0026thinsp;30mg/g. However, most affected individuals are asymptomatic until eGFR is under \u0026le;\u0026thinsp;30mL/min/1.73m2 (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe risk factors for the presentation and progression of CKD have been extensively described elsewhere (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e), and several diagnostic and prognostic algorithms have been developed. However, the lack of integration of validated models into electronic health records (EHRs) curtails their generalization. Their use is often limited to individual physician criteria in online CKD risk calculators (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). Furthermore, the complexity and numerosity of attributes used in most clinical algorithms make it unfeasible to apply them in low-resource settings due to their unavailability in EHRs and healthcare systems.\u003c/p\u003e\u003cp\u003eArtificial Intelligence (AI) has emerged as a valuable tool for the early detection of CKD. Machine learning (ML) algorithms have shown promise in identifying at-risk patients from large volumes of data, facilitating timely intervention, and improving patient outcomes (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). Although some clinicians are still sceptical about AI models' genesis, performance, and reliability, these models have proven to be as good as traditional biostatistical methods, with the advantage of processing complex and numerous data accounts faster and more precisely (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Consequently, they are being used to increase the efficiency of the screening and diagnosis processes and provide accessible risk stratification for individuals who would remain undetected until advanced stages of CKD otherwise (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eIn 2011, Tangri et al. created the clinical algorithm Kidney Failure Risk Equation (KFRE) for CKD progression probability after stage 3. It has demonstrated high discrimination and good calibration and is a reference in the field (\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). Most research using AI algorithms has followed this general structure, training on public clinical databases such as the UCI ML repository (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e) with anywhere from 24 to 100 variables. These complex models have achieved AUCs between 60\u0026ndash;70% (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e)\u003c/p\u003e\u003cp\u003eAlthough some of these models demonstrated excellent performances (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e), they rely upon a battery of specialised laboratory tests from secondary information sources, often costly, and usually incomplete. Considering the need for a more straightforward screening process and the resource limitations in LMIC (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e), we aimed at developing a machine learning algorithm capable of mass identification of patients at risk of CKD using simpler, less invasive variables, based on eGFR early alterations.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance metrics for some of the CKD ML algorithms revised.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"10\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e# of variables\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModel type\u003c/p\u003e\u003cp\u003e(architecture)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eData Source\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003ePrecision (PPV)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eF1 score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eArkangel AI (Ours)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDecision Tree\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMulticentric dataset from latinamerican health institutions\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.902\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.927\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.890\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.917\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.896\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eRavizza et al.\u003c/b\u003e (\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLogistic Regression\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eIBM Explorys database\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.794\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.682\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.726\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.217\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.721\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.329\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eKidneyIntelX\u003c/b\u003e (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRandom Forest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eBioMe Biobank at the Icahn School of Medicine at Mount Sinai and the Penn Medicine Biobank (PMBB)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.770\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eQin et al.\u003c/b\u003e (\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRandom Forest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eUCI machine learning repository\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.998\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eIslam et al.\u003c/b\u003e (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eUCI machine learning repository\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.980\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.980\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.983\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.980\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eJeong et al.\u003c/b\u003e (\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAutoencoder\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eHealth examination cohort database provided by the National Health Insurance Service (NHIS) in Korea\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.956\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.959\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.956\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eKumar et al. (66)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eGenetic Programming\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eUCI machine learning repository\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.990\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.998\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWeber et al. (67)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e11\u0026ndash;19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eArtificial Neural Network\u003c/p\u003e\u003cp\u003e(ANN)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRetrospective dataset derived from the Jena Part of the 3000 PA text corpus of the Smart Medical Information Technology for Healthcare (SMITH) consortium\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.910\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1.000\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.820\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Performance metrics, number of features, and sources used for training of some of the CKD ML algorithms revised.\u003c/p\u003e"},{"header":"MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003ePopulation and variables:\u003c/h2\u003e\n \u003cp\u003eOver 600,000 records were collected from three Latin American health institutions: Databases 1 and 2 (D1, D2) from Colombia and Database 3 (D3) from Peru (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e). The independent variables in the pooled data set included eight common features readily available in face-to-face patient consultation and used for model training. They were Age, Sex, Hypertension, Systolic and Diastolic Blood Pressure (SBP-DBP), Body mass Index (BMI), Diabetes (T2D), and Diabetes Duration (T2DD) (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). These variables were selected based on expert opinion, due to their availability and strong associations with CKD, summarized in Additional File 1a. They also constitute basic clinical information, which is commonly recorded in routine medical checkups, especially for follow-up in chronic disease. They are also non-invasive and more likely to be registered in clinical records than more sophisticated tests or laboratory results.\u003c/p\u003e\n\u003c/div\u003e\n\u003ch3\u003eData Preprocessing:\u003c/h3\u003e\n\u003cp\u003eMissing values were imputed before merging as follows (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, part a):\u003c/p\u003e\n\u003cp\u003e● Database 1: Hypertension was missing in 172 records, or 0.89% of the Database. It was imputed using the mode.\u003c/p\u003e\n\u003cp\u003e● Database 2: the Duration of Diabetes was missing in 13,513 records, corresponding to 21,6% of the dataset. We used a Machine Learning (ML) method, the k-Nearest Neighbour imputer. This method completes missing values using k-Nearest Neighbour (\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e). Each sample\u0026rsquo;s missing values are imputed using the mean value from the n nearest neighbour found. In this case, we used the default value for n\u0026thinsp;=\u0026thinsp;5.\u003c/p\u003e\n\u003cp\u003e● Database 3: The duration of diabetes was missing from 133 records, corresponding to 0.11% of the data. Here, we used a traditional imputation method, the mean.\u003c/p\u003e\n\u003cp\u003eThe eight independent variables were used in their original form. For quantitative variables, values outside physiologically plausible ranges were excluded (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). The ninth variable in the Table, SCr, was only used during the training phase. Its use will be discussed in detail in the following sections.\u003c/p\u003e\n\u003ch3\u003eSample Size\u003c/h3\u003e\n\u003cp\u003eThe sample size was selected from the pooled dataset and based on convenience, according to available records, and the following selection criteria:\u003c/p\u003e\n\u003cp\u003eInclusion:\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e1. Records from individuals older than 18 years.\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e2. Records with at least one registered serum creatinine (SCr) result, in mg/dl.\u003c/p\u003e\n\u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e3. Records from patients with or without chronic conditions (CKD, Hypertension, T2D).\u003c/p\u003e\n\u003c/span\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eExclusion:\u003c/p\u003e\n\u003cp\u003e\u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e1. No established exclusion criteria were applied after the initial exclusion of unfeasible values for quantitative variables.\u003c/p\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eVariable Distribution\u003c/h3\u003e\n\u003cp\u003eQuantitative variables were described using the mean and standard deviation (SD), and qualitative variables were described using absolute and relative frequencies. Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e summarizes the distribution of characteristics in each database and in total.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eVariable descriptions.\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFeature\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePossible Ranges\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eUnit\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNature\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOrder\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0-100\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQuantitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eSex\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e1\u0026thinsp;\u0026minus;\u0026thinsp;0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMan/Woman\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQualitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eHypertension\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e1\u0026thinsp;\u0026minus;\u0026thinsp;0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes/No\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQualitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eSystolic Blood Pressure (SBP)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e40\u0026ndash;200\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003emmHg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQuantitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eDiastolic Blood Pressure (DBP)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e40\u0026ndash;120\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003emmHg\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQuantitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eBody Mass Index (BMI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e12\u0026ndash;50\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eKg/m\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQuantitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eDiabetes (T2D)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e1\u0026thinsp;\u0026minus;\u0026thinsp;0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eYes/No\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQualitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eDiabetes Duration (T2DD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0\u0026ndash;60\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNumber of years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQualitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eSerum Creatinine (SCr)*\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003e0\u0026ndash;20\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003emg/dl\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQuantitative\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003cdiv