Using Explainable Machine Learning for Early Detection of Diabetic Kidney Disease in Rwandan Diabetic Patients

Using Explainable Machine Learning for Early Detection of Diabetic Kidney Disease in Rwandan Diabetic Patients | 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 Using Explainable Machine Learning for Early Detection of Diabetic Kidney Disease in Rwandan Diabetic Patients Silas Majyambere, Tony Lindgren, Celestin Twizere, Fiacre Rugamba Rugero This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8108979/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract The global prevalence of diabetes is increasing, often leading to complications such as diabetic kidney disease (DKD). Uncontrolled blood glucose and hypertension are key risk factors that progressively impair kidney function, potentially resulting in kidney failure or end-stage renal disease (ESRD). Early detection of DKD is crucial but challenging due to its asymptomatic onset. This study employs explainable artificial intelligence (XAI) to predict DKD risk in diabetic patients using tree-based ensemble models and SHAP (Shapley Additive exPlanations), leveraging the MIMIC-IV dataset and a dataset from hospitals in Rwanda. Among the models used, Random Forest demonstrated superior performance, achieving accuracies of 87.97% on MIMIC-IV and 91.70% on the Rwandan dataset. Models were evaluated using multiple metrics, including ROC-AUC and calibration curves. SHAP provided both global and individual-level explanations, with predictions validated using estimated glomerular filtration rate (eGFR) values. Our findings highlight the promising potential of integrating predictive modeling with explainability to develop transparent and trustworthy tools for early detection of DKD, with potential applications in clinical workflows. Diabetic Kidney Disease Diabetes Management Ensemble Models Explainable Artificial Intelligence Machine Learning SHAP Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 1 INTRODUCTION Diabetes Mellitus (DM) a chronic non-communicable disease resulting from the metabolic disorder characterized by high blood sugar levels, remains the all-time global health concern. According to the International Diabetes Federation (IDF), an estimated 537 million people were living with diabetes in 2021, and the highest proportions of undiagnosed diabetes (53.6%) were found in Africa, and diabetes cases are expected to rise to 783 million by 2045 [ 1 ]. Uncontrolled diabetes has no limit in causing burdens to diabetic patients. The progression of diabetes brings new complications, including Diabetes Kidney Disease (DKD), also known as diabetes nephropathy. DKD is a progressive condition that impairs the kidneys' ability to filter blood, primarily due to prolonged high blood glucose and hypertension, both of which are prevalent in diabetic patients, as reported in [ 2 ]. It is a leading cause of Chronic Kidney Diseases (CKD), including renal failure and the End Stage Renal Disease (ESRD) [ 3 ]. In its early stages, DKD is typically asymptomatic, making early detection difficult. As the disease progresses, symptoms such as frequent urination, foamy urine, itching, and general discomfort may appear. DKD also imposes a significant financial burden, especially in advanced stages where treatment may involve dialysis or kidney transplantation [ 4 ]. These interventions are not only costly but also place a considerable strain on healthcare systems and patients, particularly in low-resource settings. Machine learning (ML) has demonstrated strong potential in uncovering complex, non-linear patterns within data [ 5 ]. Its application in disease detection is particularly promising, offering valuable support for the early diagnosis of multifactorial conditions such as diabetes and its associated complications. The growing availability of patient health data provides an opportunity to leverage ML for the early identification of DKD, which can significantly contribute to timely prevention and intervention strategies in diabetes management [ 6 ]. However, the adoption of AI-based tools in diabetes management has been relatively slow, largely due to the black-box nature of many ML models, which rely on complex algorithms that lack transparency [ 7 ]. Although these models often deliver high predictive performance, their limited interpretability presents challenges for clinical integration and decision-making [ 8 ]. To address this, our study employs SHAP (SHapley Additive exPlanations) [ 9 ], a post-hoc model interpretation method, to enhance the transparency of the best-performing ML model. By providing clear insights into the risk factors influencing predictions, SHAP helps bridge the gap between predictive performance and clinical interpretability, enabling clinical decisions supported by evidence from