align=\"left\" class=\"colspec\"\u003e\u003cbr\u003e\u003c/div\u003e\n \u003cdiv align=\"left\" class=\"colspec\"\u003e*SCr was used ONLY during training to calculate the eGFR primary outcome, with the 2021 CKD-EPI equation. \u0026nbsp;\u003c/div\u003e\n \u003ctable id=\"Tab3\" border=\"1\" style=\"margin-right: calc(0%); width: 100%;\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eBasal Distribution of Clinical and Demographic Characteristics (before augmentation)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" rowspan=\"2\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eNumber of registries\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003eDatabase 1\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003eDatabase 2\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003eDatabase 3\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003eN\u0026thinsp;=\u0026thinsp;203,067\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003en\u0026thinsp;=\u0026thinsp;19,276\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003en\u0026thinsp;=\u0026thinsp;62,553\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\" style=\"width: 15.2453%;\"\u003e\n \u003cp\u003en\u0026thinsp;=\u0026thinsp;121,238\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eQuantitative Variables\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 4.21%;\"\u003e\n \u003cp\u003emean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 3.5721%;\"\u003e\n \u003cp\u003estd\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003emin, max\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003emean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003estd\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003emin, max\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003emean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003estd\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003emin, max\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003emean\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003estd\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003emin, max\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eAge\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.21%;\"\u003e\n \u003cp\u003e47.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5721%;\"\u003e\n \u003cp\u003e17.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e18, 100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e65.23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e11.86\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e21, 99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e64.64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e13.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e18, 100\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e36.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e13.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e18, 100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eSBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.21%;\"\u003e\n \u003cp\u003e115.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5721%;\"\u003e\n \u003cp\u003e14.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e40, 180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e129.69\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e15.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e40, 180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e122.61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e15.57\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e40, 180\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e109.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e9.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e60, 180\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eDBP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.21%;\"\u003e\n \u003cp\u003e73.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5721%;\"\u003e\n \u003cp\u003e9.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e40,12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e77.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e8.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e40, 120\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e78.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e10.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e40, 120\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e70.11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e7.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e50, 100\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eBMI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.21%;\"\u003e\n \u003cp\u003e26.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5721%;\"\u003e\n \u003cp\u003e4.33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e12, 50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e27.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e5.26\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e12.42, 49.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e26.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e5.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e12, 50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e26.92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e3.62\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e13.92, 49.67\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eDiabetes Duration (T2DD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.21%;\"\u003e\n \u003cp\u003e1.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5721%;\"\u003e\n \u003cp\u003e4.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e0, 56.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e9.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e7.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e0, 56.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e2.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e3.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e0, 9.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 4.2738%;\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 3.5083%;\"\u003e\n \u003cp\u003e0.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e0, 31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eQualitative Variables\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbsolute Frequency (n)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRelative Frequency (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbsolute Frequency (n)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRelative Frequency (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbsolute Frequency (n)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRelative Frequency (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbsolute Frequency (n)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRelative Frequency (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eSex (F)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e66,863\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e32.93%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e14,32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e74.29%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e40,043\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e64.01%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e12,5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e10.31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eHypertension (Yes)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e53,864\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e26.53%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e15,855\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e82.25%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e36,995\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e59.14%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e1,014\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e0.84\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eDiabetes (Yes)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e32,909\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e16.21%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e18,922\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e98.16%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e13,513\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e21.60%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e474\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e0.39\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" style=\"width: 8.2287%;\"\u003e\n \u003cp\u003eCKD (Yes)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e27,502\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e13.54%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e8,622\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e44.73%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e18,453\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.4632%;\"\u003e\n \u003cp\u003e29.50%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.7821%;\"\u003e\n \u003cp\u003e427\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 7.527%;\"\u003e\n \u003cp\u003e0.35\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003ch3\u003eOutcome\u003c/h3\u003e\n\u003cp\u003eThe dependent variable or primary outcome was the estimated Glomerular Filtration Rate (eGFR). Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e shows that CKD was a variable in the original database. We decided not to use this specific classification as the standard of reference for model training because we could not determine which criteria or formula was used to define it.\u003c/p\u003e\n\u003cp\u003eInstead, we followed the 2024 KDIGO Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease (KDIGO 24) (\u003cspan class=\"CitationRef\"\u003e8\u003c/span\u003e). We took the SCr measurement in the records to calculate eGFR using the 2021 updated version of the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula to standardize the measurement (Additional File 1b).\u003c/p\u003e\n\u003cp\u003eAs per the guideline, we defined a threshold of eGFR equal to 60mL/min/1.73m2 to separate individuals into two categories: on the one hand, those with risk of presenting with CKD when eGFR was \u0026lt;\u0026thinsp;60mL/min/1.73m2 (defined in Chap. \u0026nbsp;2 of the guideline as G3a - Mildly to moderately increased CKD risk). These individuals would undergo medical follow-up according to guidelines, including confirmatory tests (a second eGFR, albuminuria), and specialized kidney function evaluation. On the other hand, an eGFR\u0026thinsp;\u0026ge;\u0026thinsp;60mL/min/1.73m2 would classify the record as belonging to someone who did not need to be prioritized for immediate further testing (although it does not mean zero risk, and the guideline recommends UACR).\u003c/p\u003e\n\u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. \u003cstrong\u003eBasal Distribution of Clinical and Demographic Characteristics (before augmentation)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. Population Flowchart\u003c/p\u003e\n\u003cp\u003e*CKD Positive is defined as an eGFR\u0026thinsp;\u0026lt;\u0026thinsp;60mL/min/1.73m2. CKD Negative is an eGFR \u0026gt;/=60mL/min/1.73m2. Both were calculated using the 2021 updated version of the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula. These values were used as the reference standard for comparing the AI model and were only used during the training phase of the algorithm development. None of them were used during the testing phase.\u003c/p\u003e\n\u003cp\u003eDesigned using Canva \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.canva.com/design\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLegend\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePart a. The corresponding lines and boxes depict each step\u0026apos;s exclusion and inclusion criteria and the imputation method. Part b. Absolute and relative frequencies of excluded and imputed population, Part c. Definite sample size and outcome distribution.\u003c/p\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003eAI Methods\u003c/h2\u003e\n \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\n \u003ch2\u003eData Augmentation\u003c/h2\u003e\n \u003cp\u003eThe imbalance between positive and negative classes is common in ML algorithms. To better represent all classes (positive and negative CKD risk), we used an augmentation called Synthetic Minority Over-Sampling Technique (SMOTE) (\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e). When using SMOTE, the researcher determines a desired ratio of the samples in the minority class to the number of samples in the majority class after resampling to make the classes more equal in total numbers. Other authors have described the SMOTE augmentation technique as a valid strategy for unbalanced sample sizes in ML model training, significantly improving the classification\u0026apos;s balance and accuracy (\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eThe augmentation process was performed as follows:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Total sample size, before augmentation\u0026thinsp;=\u0026thinsp;N 203067 (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, part a).\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e2. CKD positive class before augmentation\u0026thinsp;=\u0026thinsp;n 27502 (see Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e3. CKD negative class before augmentation\u0026thinsp;=\u0026thinsp;n 175565\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e4. Using SMOTE 1 (desired ratio of 1 between classes), the CKD positive class was augmented until reaching the absolute number of the negative class, resulting in a positive class of n\u0026thinsp;=\u0026thinsp;175565 after augmentation (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e, part c)\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e5. Conversely, the CKD negative class was not modified after augmentation using SMOTE 1 (n 175565).\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e6. The absolute number of augmented data was =\u0026thinsp;n 148063.