the data and thereby improving trust and facilitating medical adoption. Patients with either type 1 or type 2 diabetes who have had the condition for five years or longer are considered at elevated risk for developing Diabetic Kidney Disease (DKD) and are typically recommended to undergo regular screening. However, because the precise onset of diabetes is not known and early stages of DKD are asymptomatic, there is a critical need for explainable machine learning approaches to support early detection, particularly among high-risk diabetic populations. Most diabetes-related complications, including DKD, are preventable through early detection, timely intervention, and effective management of blood glucose and blood pressure. This study proposes an explainable machine learning-based predictive tool for the early identification of diabetic patients at risk of developing DKD. The models were trained on the publicly available Medical Information Mart for Intensive Care (MIMIC)-IV dataset [ 10 ] and clinical datasets from three hospitals in Rwanda. Among five trained models, the best-performing one was further interpreted using the SHAP algorithm, which provides transparency into how individual predictions are made. This level of explainability enables clinicians to evaluate the model's outputs in the context of clinical judgment, thereby enhancing trust and potential adoption in real-world healthcare settings. In the context of Rwandan healthcare, explainable ML offers a dual advantage. First, it enhances the accuracy of early DKD detection by identifying individuals at risk before clinical symptoms manifest. Second, it supports data-driven decision-making by highlighting the key risk factors, thus allowing for targeted preventive strategies that preserve healthcare expenditure and maintain optimal care for diabetes patients. 1.1 Kidney Functions Figure 1 illustrates the four key stages of healthy kidney function [ 11 ]. In stage 1, blood containing various waste products enters the kidney, where the glomerular filtration system selectively filters out waste while retaining essential proteins and cells. Stage 2 involves the reabsorption of vital substances, including glucose, amino acids, sodium, potassium, and water, which helps maintain fluid and electrolyte balance in the body. In stage 3, the kidneys secrete additional waste and regulate acid-base balance, directing the final waste products to the bladder for excretion. In the final stage, the filtered, clean blood is returned to the bloodstream via the renal veins. Elevated blood glucose and hypertension damage the small blood vessels in the glomerular filtration membrane, compromising its selective permeability and allowing large molecules, such as proteins, to pass through. This progressively reduces filtration efficiency, leading to waste accumulation in the body. As a result, diabetic patients may experience symptoms such as itchiness, Frequent urination, fatigue, dry mouth, loss of appetite, and foamy urine [ 13 ]. 1.2 Rwanda Healthcare System Rwanda has undergone a significant transformation in its healthcare sector, marked by substantial improvements in infrastructure, including the adoption of information and communication technology (ICT) to support digital health and enhance the delivery of health services. Key reforms have included the expansion of the healthcare workforce, the decentralization of service provision, and the implementation of universal health coverage through the Community-Based Health Insurance (CBHI) program, which enables citizens to access healthcare services at a small cost. The national healthcare system is structured across five hierarchical levels, as illustrated in Fig. 2 . At the top are the National Referral Hospitals (NRH) and University Teaching Hospitals (UTH), while the bottom level comprises basic healthcare services delivered by Community Health Workers (CHWs). 1.3 DKD Diagnosis Diabetic Kidney Disease (DKD) is a form of Chronic Kidney Disease (CKD) that results from diabetes. Its diagnosis follows similar procedures to those used for CKD, beginning with the evaluation of clinical symptoms and laboratory assessments, including urine analysis and measurement of Glomerular Filtration Rate (GFR). The presence of macromolecules such as albumin and proteins in the urine serves as a primary indicator of DKD. Figure 3 presents the measurements of Albuminuria and GFR, along with their diagnostic interpretation. Three stages of Albuminuria (A1-A3) and DKD stages based on GFR (G1-G5). Clinically, Diabetic Kidney Disease (DKD) is characterized by a decline in the kidneys’ ability to filter waste products from the bloodstream, which is reflected in a reduced estimated Glomerular Filtration Rate (eGFR) and persistently elevated levels of albuminuria (> 300 mg/g of creatinine). 