\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e7. Therefore, the new total augmented sample was =\u0026thinsp;N 351130\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e8. After augmentation techniques using SMOTE 1, 148063 synthetic records were created. These represent 42.2% of the total sample.\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eThen, the training and following steps were performed using the augmented sample size (feature distribution in the augmented sample is exhibited in Additional File 1c).\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003ch3\u003eTraining and Testing Phases\u003c/h3\u003e\n\u003cp\u003eThe construction of an AI algorithm can be separated into two main parts: the training and testing phases. The total sample after augmentation was 351130. We split it into 70% for training, 12% for validation, and 18% for testing (Table \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e\n\u003cp\u003eDuring training, the model learns to recognize patterns. The validation set evaluates performance as the model is trained; it is considered part of the training set (consequently, the total split for the training phase was 82%). In the testing phase, the model is used to perform outcome predictions on a portion of the sample size that it has never been in contact with, which allows for the internal validation of the model.\u003c/p\u003e\n\u003cp\u003eThe training phase was performed using the supervised learning approach, in which the human programmer trains the algorithms by showing them examples marked as \u0026ldquo;positive\u0026rdquo; and \u0026ldquo;negative\u0026rdquo; or \u0026ldquo;healthy\u0026rdquo; and \u0026ldquo;ill,\u0026rdquo; etc., according to the outcome\u0026rsquo;s nature. The variables used to train the algorithms included the eight non-invasive, readily available features listed in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. The ninth, SCr, calculates baseline eGFR using the standardized 2021 CKD EPI formula.\u003c/p\u003e\n\u003cp\u003eSerum creatinine was exclusively present during the model\u0026apos;s supervised learning training. The ML\u0026apos;s automatic learning capacity uses the information of the estimated GFR to assimilate the positive and negative case criteria. Then, the model can acknowledge patterns and recognize associations among the other eight variables that are not evident to the naked eye. Consequently, it determines how different combinations or values of these features are associated with an altered eGFR when no SCr or eGFR are provided.\u003c/p\u003e\n\u003cp\u003eThis capability is assessed during the testing phase. Here, the model is evaluated using a portion of the database previously unknown to the algorithm. In this stage, the model is given records without any data on eGFR or any other information directly related to renal function measurement. It then performs the predictions or inferences, using the other eight features. As the outcome, it yields the probability of an eGFR being higher, equal, or lower than 60mL/min/1.73m2.\u003c/p\u003e\n\u003cp\u003eHowever, its benefit is not simply estimating an eGFR. Recognizing these association patterns among variables determines how different combinations or values of the features are associated with an altered eGFR (\u0026lt;\u0026thinsp;60mL/min/1.73m\u003csup\u003e2\u003c/sup\u003e) as an outcome, and how much each of the variables in the model influences it (see SHAP values ahead).\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eDistribution of data splits\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePhase\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAbsolute number (n)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePercentage (%)\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eTrain\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e245791\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e70\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eValidate\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e42136\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e12\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eTest\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e63203\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e18\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e351130\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u003cstrong\u003e100\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003eArkangel AI platform\u003c/h2\u003e\n \u003cp\u003eArkangel AI is an Auto ML tool designed to train AI models based on clinical data. It was built to simplify the human programmer\u0026apos;s training and experimentation process by automatically training and evaluating multiple candidate algorithms based on the variables provided (\u003cspan class=\"CitationRef\"\u003e30\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eThe models are trained exclusively by ML engineers with specific access credentials. Based on the datasets, they perform automatic but exhaustive training experiments using many combinations of architectures and hyperparameters to find the best-performing models. The app trains around 25 models per experiment. It is also set to automatically use cross-validation to assess the model and hold out a portion of the dataset to estimate performance and compute all relevant metrics.\u003c/p\u003e\n \u003cp\u003eAccording to these cross-validated metrics, models can be organized from highest to lowest performance. The results stem from multiple automated iterations executed by the app. Each iteration presents a confidence score, which estimates how confident the model is of its result. Finally, the platform automatically organises the algorithms and recommends the best-performing models based on their performance metrics (e.g., sensitivity, specificity, or AUC) (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\n \u003cp\u003eThe architecture search includes KNeighborsClassifier (KNC), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Gradient Boosting (GB), Random Forest Classifier (RFC), and deep learning (DL) models like Deep Fully Connected Neural Networks (DNN). Finally, these models are available so users can observe and assess their composition and performance through the user-friendly interface.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003eAI Model Training\u003c/h2\u003e\n \u003cp\u003eOur training was conducted on Arkangel\u0026rsquo;s platform. We performed several experiments. We trained models of several architectures, like DNNs, RFs, and KNCs, and explored different hyperparameters. Additional File 2 displays the extensive list of algorithms trained for this study, organized in seven sheets (using Google Sheets) by architecture, demonstrating their technical characteristics and hyperparameters. Figure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e exemplifies the Sheet file with a screenshot.\u003c/p\u003e\n \u003cp\u003eThese architectures were also trained by exploring different sample size distributions in the search for the best-performing models. For example, we examined training with the Colombian registries (Col) to test them in the Peruvian dataset. We also stratified the sample into Diabetic (T2D) and non-diabetic patient records (non-T2D), on top of the usual approach of using the sample size (AllData or Merged). Additionally, we experimented with SMOTE using ratios of 0.5 and 1.\u003c/p\u003e\n \u003cp\u003eAll the resulting models constitute Additional File 2. The projects and models were named using the terms in parentheses above, according to the sample distribution they were trained in, the hyperparameters, and the augmentation technique used (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, columns A and B).\u003c/p\u003e\n \u003cp\u003eThe file also contains each algorithm\u0026apos;s performance: accuracy, sensitivity (or recall), specificity, positive predictive value (PPV) (or precision), Area Under the Curve (AUC), and F1. Table \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e summarizes the algorithms by architecture. Table \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e displays the number of experiments or projects explored and the number of models trained on each experiment.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab5\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eMachine Learning Model Distribution by Experiment/Project Name\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eExperiment/ Project Name\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNumber of models\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAllDataAugmentedDownsampledErc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAllDataSmoteEqual0Point5Erc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAllDataSmoteEqual1Erc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCOLAugmentedpositivesx2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eColDataAugmentedDownsampledErc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eColDataImputedSmoteEqual0Point5Erc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eColDataSmoteEqual1Erc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDbSmoteEqual0Point5Erc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDbSmoteEqual1Erc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e66\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDiabeticosImputadoMerged\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMergedColombiaImputed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003emergedColombiaImputedUpsampled\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMergedColombiaImputedUpsampledDownsampled\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMergedImputed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMergedImputedUpsampled\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMergedImputedUpsampledDownsampled\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNdbSmoteEqual1Erc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003enodiabeticosMerged\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e21\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTotal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e424\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e. Screenshot of Additional File 2 displaying the sheet corresponding to the Decision Tree architecture*.\u003c/p\u003e\n \u003cp\u003e*Row 15, highlighted in yellow, presents some details of the final selected model.\u003c/p\u003e\n \u003cp\u003eDeveloped with Google Sheets \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://docs.google.com/spreadsheets/d/16OAaJ5FbgINUhguW_DtNIjJlhFNltXExDcp6dXCGaXw/edit?gid=1799842340#gid=1799842340\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003eAI Model Prioritization\u003c/h2\u003e\n \u003cp\u003eThe trained models are stored in the App and grouped by project name (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e). Because this is a screening use case, sensitivity (true-positive rate) was the primary selection criterion for prioritizing the models. At the same time, other metrics were required to remain within predefined acceptable ranges. They are stated below. Figure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e illustrates how models can be organized from higher to lower sensitivity or the user\u0026rsquo;s preferred metric. Figure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e shows how the results of each algorithm are presented in the App.\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e. Screenshot of the app\u0026rsquo;s interface when accessing a Project*.\u003c/p\u003e\n \u003cp\u003eTaken from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://hippocrates.arkangel.a\u003c/span\u003e\u003c/span\u003e. All rights reserved.\u003c/p\u003e\n \u003cp\u003e*Project\u0026rsquo;s name in the figure: alldataSMOTEequal1erc, at the top left of the screen.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eLegend\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eThe metrics can be organized from highest to lowest performance, choosing any metric from the dropdown menu.\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. Screenshot of the information displayed on the App for every model trained or tested\u003c/p\u003e\n \u003cp\u003eTaken from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://hippocrates.arkangel.a\u003c/span\u003e\u003c/span\u003e. All rights reserved.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eLegend\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003ePerformance metrics are summarized in the screen\u0026apos;s upper left, while the model\u0026rsquo;s technical characteristics are displayed in the upper right. The confusion matrix is represented in the lower right quadrant. The Shap values, which quantify the influence of each factor on the model\u0026apos;s outcome, are also depicted (not shown in the Figure).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003eAI algorithm selection\u003c/h2\u003e\n \u003cp\u003eOnce all projects were finalized, we evaluated and organized them from highest to lowest sensitivity. We grouped them by project, as can be replicated in the second sheet inside Additional File 2. We did not use the grouping by architecture because not all of them produced algorithms that were competitive in performance.\u003c/p\u003e\n \u003cp\u003eThen, we selected the top-performing models following predefined rules, agreed upon by the research and medical team, as follows:\u003c/p\u003e\u003cspan\u003e\n \u003cp\u003e1. Select the models with a sensitivity/recall higher than 90%\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e2. From those, the ones with an AUC higher than 90%\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e3. Then, those with a PPV/precision higher than 90%\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e4. From those, the ones with specificity higher than 85%\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e5. F1 higher than 90%\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e6. If the best models belonged to those trained in the stratified sample size, from both T2D and non-T2D, then an ensemble model of the two strata would be explored.