1.4 DKD Treatment and Prevention Strategies The management of Diabetic Kidney Disease (DKD) typically involves lifestyle modifications, as well as the regulation of blood glucose and blood pressure levels. Effective prevention relies on the early detection of DKD in high-risk diabetic individuals. This study aims to predict the risk of DKD among diabetic patients using machine learning by leveraging demographic information, vital signs, and laboratory results extracted from Electronic Health Records (EHRs). This study is guided by two primary objectives: (1) to train and evaluate machine learning models using a publicly available dataset (MIMIC-IV) and diabetic kidney disease (DKD) dataset obtained from three Rwandan hospitals, and (2) to interpret the most effective model for early DKD detection using SHAP (SHapley Additive exPlanations) techniques. The rest of the paper is structured as follows: Section 2 details the methodology, including data sources, preprocessing, model development, and evaluation techniques. Section 3 presents the experimental results. Section 4 discusses the implications of the findings in the context of DKD prediction and healthcare in Rwanda. Finally, Section 5 concludes the study and outlines directions for future research. 2 METHODOLOGY This study was conducted in five phases: the study design began with data extraction and dataset preprocessing, followed by the training of five tree-based ensemble models, evaluation of model performance, and interpretation of model predictions using SHAP methods on both the public dataset and the dataset from Rwanda. The primary purpose of training and testing machine learning models on datasets from two sources is to analyze how the models generalize to data recorded in different contexts and distributions. The overall study design, including all research phases, is illustrated in Fig. 4 . 2.1 Study Design This study utilized two datasets: one extracted from the Medical Information Mart for Intensive Care (MIMIC-IV) database provided by the Beth Israel Deaconess Medical Center, and another comprising aggregated data from three hospitals in Rwanda. Data preprocessing techniques were applied to prepare both datasets for machine learning applications. Following preprocessing, each dataset was partitioned into training and testing subsets. Five tree-based ensemble machine learning models were developed, comprising one bagged ensemble model and four boosted ensemble models. These models were trained on the training data and evaluated on the test data using five performance metrics. The best-performing model was subsequently interpreted using SHAP methods. 2.2 Data Extraction Data were extracted from the Electronic Health Record (EHR) systems. Patient data were identified using the 10th Revision of the International Classification of Diseases (ICD-10) codes for type 1 and type 2 diabetes mellitus. Specifically, the format E10.XXX was used for Type 1 Diabetes Mellitus (T1DM), and E11.XXX for Type 2 Diabetes Mellitus (T2DM), where X denotes a digit from 0 to 9 indicating the stage or complication of the disease [ 16 ]. Diabetic Kidney Disease (DKD) cases were identified using codes E10.2X for T1DM patients with kidney complications and E11.2X for T2DM patients with kidney disease. For the Rwandan dataset, the same data extraction method was applied. In instances where the relevant ICD-10 code was not available, clinical notes accompanying the diagnosis were reviewed, and in most cases, they provided sufficient information to confirm DKD status. The MIMIC-IV dataset contains data of diabetic patients recorded between 2008 and 2019. Diabetic Kidney Disease (DKD) is a progressive complication that typically affects both type 1 and type 2 diabetic patients after at least five years following a confirmed diabetes diagnosis. However, due to the undiagnosed diabetes, the precise onset dates are not known; some patients may develop DKD earlier within one year of diagnosis or even at the time of initial diabetes screening. To address this uncertainty and improve the reliability of the analysis, only patients with a minimum of one year of follow-up were included in the study. The final datasets consisted of 949 samples from the MIMIC-IV database and 810 samples from the Rwandan dataset. The Rwandan data were collected from three hospitals: 383 samples from the National Referral Hospital (CHUK), 204 from a Teaching University Hospital (Ruhengeri Hospital), and 223 from a District Hospital (Nyamata Hospital). Both datasets included 20 predictive features for DKD, comprising 16 numerical and 4 categorical variables, along with a single class label indicating the presence or absence of DKD. The dataset from Rwanda contains data of diabetes patients recorded between 2012 and 2024 2.3 Data Preprocessing Data preprocessing is a