\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e7. If a model trained in the Colombian population were classified as top-performing, its further assessment would depend on how it performed using the Peruvian sample.\u003c/p\u003e\n \u003c/span\u003e\u003cspan\u003e\n \u003cp\u003e8. Models trained using the entire merged sample size would be preferred if the other best-performing models were those trained separately in the T2D or the non-T2D stratified samples.\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003eCriteria six to eight were agreed upon in case several algorithms fulfilled the first five criteria.\u003c/p\u003e\n \u003cp\u003eBy applying the criterion, we ensured a highly sensitive model with optimal discrimination ability (AUC) and capacity to predict a positive case (PPV) despite lower specificity. An algorithm with satisfactory, balanced global metrics (F1). The final step of this process is depicted in Table \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e. The detailed steps are displayed on each sheet of Additional File 3.\u003c/p\u003e\n \u003cp\u003eThe chosen algorithm is presented using simple tables and figures: technical characteristics (Fig. \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e), performance metrics, and a confusion matrix to calculate them (Table \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e), and the Shapley Additive Values (Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e), which quantify the weight of each variable in the algorithm\u0026rsquo;s inference, supporting model explainability. Figures \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e to \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e display screenshots of how results are presented in the Arkangel AI web app.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003eModel Validation\u003c/h2\u003e\n \u003cp\u003eThe Arkangel app uses iterative cross-validation to automatically evaluate the model\u0026apos;s results; it is a functionality embedded in its algorithm. Sensitivity analysis is further assessed by appraising the model\u0026apos;s behaviour on the records of diabetic and non-diabetic individuals in the test set (Additional File 1m-1o). Additionally, we performed stratified analyses by gender and age group, framing it into three segments: younger than forty, forty to sixty, and older than sixty years old (Additional File 1p-1r). We also present the 95% Confidence Intervals (CI 95%) for performance metrics.\u003c/p\u003e\n \u003cp\u003eIn addition to AI and ML analyses, we undertook traditional statistical analyses. First, we conducted bivariate hypothesis tests to obtain p-values to explore statistical differences (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Then we used regression models to evaluate associations, model variation explainability, and to control for confusion. The prevalence ratio was also determined to explore associations not evident in the AI model.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003eOverfitting Control\u003c/h2\u003e\n \u003cp\u003ePossible overfitting was assessed on two fronts. The first was comparing the training and testing performance metrics and their respective confusion matrices (Tables \u003cspan class=\"InternalRef\"\u003e8\u003c/span\u003e to \u003cspan class=\"InternalRef\"\u003e10\u003c/span\u003e). The second was the evaluation of the maximum depth graph, which depicts the model\u0026apos;s learning behaviour when training and testing, in terms of accuracy (\u003cspan class=\"CitationRef\"\u003e31\u003c/span\u003e) (Fig. \u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e). As approaches to control for this factor we used automatic cross validation for all results and confidence intervals, de-duplication in the row dataset (\u003cspan class=\"CitationRef\"\u003e32\u003c/span\u003e), regularization (using L1, L2 to prevent too large coefficients), early stopping (by monitoring the validation set during training for when it starts deteriorating) (\u003cspan class=\"CitationRef\"\u003e31\u003c/span\u003e) and data augmentation (to make the model more robust and less likely to overfit) (\u003cspan class=\"CitationRef\"\u003e33\u003c/span\u003e).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\n \u003ch2\u003eMethods for Traditional Statistical Analyses\u003c/h2\u003e\n \u003cp\u003eWe performed traditional statistical data analyses to support and strengthen our methodology and conclusions. We performed the Shapiro-Wilk test and conducted a bivariate analysis accordingly. We used Pearson/Spearman\u0026rsquo;s correlations for independent quantitative variables and the Chi-squared test for qualitative variables (Additional File AF 1dt o 1f). Student\u0026acute;s T/Mann-Whitney\u0026rsquo;s U test (Additional File 1g) was used to explore differences in means or medians in quantitative-qualitative variable associations. Then we confirmed the association between each independent variable and the outcome (Additional File 1h).\u003c/p\u003e\n \u003cp\u003eFinally, multivariate analysis (Additional File 1j) was performed to evaluate associations in the presence of all model variables to control for confusion; the assumptions of independence of errors and homoscedasticity were previously evaluated (Additional File 1i). Only variables with p-values of \u0026gt;\u0026thinsp;0.05 were maintained in the model. Variance due to the regression models was assessed using Nagelkerke\u0026rsquo;s R-squared and AIC (Additional File 1k). Additional File 1l shows the explainability for different models, exploring the model\u0026rsquo;s variability when one variable was included or dropped at a time.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\n \u003ch2\u003eETHICAL CONSIDERATIONS\u003c/h2\u003e\n \u003cp\u003eThis paper was developed in accordance with the stipulations and rules cited in the Declaration of Helsinki and the CIOMS International Ethical Guidelines for Health-related Research Involving Humans in their last version from 2016. It describes the algorithm training, testing, and selection process as thoroughly as possible to ensure transparency. Databases were voluntarily provided and completely anonymised. The study was approved by an external ethics committee in Peru and by the ethics committees at each of the health institutions in Colombia. Additional File 4 compiles each committee\u0026apos;s approved research protocols and acceptance letters. No exclusions were made based on gender, race, sexual orientation, or nationality.\u003c/p\u003e\n \u003cp\u003eFigure \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. User interface for performing single predictions based on the algorithm selected and the patient\u0026rsquo;s variables\u003c/p\u003e\n \u003cp\u003eTaken from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://hippocrates.arkangel.a\u003c/span\u003e\u003c/span\u003e. All rights reserved.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eLegend\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eUsing patient information as input (right upper quadrant), the algorithm predicts his/her CKD risk. CKD risk is defined as an eGFR\u0026thinsp;\u0026lt;\u0026thinsp;60mL/min/1.73m2. Prediction (1\u0026thinsp;=\u0026thinsp;risk, 0\u0026thinsp;=\u0026thinsp;not at risk) is displayed on the left as the primary outcome, accompanied by the SHAP values for each variable.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\u003ch2\u003eStudy population\u003c/h2\u003e\u003cp\u003eThree databases from three healthcare institutions were obtained and merged, giving 645412 initial registries (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, part a). After applying quality control and selection criteria to the merged database (part b), we obtained a sample size of 203067 registries. In the baseline sample, CKD prevalence was 13.54%, since 27502 records presented with an eGFR\u0026thinsp;\u0026lt;\u0026thinsp;60 ml/min/1.73m2, as defined for this study\u0026rsquo;s outcome, and calculated using the CKD EPI 2021 updated formula. Diabetic individuals accounted for 16.2% of the sample (32909). The mean age was 47.55 years, and the mean time since T2D diagnosis was 1.55 years. Thirty-three percent of registries were female. Twenty-seven were hypertensive. The average record belonged to a normotensive, overweight patient (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eWhen evaluated separately, Diabetes was almost five times more prevalent in D1 than in D2. This difference was even more marked in D3, where virtually no records reported T2D. CKD and hypertension were also minimal in D3, while more than half of the records in D1 presented with these chronic diseases. Blood pressure and BMI were similar among the three databases. However, D3 was predominantly composed of young male adults, while D1 and D2 were twice the age and more balanced in the sex distribution.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\u003ch2\u003eAlgorithm Testing and Selection\u003c/h2\u003e\u003cp\u003eThe total number of models trained was 424 (Additional File 2). Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e5\u003c/span\u003e summarizes them according to their distribution by architecture. Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e6\u003c/span\u003e demonstrates the number of experiments performed (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e) and the number of models trained on each (22 for most). After grouping them by experiment, the algorithms with sensitivities\u0026thinsp;\u0026gt;\u0026thinsp;90% were chosen, resulting in 120 preselected high-performing models (Additional File 3). Then we selected the ones with AUC\u0026thinsp;\u0026gt;\u0026thinsp;90% (24 algorithms), and then those with precision\u0026thinsp;\u0026gt;\u0026thinsp;90% (2 algorithms remaining) (Table\u0026nbsp;\u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e7\u003c/span\u003e). Specificity\u0026thinsp;\u0026gt;\u0026thinsp;85% was the following criterion, resulting in the selection of the final model. Criteria 5 to 8 were not needed.\u003c/p\u003e\u003cp\u003e We observed general poor performance in the models trained using only the Colombian (Col) sample; they never obtained sensitivities higher than 86% and presented a low average AUC (Additional File 3). Therefore, they were never tested in the Peruvian population. The performances derived from models trained on stratified data were higher for non-T2D. On the contrary, models trained in T2D records presented with lower general metrics, and sensitivities did not surpass 81% for any of them (Additional File 3, sheet 2). Consequently, none of the diabetic-strata trained models made the top performing, according to the criteria stated in the Methodology section (Tables\u0026nbsp;\u003cspan refid=\"Tab11\" class=\"InternalRef\"\u003e11\u003c/span\u003e and \u003cspan refid=\"Tab12\" class=\"InternalRef\"\u003e12\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe final selected model was trained on the undivided sample size, using an augmentation of SMOTE 1. It had a Decision Tree architecture with a depth of 33 layers (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The augmentation technique was explained in detail in the Methods section under \u0026ldquo;Data Augmentation.\u0026rdquo; The performance metrics and confusion matrix are displayed in Table\u0026nbsp;\u003cspan refid=\"Tab8\" class=\"InternalRef\"\u003e8\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003ch2\u003eAlgorithm Performance and Validation\u003c/h2\u003e\u003cp\u003eCompared to the standard of reference for risk of CKD, which was defined as an eGFR\u0026thinsp;\u0026lt;\u0026thinsp;60 ml/min/1.73m2 and calculated using the standardized 2021 CKD EPI equation, our model achieved a sensitivity of 91.4%, specificity of 90.1%, precision of 90.2%, AUC of 90.9%, f1 score of 90.8%, and accuracy of 90.7%. SHAP values were calculated and are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. The most influential variable was age, followed by T2DD and BMI. The SHAP value for the binary T2D was consistently zero for most of the best-performing models.\u003c/p\u003e\u003cp\u003ePerformance in the diabetes-status stratified sample shows an equilibrium between metrics in T2D and non-T2D (Additional files 1m-1o). Sensitivity and precision were higher for diabetics, and specificity and accuracy were higher for non-diabetics. It is important not to confuse this stratified by diabetes status assessment of the chosen model with evaluating the performance of algorithms trained in those divided populations, which yielded poor-performing models, before final model selection. The Z test p-value demonstrates no statistical differences between the performance metrics of both groups.\u003c/p\u003e\u003cp\u003eAdditionally, we performed stratified analyses by sex and age (Additional Files 1p to 1r). The prevalence rates for males and females were 53% and 47%, respectively. The analysis of age stratified as \u0026lt;\u0026thinsp;40, 40 to 60, and \u0026gt;\u0026thinsp;60 years old provided evidence that CKD risk in patients younger than sixty years was similar (24% prevalence). For those older than sixty, it was twice as likely to be present (53%) in our population.\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab10\" class=\"InternalRef\"\u003e10\u003c/span\u003e presents absolute differences in performance metrics in training and testing datasets. The absolute difference in accuracy is lower than 10 (9.4), which may or may not indicate overfitting, since there is no established threshold (see Discussion) (\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e). The maximum depth curve (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e) also presents some indicators of overfitting, such as a validation curve surpassed by the training curve. Despite this, it evidences a plateau from depth\u0026thinsp;=\u0026thinsp;29, which means it reaches stability and non-degradation with increases in depth, which is an argument in favour of the non-overfitted model.\u003c/p\u003e\u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\u003ch2\u003eTraditional Statistics Results\u003c/h2\u003e\u003cp\u003eAdditional Files 1d to 1l provide evidence of the tests described in Methods. We found a statistical association among all the independent variables and the outcome. We confirmed the assumption of independence of errors but found a high probability of heteroscedasticity (Additional File 1i). Consequently, we performed a multivariate analysis with a binomial regression (Additional File 1j). All the features tested had a p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05. We evaluated the model, including the eight variables in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, and dropping one at a time, to assess performance and p-values (Additional File 1l). These were compared based on McFadden and Nagelkerke\u0026rsquo;s R-squared and AIC (Additional File 1k). The model with all features presented the highest values. Confusion was controlled in the presence of all variables, and statistical associations were sustained. Consequently, all eight variables were statistically and clinically relevant and were kept in the model. Additional File 1a summarizes the clinical associations described and supported in the literature.\u003c/p\u003e\u003cp\u003eWe also calculated and interpreted the Prevalence Ratio (PR). Its value was closer to the unit (1\u0026thinsp;=\u0026thinsp;null association) for SBP and DBP, followed by BMI, Diabetes Duration, and age, in that order. The highest PR in our sample was for sex (male) with 2.6, followed by the presence of Diabetes (PR 2) and Hypertension (PR 1.9). Nagelkerke\u0026rsquo;s R-squared indicates that this regression explains 53% of the outcome variability under the conditions of our study.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eMachine Learning Model Distribution by Architecture\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eArquitecture\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAbsolute Frequency\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRelative Frequency\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDecisionTree\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e38\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e8.96%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e360\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e84.91%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eKNC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e4.95%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLogRegression\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.47%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRFC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.47%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSVC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.24%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTotal\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e424\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e100.00%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Machine Learning Model Distribution by Experiment/Project Name\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eThe top two performing models*\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"10\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eProject name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSample Distribution\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModel Name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eArchitecture\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSensitivity/ Recall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePPV/ Precision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eF1Score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAllDataSmoteEqual1Erc\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eAll data\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003eDecisionTreeC_max_depth-33_min_samples_split-2_RandomSearch\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003eDecisionTree\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.914\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.909\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.902\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.901\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e\u003cb\u003e0.908\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e\u003cb\u003e0.907\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDbSmoteEqual1Erc\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eStratified data - db\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRFC_n_estimators-388_max_depth-35_RandomSearch\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRFC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.966\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.9378\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.906\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.6423\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.935\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.895\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e*The defining step was to apply criterion 4: \u0026ldquo;specificity higher than 85%\u0026rdquo;, which led to the model in bold.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab8\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 8\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eTesting phase dataset\u0026rsquo;s confusion matrix and performance metrics for the selected best-performing model\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eArkangel app\u0026nbsp; Classification\u003c/p\u003e\u003cp\u003eTesting dataset\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eCKD EPI\u0026thinsp;\u0026lt;\u0026thinsp;60mL/min/1.73m\u003csup\u003e2\u003c/sup\u003e (true case)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eCKD EPI\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;60mL/min/1.73m\u003csup\u003e2\u003c/sup\u003e (false case)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eTotal\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eValue\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eIC 0.95% - Lower\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eIC 0.95% - Upper\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003eSensitivity\u003c/b\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.911\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.917\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003eSpecificity\u003c/b\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.901\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.904\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003ePrecision\u003c/b\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.902\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.898\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAI Algorithm true case\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e28797\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e3142\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e31939\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.907\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.910\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAI Algorithm false case\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e2718\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e28546\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e31264\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003eF1 score\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.905\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.911\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTotal\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e31515\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e31688\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e63203\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003eAUC\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.909\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.910\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.910\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab9\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 9\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eTraining phase dataset\u0026rsquo;s confusion matrix and performance metrics for the chosen model\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eArkangel app\u003c/p\u003e\u003cp\u003eClassification\u003c/p\u003e\u003cp\u003eTraining dataset\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eCKD EPI\u0026thinsp;\u0026lt;\u0026thinsp;60mL/min/1.73m\u003csup\u003e2\u003c/sup\u003e (true case)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eCKD EPI\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;60mL/min/1.73m\u003csup\u003e2\u003c/sup\u003e (false case)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\" morerows=\"3\" rowspan=\"4\"\u003e\u003cp\u003eTotal\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eMetric\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eValue\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eIC 0.95% - Lower\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eIC 0.95% - Upper\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003eSensitivity\u003c/b\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003eSpecificity\u003c/b\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003ePrecision\u003c/b\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAI Algorithm true case\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e143619\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e505\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e144124\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAI Algorithm false case\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e431\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e143372\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e143803\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003eF1 score\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTotal\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e144050\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e143877\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e287927\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003eAUC\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab10\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 10\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison between training and testing performances\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eF1 score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTraining\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.996\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.997\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTesting\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.901\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.909\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.902\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDifference\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.083\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.095\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.088\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.094\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.027\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.089\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eZ test p-value\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.078\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.077\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.079\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.065\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.600\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Shapley Additive Values for each attribute on the model\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eLegend\u003c/strong\u003e\u003cp\u003eThe Shapley Additive Values are a measurement used in ML models to quantify each variable's weight in the model, on top of performance metrics. Despite its documented strong association, T2D appeared as the least influential variable for the algorithm's outcome. See the Discussion section for in-depth analysis.\u003c/p\u003e\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. Maximum Depth of Individual Estimators\u003c/p\u003e\u003cp\u003eTaken from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://hippocrates.arkangel.a\u003c/span\u003e\u003cspan address=\"https://hippocrates.arkangel.a\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. All rights reserved.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eLegend\u003c/strong\u003e\u003cp\u003eThe chosen model, DecisionTreeC_max_depth-33_min_samples_split-2_RandomSearch, reveals clinically optimal performance despite showing traditional indicators of overfitting, such as the training curve surpassing the validation curve. The model achieved 99.7% accuracy in the training set and 91% in validation, presenting a gap of 9 percentage points between the two sets. The stability of validation performance, evidenced by the plateau reached from depth\u0026thinsp;=\u0026thinsp;29, and the absence of significant model degradation, suggests that the observed difference does not compromise clinical generalization. The conclusion is that the model does not present substantial overfitting and can be considered viable for implementation in medical screening settings, where practical utility and case identification capability outweigh strict criteria for statistical perfection.