crucial step in any machine learning pipeline, as the performance of predictive models, such as those used to detect DKD risk, largely depends on the quality and relevance of the data used during training and evaluation. Figure 5 illustrates the correlation between various features and the presence of DKD in the MIMIC-IV dataset. In Fig. 5 , lighter orange shades represent features with weaker correlations to DKD risk, while darker orange shades indicate stronger correlations. Among the features, Complication, Urea (Blood Urea Nitrogen), and Creatinine exhibit the strongest associations with the risk of developing DKD in this dataset. Figure 6 and Fig. 7 illustrate the class imbalance present in both datasets, where the minority class corresponds to patients at risk of developing Diabetic Kidney Disease (DKD). This imbalance reflects a common clinical reality, as DKD typically affects approximately one-third of the diabetic population. To address this issue, the Synthetic Minority Oversampling Technique (SMOTE) [ 17 ] was applied. SMOTE generates new synthetic instances of the minority class through interpolation between existing samples, thereby increasing the representation of the minority class and enhancing the model's ability to learn from imbalanced data. The dataset contained missing values across all numerical features except for the Age feature. The highest proportion of missing data was observed in Total Cholesterol (TC) at 38.04%, followed by Glycated Hemoglobin (HbA1c) at 28.33%. To address this, missing values were imputed using the mean of each respective feature. Additionally, samples containing extreme outlier data points that significantly deviated from the overall distribution were considered irrelevant and subsequently removed. For instance, records with a Body Mass Index (BMI) of 3658.5 and a Blood Glucose (BG) level of 1120 mg/dL were excluded from the analysis. All numerical features were normalized using the Min-Max scaling technique to ensure that all variables were on the same scale, thereby preventing the models from under learning features with smaller value ranges. To prevent data leakage and ensure a fair evaluation of model performance, the Synthetic Minority Oversampling Technique (SMOTE) was applied exclusively to the training dataset, while the test dataset remained in its original form. The MIMIC-IV DKD dataset comprises 274 patients diagnosed with Diabetic Kidney Disease (DKD) and 675 patients without DKD. Among the total cohort, 577 patients present with at least one diabetes-related complication other than nephropathy, while 372 have no recorded complications. In terms of diabetes type, 603 patients have Type 2 Diabetes Mellitus (T2DM) and 346 have Type 1 Diabetes Mellitus (T1DM). Additionally, 306 patients are diagnosed with chronic hypertension, whereas 643 have no history of hypertension. The dataset also includes 482 male and 467 female patients. The dataset from Rwanda exhibits an imbalanced distribution of class labels, comprising 256 patients diagnosed with Diabetic Kidney Disease (DKD) and 554 patients without DKD. The dataset counts 164 patients with diabetes-related complications and 646 without complications. Additionally, 312 patients present with hypertension (HT), whereas 498 are without this condition. In terms of diabetes type, there are 558 patients with Type 2 Diabetes Mellitus (T2DM) and 252 with Type 1 Diabetes Mellitus (T1DM). The gender distribution includes 533 female and 277 male patients. Figure 8 illustrates the distribution of DKD cases by age group. Figure 8 (a) shows that DKD steadily increases with age. Figure 8 (b) reveals a peak incidence between the ages of 30 and 69. The number of DKD cases slightly declines after the age of 70, which corresponds with Rwanda’s life expectancy of 69.1 years as reported by the National Institute of Statistics of Rwanda (NISR) in 2022 [ 18 ]. Table 1 presents the descriptive analysis of the numerical variables in the Rwandan DKD dataset. The analysis reveals that several variables have a significant proportion of missing values, exceeding 10%. Specifically, Body Mass Index (BMI) has 16.54% missing data, glycated hemoglobin (HbA1c) 12.96%, urea 14.2%, systolic blood pressure (SBP) 11.85%, diastolic blood pressure (DBP) 11.98%, and total cholesterol (TC) 19.88%. The missing values were handled using mean imputation methods. Each missing value was replaced by the column mean value. Table 1 Descriptive analysis of the numerical variables in the Rwandan DKD dataset Feature Name Counts Mean Std Min Max Missing Values (%) Normal Range Age 810 56.51 16.01 11 95 0 N/A BMI 676 27.03 5.07 15.11 48.49 16.54 18.5–24.9 HbA1C 705 10.15 2.65 5.00 19.10 12.96 < 5.7 BG 805 229.45 106.32 53.69 833.88 0.62 70–100 Creatinine 766 1.16 0.99 0.20 7.96 5.43 0.6–1.3 Urea 695 10.14 4.50 3.08 