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab11\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 11\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eBest performing model for those trained in the Colombian sample exclusively\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"10\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eProject name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTraining Sample Distribution\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModel Name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eArchitecture\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSensitivity/ Recall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePPV/ Precision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003cp\u003eScore\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCOLAugmentedpositivesx2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eColombian Sample\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDeepFullyConnectedNeuralNetwork_Adagrad_l1_dropoutFalse_numHiddenLayers5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eDeep Fully Connected Neural Network\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.999\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.828*\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.672\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.001\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.804\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.672\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e*The metric in bold signifies the selection criteria for which the respective algorithm was disqualified (see Methods). In this case, AUC should be \u0026gt;\u0026thinsp;90% to qualify.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab12\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 12\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eBest performing model for those trained in the diabetes/non diabetes stratified sample\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"10\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eProject name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTraining Sample Distribution\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eModel Name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eArchitecture\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSensitivity/ Recall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ePPV/ Precision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003cp\u003eScore\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDbSmoteEqual1Erc\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDiabetic Sample\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRFC_n_estimators-388_max_depth-35_RandomSearch\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRandom Forest Classifier\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.967\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.938\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.906\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e\u003cb\u003e0.643*\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.935\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.895\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNdbSmoteEqual1Erc\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNon-diabetic sample\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eDecisionTreeC_max_depth-26_min_samples_split-12_RandomSearch\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eDecisionTreeClassifier\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.901\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.951\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.885*\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e\u003cp\u003e0.923\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e\u003cp\u003e0.893\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e\u003cp\u003e0.914\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e*The metric bold signifies the selection criteria for which the respective algorithm was disqualified (see Methods). In this case, PPV should be \u0026gt;\u0026thinsp;90% (non-diabetics), and Specificity\u0026thinsp;\u0026gt;\u0026thinsp;85% (diabetics) to qualify.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eWe trained AI algorithms using anonymised data from over 200.000 Colombian and Peruvian EHRs, attempting to reliably classify their CKD risk by predicting the probability of having a eGFR \u0026lt;60mL/min/1.73m2, based on eight readily available, non-invasive, and measurable at the point-of-care attributes: age, sex, SBP, DBP, BMI, hypertension, presence of diabetes (T2D), and diabetes duration (T2DD). The study responds to the need for an extremely sensitive screening strategy for CKD, especially for LMIC, where human and technical resources for finding and prioritizing patients are scarce. \u003c/p\u003e\n\u003cp\u003eResearch in LMIC has emphasised producing models using fewer variables than other good-performing ML models, without compromising their quality. Rashed-Al-Mahfuz et al. trained AI algorithms for CKD prediction in Bangladesh, with an accuracy of 99.50%, sensitivity of 98.75%, specificity of 100%, precision 100%, and AUC of 99.38%. Their best-performing model was a Random Forest, with 13 features from those in the UCI-ML repository. Using SHAP values, they determined that the most influential variables for their model were haemoglobin levels, hypertension, and blood sugar levels (21). \u003c/p\u003e\n\u003cp\u003eIn Colombia, Isaza-Ruget et al. (2024) developed an AI model to predict the risk of CKD progression in patients in stages 3-5, using Gradient Boosting and ten features: sex, residence, diabetes, hypertension, haemoglobin, creatinine, HDL, LDL, and eGFR (14). However, the population with early alterations in eGFR would still go unnoticed. \u003c/p\u003e\n\u003cp\u003eWe tested and internally validated our algorithm, evidencing its optimal performance, using even fewer attributes. The chosen model was a Decision Tree with 33 layers, trained on an augmented sample size using SMOTE=1 (351130 EHRs). Within the study’s conditions, our model detects individuals at risk of presenting altered eGFR defined as \u0026lt;60mL/min/1.73m2 (mildly to moderate increased CKD risk, according to 2024 KIDGO guidelines), with sensitivity of 91.4%, specificity of 90.1%, a precision 90.2%, AUC of 90.9% and f1 score of 90.8%, only using eight variables.\u003c/p\u003e\n\u003cp\u003eAutomatic cross-validation was carried out on the performance metrics in the Arkangel App. The narrow 95% CIs in Table 8 support the validity of the results, on top of the hypothesis test comparisons (no statistical differences, p-value \u0026gt;0.05) of the proportions of performance metrics in the training and testing datasets. Subgroup analyses, implemented by running the model in the sample stratified by diabetes status, evidenced a balanced, consistent, high-performing model in both populations (AF 1l to 1n). \u003c/p\u003e\n\u003cp\u003eThis stratified analysis is also interesting given the difference in absolute numbers of diabetic patients compared to non-diabetic patients in the base population. Only 16% of EHRs belonged to diabetic patients, predominantly from the Colombian population (D1-D2). They were older and more chronically ill than the Peruvian population (D3) (mean ages 65 and 36, respectively). \u003c/p\u003e\n\u003cp\u003eThere are two important things to emphasise regarding this issue: the first is that the nature of D3 most likely explains the baseline population differences; the Peruvian healthcare company was mainly an insurer for vehicle-related accidents, which are more common in the ages under 35 (35,36). Conversely, Colombian health institutions, although they also acted as insurance companies, had a broader spectrum of health-related events, especially chronic disease care.\u003c/p\u003e\n\u003cp\u003eThe second fact is that this imbalance is common in retrospective, secondary source research (37), especially in the face of AI capabilities (38–41), where enormous quantities of data are expected to be processed at once. This bottleneck is worsened when trying to find specific literature about hetero/homogeneity of sample sizes in ML-trained models; there is virtually none. For the results to be valid, the sample size should be as similar as possible in all characteristics, except the one expected as the outcome. This traditional epidemiological view is supported by statistical theorems and methods amply described elsewhere (42–47). Nevertheless, one could also state that to achieve maximum efficiency with AI models, it is not a bad practice to include different base populations in the studies and show the model’s possible variations of relevant characteristics (48,49).\u003c/p\u003e\n\u003cp\u003eThe latter was the rationale behind exploring the effect of sample size’s hetero/homogeneity by training using different sample distributions. We trained 88 models using only the Colombian and evaluated their metrics before testing them on the Peruvian sample as an external validation. We never got to that point because, as mentioned in Results, the model's performances in this “more homogeneous” sample resulted in poor metrics that did not comply with the selection criteria stated in Methods. An example is the DNN model shown in Table 11, which was the top performer in this group of experiments. \u003c/p\u003e\n\u003cp\u003eThe same goes for the models trained on the diabetic/non-diabetic population (152 algorithms). Table 12 shows the best-performing model for these experiments, in which a model was trained on each group. The algorithms trained in diabetics had even poorer performance, which could be explained by the class imbalance, in favour of the non-diabetics, even after SMOTE augmentation of the diabetic strata. This evidence meant that training with a stratified sample was not satisfactory for addressing the imbalance. The performance of the non-diabetic trained model was better, but still not up to established standards.\u003c/p\u003e\n\u003cp\u003eIn addition, these two models needed an extra step: an ensemble learning model approach, where each algorithm is weighted and computed to produce a final, single result. Although we explored this avenue for the present study, the lack of a strong rationale for dividing the sample into diabetes strata, the evidence of better results by controlling for class imbalance using ML augmentation techniques alone, and the complexity of the method itself, determined that these experiments were not included in the tables displayed. Their results would be available upon reasonable request.\u003c/p\u003e\n\u003cp\u003eAnother essential topic to discuss is the SHAP values and their meaning compared to traditional statistical analyses. According to SHAP, the ranking of features by their importance or weight for the final model’s decision is Age, T2DD, BMI, SBP, sex, DBP, hypertension, and T2D, in that order. At first glance, this is remarkable because there is a strong association between diabetes and CKD, as documented in countless scientific articles (Additional File 1a). \u003c/p\u003e\n\u003cp\u003eAfter a closer look, this association might be observed with T2DD in second place, representing the burden of the association with CKD. Nevertheless, it ranks first since it provides more information about the condition (years of disease progression vs. presence or absence of T2D). \u003c/p\u003e\n\u003cp\u003eTo test this hypothesis, we evaluated collinearity using the Variance Inflation Factor (VIF) (Additional File 1s). General recommendation is that a VIF value lower than five is acceptable (50–52), according to the study's nature and the variable's clinical relevance. T2DD and T2D have a VIF of three, so we believe it is prudent to declare potential collinearity. However, due to the clinical value of the variables, this is not enough to warrant exclusion. \u003c/p\u003e\n\u003cp\u003eBesides, literature sustains that ML architectures, particularly Random Forest and Decision Trees, are unaffected by collinearity (53–55). They are uniquely qualified to deal with multicollinearity through the conditional subgroups technique, a strategy used in Explainable Artificial Intelligence (XAI) methods. This method manages multicollinearity using conditional subgroups with permutation feature importance and partial dependency plot. In Decision Trees, data splits are made based on individual variables, rather than coefficients, into groups that secure a more homogeneous distribution in one group and a more heterogeneous distribution in the rest. The rationale behind this method is to create groups so that the feature of interest is less dependent on all other features in all the subgroups (56).\u003c/p\u003e\n\u003cp\u003eTo further evaluate the behaviour of variables in the AI model, we performed bi- and multivariate analyses with traditional statistical methods. All features were confirmed as statistically associated with CKD in the presence of the others. The best model included the eight variables, which explained 53.1% of our sample's model variability. The magnitude of the positive association was evidenced by the Prevalence Ratios (PR): T2DD was among the weaker associations, while T2D (PR 2) and sex (PR 2.6) displayed the highest associations.\u003c/p\u003e\n\u003cp\u003eThe latter ranking is not equivalent in meaning to the ranking established by SHAP. These values are a machine learning interpretability tool that quantifies the contribution of each feature to predicting an individual instance in a model (57–59). They provide a local explanation of model output, showing how much each feature increases or decreases the predicted value relative to a baseline. Conversely, PRs are epidemiological/statistical measures that compare the prevalence of an outcome in two groups (60). They quantify the relative risk or association between exposure and outcome in population-level data (61).\u003c/p\u003e\n\u003cp\u003eConsequently, all results point to the maintenance of both variables in the model: their documented association with the outcome, the relevance of T2DD in the ML model, the strong positive association of T2D proven by its PR in statistical analysis, and the better statistical explainability of the variations in outcome when including all eight variables in the binary regression. \u003c/p\u003e\n\u003cp\u003eThe fourth significant issue is the possibility of overfitting. According to accuracy, Figure 7 shows the model's learning behaviour over time. Overfitting is generally identified when a model performs significantly better on the training data than on the testing data, indicating poor generalization (31,33,34). The ideal graph would display an increasing curve that stabilizes at the threshold. The training curve should not surpass the validation curve. However, there is no threshold to define this difference in performance. It has been described that this difference in accuracy should not be over 10%. \u003c/p\u003e\n\u003cp\u003eAlthough our model differs in accuracy by 9.4 percentage points, its performance in both datasets and the subgroup assessment is optimal. Additionally, we employed multiple approaches to control for it (see Methods). Specifically, for the augmentation strategy, we used SMOTE, which is an oversampling technique that is less prone to introducing overfitting than others because synthetic data is created by an algorithm (K-nearest neighbour) that seeks to balance the dataset (33). This motivates us to share our algorithm's development and internal validation through this manuscript.\u003c/p\u003e\n\u003cp\u003eThis algorithm needs external validation to appraise whether its performance metrics are maintained in populations with entirely different characteristics. The study is being developed thanks to our access to the iCaReMe (CardioRenal and Metabolic) global registry (NCT03549754). The dataset contains healthcare data of individuals across 11 countries (62,63).\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eOur study demonstrates that a decision tree model trained with eight non-invasive clinical variables can accurately identify individuals at risk of CKD without requiring specialized tests. This approach is feasible for large-scale screening in low-resource settings and supports integration into electronic health records to prioritize confirmatory testing and timely care.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eClinical Trial Number\u003c/strong\u003e: not applicable\u0026nbsp;\u003c/p\u003e\n\u003ch4\u003e\u003cstrong\u003eETHICS APPROVAL AND CONSENT TO PARTICIPATE\u003c/strong\u003e\u003c/h4\u003e\n\u003cp\u003eThis study is classified as minimal risk according to Resolution 8430 of 1993\u0026nbsp;(Colombia), which stipulates that minimal risk-observational studies, based on secondary data, exclusively from clinical sources, do not need individual consent\u0026nbsp;(64). In addition, the Declaration of Helsinki (2022) and the CIOMS Guidelines (2016) state that individual informed consent may be waived when the conditions of minimal risk, impracticability of obtaining consent, and high social and scientific value are met, while always ensuring strict confidentiality of the information\u0026nbsp;(65). Furthermore, the study was reviewed and approved by an external ethics committee in Peru and the ethics committees at each participating health institution in Colombia (see Additional File 4, which includes the letters of ethics approval, the research protocols, and the letter we sent to the external committee in Perú, in PDF format into one compressed file)\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCONSENT FOR PUBLICATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. In accordance with the Resolution 8430 of 1993 from Colombia, the Declaration of Helsinki (2022) and the CIOMS Guidelines (2016), individual consent for publication was waived, given that the research was based on anonymized secondary data, was classified as minimal risk, it was not operationally feasible to obtain individual consent, and strict confidentiality of the information was guaranteed at all times. There are no images or any other sensitive information that could compromise anonymity.\u0026nbsp;\u003c/p\u003e\n\u003ch4\u003e\u003cstrong\u003ePUBLICATION AVAILABILITY OF DATA AND MATERIALS\u003c/strong\u003e\u003c/h4\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study are not publicly available due to restrictions related to patient confidentiality and institutional agreements. However, anonymised data may be made available from the corresponding author on reasonable request and with permission of the participating health institutions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCOMPETING INTERESTS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors completed the International Committee of Medical Journal Editors\u0026nbsp;(ICMJE)\u0026nbsp;disclosure form from https://www.icmje.org/disclosure-of-interest/\u003c/p\u003e\n\u003cp\u003eJM, AP, DJ, JZ, IL, and NCV work at Arkangel AI.\u0026nbsp;\u003cbr\u003e\u0026nbsp;IL received support from AstraZeneca to attend and present a poster at the Colombian Nephrology Conference from ASOCOLNEF on August 1, 2, and 3, 2024.\u0026nbsp;\u003cbr\u003e\u0026nbsp;DC works as Medical Therapeutic Area Lead at AstraZeneca, responsible for medical strategy for company products, including Dapagliflozin. He has received support from AstraZeneca to attend ERA 2023; EASD 2022, 2023, 2024; ESC 2025 and ESCMID 2022 as part of his job and participated as an organizer of AdBoards for different products in AstraZeneca\u003c/p\u003e\n\u003cp\u003eJJA, WB, and WC don’t report any conflict of interest.\u003c/p\u003e\n\u003cp\u003eVE is an expert advisor and speaker for Novo Nordisk and an expert advisor for some pharmaceutical companies namely Novo Nordisk and Astrazeneca and received support from Boehringer Ingelheim to attend ERA 2024.\u003c/p\u003e\n\u003cp\u003eAML is the brand ambassador and is a Corporate medical manager for Pulso Salud Corporación Médica,\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFUNDING\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe project was funded by AstraZeneca.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHORS' CONTRIBUTIONS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJM: Methodology design and supervision of information of data processing, AI design and training and model output\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAP: Methodology execution, including information processing, AI design, experimentation and training, model output, results, and interpretation\u003c/p\u003e\n\u003cp\u003eDJ: Methodology execution, including information processing, AI design, experimentation and training, model output, results, and interpretation\u003c/p\u003e\n\u003cp\u003eJZ: Project management and supervision\u003c/p\u003e\n\u003cp\u003eIL: Scientific paper writing and editing, including abstract, introduction, methodology, results, and discussion\u003c/p\u003e\n\u003cp\u003eNCV: In charge of scientific paper writing and editing, including abstract, introduction, methodology, results, and discussion\u003c/p\u003e\n\u003cp\u003eDC: Medical expert, revisor, and editor\u003c/p\u003e\n\u003cp\u003eJJA: Expert reviewer of final manuscript\u003c/p\u003e\n\u003cp\u003eWB: Expert reviewer of the final manuscript and provided anonymized data for the study\u003c/p\u003e\n\u003cp\u003eVE: Expert reviewer of the final manuscript and provided anonymized data for the study\u003c/p\u003e\n\u003cp\u003eAML: Expert reviewer of the final manuscript and provided anonymized data for the study\u003c/p\u003e\n\u003cp\u003eWC: Expert reviewer of the final manuscript and provided anonymized data for the study\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eACKNOWLEDGEMENTS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank Somedyt, Medisinu, and Pulso Salud for providing the data necessary for the development of this project. We also thank AstraZeneca for their medical insight and collaboration.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eGBD 2019 Mental Disorders Collaborators. Global, regional, and national burden of 12 mental disorders in 204 countries and territories, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Psychiatry. 2022 Feb;9(2):137\u0026ndash;50.\u003c/li\u003e\n \u003cli\u003eForecasting life expectancy, years of life lost, and all-cause and cause-specific mortality for 250 causes of death: reference and alternative scenarios for 2016\u0026ndash;40 for 195 countries and territories - The Lancet [Internet]. [cited 2025 May 30]. Available from: https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(18)31694-5/fulltext\u003c/li\u003e\n \u003cli\u003eLevey AS, Astor BC, Stevens LA, Coresh J. Chronic kidney disease, diabetes, and hypertension: what\u0026rsquo;s in a name? Kidney Int. 2010 Jul;78(1):19\u0026ndash;22.\u003c/li\u003e\n \u003cli\u003eKovesdy CP. Epidemiology of chronic kidney disease: an update 2022. Kidney Int Suppl. 2022 Apr 1;12(1):7\u0026ndash;11.\u003c/li\u003e\n \u003cli\u003eKalantar-Zadeh K, Jafar TH, Nitsch D, Neuen BL, Perkovic V. Chronic kidney disease. The Lancet. 2021 Aug;398(10302):786\u0026ndash;802.\u003c/li\u003e\n \u003cli\u003eINTERNATIONAL SOCIETY OF NEPHROLOGY. KDIGO Clinical Practice Guideline for Acute Kidney Injury [Internet]. INTERNATIONAL SOCIETY OF NEPHROLOGY; 2012. Available from: https://kdigo.org/wp-content/uploads/2016/10/KDIGO-2012-AKI-Guideline-English.pdf\u003c/li\u003e\n \u003cli\u003eCardiovascular Implications of the 2021 KDIGO Blood Pressure Guideline for Adults With Chronic Kidney Disease [Internet]. [cited 2025 Jul 7]. Available from: https://www.jacc.org/doi/epdf/10.1016/j.jacc.2022.02.040\u003c/li\u003e\n \u003cli\u003eInternationl Society of Nephrology. KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease [Internet]. 2024. Available from: https://www.researchgate.net/profile/Ifeoma-Ulasi/publication/379470748_KDIGO_2024_Clinical_Practice_Guideline_for_the_Evaluation_and_Management_of_Chronic_Kidney_Disease/links/660e35f0b839e05a20bd32b7/KDIGO-2024-Clinical-Practice-Guideline-for-the-Evaluation-and-Management-of-Chronic-Kidney-Disease.pdf\u003c/li\u003e\n \u003cli\u003eEvans M, Lewis RD, Morgan AR, Whyte MB, Hanif W, Bain SC, et al. A Narrative Review of Chronic Kidney Disease in Clinical Practice: Current Challenges and Future Perspectives. Adv Ther. 2022 Jan 1;39(1):33\u0026ndash;43.\u003c/li\u003e\n \u003cli\u003eEchouffo-Tcheugui JB, Kengne AP. Risk Models to Predict Chronic Kidney Disease and Its Progression: A Systematic Review. PLOS Med. 2012 Nov 20;9(11):e1001344.\u003c/li\u003e\n \u003cli\u003ePadmanabhan AA, Balczewski EA, Singh K. Artificial Intelligence Systems in CKD: Where Do We Stand and What Will the Future Bring? Adv Chronic Kidney Dis. 2022 Sep 1;29(5):461\u0026ndash;4.\u003c/li\u003e\n \u003cli\u003eWu CC, Islam MM, Poly TN, Weng YC. Artificial Intelligence in Kidney Disease: A Comprehensive Study and Directions for Future Research. Diagnostics. 2024 Jan;14(4):397.\u003c/li\u003e\n \u003cli\u003eIslam MdA, Majumder MdZH, Hussein MdA. Chronic kidney disease prediction based on machine learning algorithms. J Pathol Inform. 2023 Jan 12;14:100189.\u003c/li\u003e\n \u003cli\u003eIsaza-Ruget MA, Yomayusa N, Gonz\u0026aacute;lez CA, H CA, de Oro V FA, Cely A, et al. Predicting chronic kidney disease progression with artificial intelligence. BMC Nephrol. 2024 Apr 26;25(1):148.\u003c/li\u003e\n \u003cli\u003eTangri N, Peach EJ, Franz\u0026eacute;n S, Barone S, Kushner PR. Patient Management and Clinical Outcomes Associated with a Recorded Diagnosis of Stage 3 Chronic Kidney Disease: The REVEAL-CKD Study. Adv Ther. 2023 Jun;40(6):2869\u0026ndash;85.\u003c/li\u003e\n \u003cli\u003eTangri N, Ferguson TW. Artificial Intelligence in the Identification, Management, and Follow-Up of CKD. Kidney360. 2022 Mar 31;3(3):554\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eTangri N, Grams ME, Levey AS, Coresh J, Appel LJ, Astor BC, et al. Multinational Assessment of Accuracy of Equations for Predicting Risk of Kidney Failure: A Meta-analysis. JAMA. 2016 Jan 12;315(2):164\u0026ndash;74.\u003c/li\u003e\n \u003cli\u003eL. Rubini PS. Chronic Kidney Disease [Internet]. UCI Machine Learning Repository; 2015 [cited 2024 May 22]. Available from: https://archive.ics.uci.edu/dataset/336\u003c/li\u003e\n \u003cli\u003eChan L, Nadkarni GN, Fleming F, McCullough JR, Connolly P, Mosoyan G, et al. Derivation and validation of a machine learning risk score using biomarker and electronic patient data to predict progression of diabetic kidney disease. Diabetologia. 2021 Jul;64(7):1504\u0026ndash;15.\u003c/li\u003e\n \u003cli\u003eRavizza S, Huschto T, Adamov A, B\u0026ouml;hm L, B\u0026uuml;sser A, Fl\u0026ouml;ther FF, et al. Predicting the early risk of chronic kidney disease in patients with diabetes using real-world data. Nat Med. 2019 Jan;25(1):57\u0026ndash;9.\u003c/li\u003e\n \u003cli\u003eRashed-Al-Mahfuz Md, Haque A, Azad A, Alyami SA, Quinn JMW, Moni MA. Clinically Applicable Machine Learning Approaches to Identify Attributes of Chronic Kidney Disease (CKD) for Use in Low-Cost Diagnostic Screening. IEEE J Transl Eng Health Med. 2021;9:1\u0026ndash;11.\u003c/li\u003e\n \u003cli\u003eSwain D, Mehta U, Bhatt A, Patel H, Patel K, Mehta D, et al. A Robust Chronic Kidney Disease Classifier Using Machine Learning. Electronics. 2023 Jan;12(1):212.\u003c/li\u003e\n \u003cli\u003eChittora P, Chaurasia S, Chakrabarti P, Kumawat G, Chakrabarti T, Leonowicz Z, et al. Prediction of Chronic Kidney Disease - A Machine Learning Perspective. IEEE Access. 2021;9:17312\u0026ndash;34.\u003c/li\u003e\n \u003cli\u003eStevens P, Ahmed S, Carrero J, Foster B, Francis A, Hall R, et al. KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease. Kidney Int. 2024 Apr 1;105:S117\u0026ndash;314.\u003c/li\u003e\n \u003cli\u003escikit-learn.org. KNNImputer [Internet]. Available from: https://scikit-learn.org/1.5/modules/generated/sklearn.impute.KNNImputer.html\u003c/li\u003e\n \u003cli\u003eimbalanced-learn.org. SMOTE [Internet]. [cited 2025 Feb 20]. Available from: https://imbalanced-learn.org/stable/references/generated/imblearn.over_sampling.SMOTE.html\u003c/li\u003e\n \u003cli\u003eKummer A, Ruppert T, Medvegy T, Abonyi J. Machine learning-based software sensors for machine state monitoring - The role of SMOTE-based data augmentation. Results Eng. 2022 Dec 1;16:100778.