37.00 14.20 10–20 Sodium 789 138.05 4.64 121.00 155.00 2.59 135–145 Potassium 749 4.41 0.78 2.70 8.90 2.72 3.6–5.2 Hemoglobin 749 13.94 2.52 6.70 21.40 7.57 13.2–16.6 RBC 758 4.74 1.26 1.51 10.80 6.42 4.2–6.1 WBC 757 7.59 3.87 2.36 30.66 6.54 4.5–11.0 Neutrophils 758 54.83 15.35 16.80 92.00 6.42 40–60 Lymphocytes 753 32.79 13.94 10.00 69.90 7.04 20–40 SBP 714 137.56 25.99 92.00 210.00 11.85 < 120 DBP 713 80.90 14.70 38.00 145.00 11.98 < 80 TC 649 5.13 1.58 0.98 11.30 19.88 3.9–5.5 The descriptive analysis reveals that some diabetic patients exhibit vital signs and laboratory results that deviate significantly from clinically accepted normal ranges. Since age does not have a standardized normal range, it was marked as not applicable (N/A) in this context. To further explore these deviations, a box plot of the numerical features is presented in Fig. 9 , providing a visual representation for identifying potential outliers within the dataset. Figure 9 displays box plots that highlight data points lying above or below the whiskers, suggesting the presence of potential outliers. Upon further analysis, it was observed that patients diagnosed with DKD often exhibited abnormal clinical values. While these abnormal values were retained due to their clinical relevance, extreme outliers, which fall far outside the overall data distribution, were excluded. For instance, blood glucose levels exceeding 600 mg/dL were considered extreme outliers and were subsequently removed from the dataset. The final Rwandan dataset comprises 801 samples. Prior to model training, the dataset was normalized using the Min-Max scaling technique. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training subset, while the test subset remained unaltered to ensure an unbiased evaluation. 2.4 Machine Learning Models Diabetes Kidney Disease (DKD) risk prediction is a classification task that uses supervised machine learning technique. In this study, ensemble machine learning models were employed to predict the risk of DKD among diabetes patients, given their demonstrated effectiveness in handling complex datasets, particularly those comprising both numerical and categorical data. These models are widely recognized for their strong performance in medical applications. Specifically, we utilized one bagging-based model (Random Forest) and four boosting-based models: Extreme Gradient Boosting (XGBoost), CatBoost, AdaBoost, and LightGBM. Random Forest. Random Forest is a bagging-based ensemble model that leverages bootstrapping and aggregation techniques for learning. It constructs multiple decision trees that are trained in parallel, each trained on randomly selected subsets of the original dataset, using either instance-based or feature-based sampling. This allows individual samples or features to appear multiple times within a training subset. The model aggregates the predictions from all individual trees to produce a final output by averaging in regression tasks and majority voting in classification tasks. Random Forest has been widely adopted in medical research, particularly for disease detection. For example, it was utilized in [ 19 ] for predictive modeling of early-stage diabetes, achieving a classification accuracy of 97.03%. Extreme Gradient Boosting (XGBoost) Extreme Gradient Boosting (XGBoost) is a sequential ensemble learning method that builds decision trees iteratively, with each subsequent tree aiming to correct the prediction errors of its predecessors through gradient-based optimization. As an open-source library, XGBoost implements a distributed gradient-boosted decision tree algorithm designed for both classification and regression tasks. It integrates regularization techniques with advanced optimization strategies to enhance predictive accuracy while reducing training time. The core learner in XGBoost is the Classification and Regression Tree (CART). Due to its high accuracy and computational efficiency, XGBoost is widely applied in solving problems involving structured data. For instance, in [ 20 ], XGBoost was used to predict chronic kidney disease, achieving a classification accuracy of 93.29% and ranking it as the best model. CatBoost Model CatBoost is an open-source gradient boosting library specifically designed to handle categorical features effectively without extensive preprocessing or transformation. Developed by Yandex, it is particularly well-suited for machine learning tasks involving heterogeneous data types. CatBoost constructs its model iteratively using decision trees as weak learners, optimizing performance through gradient boosting. A key advantage of CatBoost lies in its use of ordered boosting and random permutations, which help prevent overfitting and improve generalization. Additionally, it incorporates gradient-based optimization to enhance predictive accuracy