\u003c/li\u003e\n \u003cli\u003eChawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: Synthetic Minority Over-sampling Technique. J Artif Intell Res. 2002 Jun 1;16:321\u0026ndash;57.\u003c/li\u003e\n \u003cli\u003eUdu AG, Salman MT, Ghalati MK, Lecchini-Visintini A, Siddle DR, Dong H. Emerging SMOTE and GAN Variants for Data Augmentation in Imbalance Machine Learning Tasks: A Review. IEEE Access. 2025;13:113838\u0026ndash;53.\u003c/li\u003e\n \u003cli\u003eArkangel AI. Hippocrates [Internet]. Arkangel AI; [cited 2024 Dec 20]. Available from: https://arkprod.us.auth0.com/u/login/identifier?state=hKFo2SBTelNBSXBFaWJtLVhIcHU3R29ZRERWeHAyUkJSUDRQOaFur3VuaXZlcnNhbC1sb2dpbqN0a\u003cbr\u003eWTZIEtUMjRnSF9JTGNzTkc5eTNQY3hOZFFKX2w1eUUwRm5Qo2NpZNkgNzNXMFZVQWdzdzZLbGx0SXpsbUd3RUcwN0Nza0lRYjU\u003c/li\u003e\n \u003cli\u003eYing X. An Overview of Overfitting and its Solutions. J Phys Conf Ser. 2019 Feb 1;1168:022022.\u003c/li\u003e\n \u003cli\u003eMohammed S, Budach L, Feuerpfeil M, Ihde N, Nathansen A, Noack N, et al. The effects of data quality on machine learning performance on tabular data. Inf Syst. 2025 Jul 1;132:102549.\u003c/li\u003e\n \u003cli\u003eAburass S, Rumman MA. Quantifying Overfitting: Introducing the Overfitting Index. In: 2024 International Conference on Electrical, Computer and Energy Technologies (ICECET [Internet]. Sydney, Australia: IEEE; 2024 [cited 2025 Aug 29]. p. 1\u0026ndash;7. Available from: https://ieeexplore.ieee.org/document/10698575/\u003c/li\u003e\n \u003cli\u003eMitigating Overfitting in Decision Trees [Internet]. [cited 2025 May 24]. Available from: https://tilburgsciencehub.com/topics/analyze/machine-learning/supervised/decision_trees/\u003c/li\u003e\n \u003cli\u003eCalle-Toro J, Ordo\u0026ntilde;ez C, S\u0026aacute;nchez A, Sanjuan J, Badiel M, Pino L, et al. Epidemiologia de Lesiones Relacionadas con Colisiones de Vehiculos Motorizados en dos Centros de Referencia del Suroccidente Colombiano. Reporte del Registro Internacional de Trauma de la Sociedad P. Panam J Trauma Crit Care Emerg Surg. 2014 Apr 1;3:16\u0026ndash;22.\u003c/li\u003e\n \u003cli\u003eAlberdi F, Garc\u0026iacute;a I, Atutxa L, Zabarte M. Epidemiolog\u0026iacute;a del trauma grave. Med Intensiva. 2014 Dec 1;38(9):580\u0026ndash;8.\u003c/li\u003e\n \u003cli\u003eAbdulsadig RS, Rodriguez-Villegas E. A comparative study in class imbalance mitigation when working with physiological signals. Front Digit Health. 2024 Mar 26;6:1377165.\u003c/li\u003e\n \u003cli\u003eChen W, Yang K, Yu Z, Shi Y, Chen CLP. A survey on imbalanced learning: latest research, applications and future directions. Artif Intell Rev. 2024 May 9;57(6):137.\u003c/li\u003e\n \u003cli\u003eFerreiro, Gonzalo. Medium. [cited 2024 Apr 17]. Class Imbalance: a classification headache | by Gonzalo Ferreiro Volpi | Towards Data Science. Available from: https://towardsdatascience.com/class-imbalance-a-classification-headache-1939297ff4a4\u003c/li\u003e\n \u003cli\u003eTh\u0026ouml;lke P, Mantilla-Ramos YJ, Abdelhedi H, Maschke C, Dehgan A, Harel Y, et al. Class imbalance should not throw you off balance: Choosing the right classifiers and performance metrics for brain decoding with imbalanced data. NeuroImage. 2023 Aug 15;277:120253.\u003c/li\u003e\n \u003cli\u003eResearch on expansion and classification of imbalanced data based on SMOTE algorithm | Scientific Reports [Internet]. [cited 2025 Aug 29]. Available from: https://www.nature.com/articles/s41598-021-03430-5\u003c/li\u003e\n \u003cli\u003ePalinkas LA, Horwitz SM, Green CA, Wisdom JP, Duan N, Hoagwood K. Purposeful sampling for qualitative data collection and analysis in mixed method implementation research. Adm Policy Ment Health. 2015 Sep;42(5):533\u0026ndash;44.\u003c/li\u003e\n \u003cli\u003eJager J, Putnick DL, Bornstein MH. More than Just Convenient: The Scientific Merits of Homogeneous Convenience Samples. Monogr Soc Res Child Dev. 2017 Jun;82(2):13\u0026ndash;30.\u003c/li\u003e\n \u003cli\u003eGuest G, Namey EE, Mitchell ML. Sampling in Qualitative Research. In: Collecting Qualitative Data: A Field Manual for Applied Research [Internet]. SAGE Publications, Ltd; 2013 [cited 2025 Aug 29]. p. 41\u0026ndash;74. Available from: https://methods.sagepub.com/book/mono/collecting-qualitative-data/chpt/2-sampling-qualitative-research\u003c/li\u003e\n \u003cli\u003eHennink M, Kaiser BN. Sample sizes for saturation in qualitative research: A systematic review of empirical tests. Soc Sci Med. 2022 Jan 1;292:114523.\u003c/li\u003e\n \u003cli\u003ePerez N. \u003cem\u003en\u003c/em\u003e-Sizes, Attributes, and A Priori Sampling: A Qualitative Sampling Model for Large, Heterogeneous Populations. Am J Qual Res. 2024 Jul 26;8(3):193\u0026ndash;207.\u003c/li\u003e\n \u003cli\u003eLewis DD, Catlett J. Heterogeneous Uncertainty Sampling for Supervised Learning. In: Cohen WW, Hirsh H, editors. Machine Learning Proceedings 1994 [Internet]. San Francisco (CA): Morgan Kaufmann; 1994 [cited 2025 Aug 29]. p. 148\u0026ndash;56. Available from: https://www.sciencedirect.com/science/article/pii/B978155860335650026X\u003c/li\u003e\n \u003cli\u003eSchinkel M, Bennis FC, Boerman AW, Wiersinga WJ, Nanayakkara PWB. Embracing cohort heterogeneity in clinical machine learning development: a step toward generalizable models. Sci Rep. 2023 May 24;13(1):8363.\u003c/li\u003e\n \u003cli\u003eZhao Q, Nooner KB, Tapert SF, Adeli E, Pohl KM, Kuceyeski A, et al. The Transition From Homogeneous to Heterogeneous Machine Learning in Neuropsychiatric Research. Biol Psychiatry Glob Open Sci. 2025 Jan;5(1):100397.\u003c/li\u003e\n \u003cli\u003eBenito BM. Blas M. Benito, PhD. 2023 [cited 2025 Jul 21]. Multicollinearity Hinders Model Interpretability. Available from: https://blasbenito.com/post/multicollinearity-model-interpretability/\u003c/li\u003e\n \u003cli\u003eDetecting Multicollinearity Using Variance Inflation Factors | STAT 462 [Internet]. [cited 2025 Jul 21]. Available from: https://online.stat.psu.edu/stat462/node/180/\u003c/li\u003e\n \u003cli\u003eAkinwande MO, Dikko HG, Samson A. Variance Inflation Factor: As a Condition for the Inclusion of Suppressor Variable(s) in Regression Analysis. Open J Stat. 2015 Dec 11;5(7):754\u0026ndash;67.\u003c/li\u003e\n \u003cli\u003eSaslow E. Collinearity - What it means, Why its bad, and How does it affect other models? [Internet]. Future Vision. 2018 [cited 2025 Jul 21]. Available from: https://medium.com/future-vision/collinearity-what-it-means-why-its-bad-and-how-does-it-affect-other-models-94e1db984168\u003c/li\u003e\n \u003cli\u003eMulti Collinearity in the Tree Based Models | Kaggle [Internet]. [cited 2025 Jul 21]. Available from: https://www.kaggle.com/questions-and-answers/444752\u003c/li\u003e\n \u003cli\u003eYıldırım H. The Multicollinearity Effect on the Performance of Machine Learning Algorithms: Case Examples in Healthcare Modelling. Acad Platf J Eng Smart Syst. 2024 Sep 25;12(3):68\u0026ndash;80.\u003c/li\u003e\n \u003cli\u003eSalih AM. Explainable Artificial Intelligence and Multicollinearity : A Mini Review of Current Approaches [Internet]. arXiv; 2024 [cited 2025 Jul 21]. Available from: http://arxiv.org/abs/2406.11524\u003c/li\u003e\n \u003cli\u003eHamilton RI, Papadopoulos PN. Using SHAP Values and Machine Learning to Understand Trends in the Transient Stability Limit. IEEE Trans Power Syst. 2024 Jan;39(1):1384\u0026ndash;97.\u003c/li\u003e\n \u003cli\u003eNohara Y, Matsumoto K, Soejima H, Nakashima N. Explanation of machine learning models using shapley additive explanation and application for real data in hospital. Comput Methods Programs Biomed. 2022 Feb 1;214:106584.\u003c/li\u003e\n \u003cli\u003ePonce-Bobadilla AV, Schmitt V, Maier CS, Mensing S, Stodtmann S. Practical guide to SHAP analysis: Explaining supervised machine learning model predictions in drug development. Clin Transl Sci. 2024;17(11):e70056.\u003c/li\u003e\n \u003cli\u003eDeddens J, Petersen MR. Approaches for estimating prevalence ratios. Occup Environ Med. 2008 Aug 1;65:481, 501\u0026ndash;6.\u003c/li\u003e\n \u003cli\u003eFajardo-Guti\u0026eacute;rrez A. [Measurement in epidemiology: prevalence, incidence, risk, impact measures]. Rev Alerg Mex Tecamachalco Puebla Mex 1993. 2017;64(1):109\u0026ndash;20.\u003c/li\u003e\n \u003cli\u003eClinConnect [Internet]. 2020 [cited 2025 Sep 17]. ClinConnect | iCaReMe Global Registry (NCT03549754). Available from: https://clinconnect.io/trials/NCT03549754?queryID=bb5320cb7c05993d60da\u0026amp;apos=20\u003c/li\u003e\n \u003cli\u003eiCaReMe Global Registry [Internet]. [cited 2025 Apr 4]. Available from: https://www.astrazenecaclinicaltrials.com/study/D1690R00044/\u003c/li\u003e\n \u003cli\u003eMinisterio de Salud de Colombia. Resoluci\u0026oacute;n n\u0026uacute;mero 8430 de 1993 (4 de octubre), por la cual se establecen las normas cient\u0026iacute;ficas, t\u0026eacute;cnicas y administrativas para la investigaci\u0026oacute;n en salud [Internet]. Bogot\u0026aacute;, Colombia: Diario Oficial No. 41.215; 1993 [cited 2025 Feb 19]. Available from: https://www.minsalud.gov.co/sites/rid/lists/bibliotecadigital/ride/de/dij/resolucion-8430-de-1993.pdf\u003c/li\u003e\n \u003cli\u003eAsociaci\u0026oacute;n M\u0026eacute;dica Mundial. Declaraci\u0026oacute;n de Helsinki \u0026ndash; Principios \u0026eacute;ticos para las investigaciones m\u0026eacute;dicas en seres humanos [Internet]. Fortaleza, Brasil: Asociaci\u0026oacute;n M\u0026eacute;dica Mundial; 2013 [cited 2025 Feb 24]. Available from: https://www.wma.net/policies-post/wma-declaration-of-helsinki-ethical-principles-for-medical-research-involving-human-subjects/\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-nephrology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bnep","sideBox":"Learn more about [BMC Nephrology](http://bmcnephrol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bnep/default.aspx","title":"BMC Nephrology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Renal Insufficiency, Chronic, Machine Learning, Artificial Intelligence, Risk Assessment, Early diagnosis, Latin America","lastPublishedDoi":"10.21203/rs.3.rs-7888843/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7888843/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e Chronic kidney disease (CKD) is a leading global cause of morbidity and mortality, particularly in low- and middle-income countries (LMIC) where access to specialized laboratory tests is limited. Early detection is essential but often delayed due to reliance on serum creatinine-based estimated glomerular filtration rate (eGFR). Artificial intelligence (AI) offers opportunities for simple, sensitive screening models using routinely available variables.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eWe trained and tested a low-cost machine learning algorithm in a multicenter Latin American dataset of 203,067 anonymized records to identify patients at risk of CKD, defined as an eGFR \u0026lt;60 mL/min/1.73m² (CKD-EPI 2021). Eight routinely available, non-invasive variables were used: age, sex, systolic and diastolic blood pressure, body mass index, hypertension, presence of type 2 diabetes (T2D), and diabetes duration (T2DD). To address the imbalance between CKD-positive and CKD-negative cases, oversampling techniques were applied before splitting the dataset into training (70%), validation (12%), and testing (18%). Using the Arkangel AutoML platform, 424 candidate models were generated, including decision trees, random forests, support vector machines, XGBoost, and deep neural networks. Models were prioritized based on predefined criteria: sensitivity \u0026gt;90%, followed by AUC, precision, specificity, and F1 score.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e The final model was a decision tree trained in a non-stratified sample with the SMOTE augmentation technique. Sensitivity was 90.2%, specificity 92.7%, precision (PPV) 89%, and AUC 91.4%. Binary regression demonstrated the statistical relevance of all the model’s features in predicting CKD risk in our sample. SHAP analysis identified age and diabetes duration as the most influential features in the final ML model.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e: A decision tree model trained with eight routine clinical variables accurately identified individuals at risk of CKD, achieving high sensitivity and balanced performance without requiring specialized tests. This approach is feasible for large-scale screening in low-resource settings and can be integrated into electronic health records to prioritize confirmatory diagnostics and timely care. It also represents one of the first approximations to CKD diagnosis using ML models trained exclusively on Latin American data.\u003c/p\u003e","manuscriptTitle":"Search AI: a Machine Learning algorithm for chronic kidney disease risk detection using eight readily available clinical features","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-13 14:46:38","doi":"10.21203/rs.3.rs-7888843/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"171951971434421618520275332981548350579","date":"2026-01-07T13:05:38+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-03T07:16:45+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-03T07:13:08+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-23T09:39:50+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-22T19:44:25+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Nephrology","date":"2025-10-22T19:40:12+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-nephrology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bnep","sideBox":"Learn more about [BMC Nephrology](http://bmcnephrol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bnep/default.aspx","title":"BMC Nephrology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9c338e6a-a1ab-46c2-a804-f099bfef0375","owner":[],"postedDate":"November 13th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-11-13T14:46:38+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-13 14:46:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7888843","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7888843","identity":"rs-7888843","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}