on complex datasets. In [ 21 ], CatBoost was applied to predict diabetic kidney disease in patients with type 2 diabetes, achieving a classification accuracy of 75.5%. AdaBoost Model Adaptive Boosting (AdaBoost) is a boosting ensemble algorithm that combines multiple weak decision tree classifiers to form a single, strong predictive model. It operates by training weak learners sequentially, starting from one decision tree, also known as a decision stump, with each iteration placing increased emphasis on the instances that were previously misclassified. This adaptive weighting mechanism enables the model to focus on difficult cases, thereby improving overall performance. AdaBoost is particularly effective for binary classification and regression tasks, with its strength lying in its iterative learning approach. While it performs well on smaller datasets, the sequential nature of its training process can lead to longer training times when applied to larger datasets. In [ 22 ], AdaBoost was utilized for predicting diabetic kidney disease staging, achieving a classification accuracy of 83.5%. LightGBM Algorithm Light Gradient Boosting Machine (LightGBM) is an open-source gradient boosting framework developed by Microsoft for classification and regression tasks. It constructs decision trees using a leaf-wise strategy, which enhances training speed and accuracy. LightGBM’s efficiency is largely attributed to two key techniques: Gradient-Based One-Side Sampling (GOSS), which improves computational efficiency by prioritizing instances with high gradients, and Exclusive Feature Bundling (EFB), which reduces feature dimensionality by combining mutually exclusive features. These innovations enable LightGBM to train effectively on large datasets with reduced resource demands. In [ 23 ], LightGBM achieved an area under the curve (AUC) score of 0.815 in predicting diabetic kidney disease, outperforming other models used in the study. 2.5 Models Performance Evaluation The selected ensemble models were trained on 70% of each dataset, with the remaining 30% used for testing. Model performance was evaluated using five metrics: accuracy, sensitivity, specificity, F1-score, and area under the curve (AUC). For each dataset, the confusion matrix of the best-performing model was used to visualize its ability to identify the risk of diabetic kidney disease (DKD). Additionally, ROC-AUC curves were plotted to illustrate the trade-off between the true positive rate (TPR) and false positive rate (FPR), and a calibration curve was used to visualize the reliability of models’ confidence in predicting DKD risk. The equations (1) to (4) explain how accuracy, sensitivity, specificity, and F1-score are calculated. $$\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:Accuracy=\:\:\frac{TP\:+\:TN}{TP\:+\:TN\:+FP\:+\:FN}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(1\right)$$ $$\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:Sensitivity=\:\:\frac{TP}{TP\:+\:FN}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(2\right)$$ $$\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:Specificity=\frac{TN}{TN\:+\:FP}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(3\right)$$ $$\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:F1-score\:\:=\frac{2*TP}{2*TP\:+\:FP\:+\:FN}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\left(4\right)$$ Here, TP (true positives) denotes correctly identified positive cases, TN (true negatives) indicates correctly identified negative cases, FP (false positives) refers to negative cases incorrectly classified as positive, and FN (false negatives) represents positive cases incorrectly classified as negative. Sensitivity reflects the model’s ability to detect positive cases, while specificity measures its ability to correctly identify negatives. 2.6 Explainable Artificial Intelligence (XAI) Methods Ensemble models offer an improved predictive performance by combining multiple learners, but often result in increased complexity, making them difficult to interpret. In critical domains such as medicine, the lack of transparency in black-box models can hinder their adoption, as a model that does not offer details of its predictions may lead to incorrect clinical decisions and compromise patient safety. Explainable Artificial Intelligence (XAI) addresses this challenge by developing methods to interpret model behavior. In this study, a model-agnostic approach based on feature attribution is employed. Specifically, Shapley Additive exPlanations (SHAP), which use the principles of game theory to compute the contribution of each feature to the prediction. While SHAP can be computationally intensive, optimized variants have been developed to improve efficiency. We utilize FastTreeSHAP [ 24 ], an advanced version of TreeSHAP, to compute SHAP values for tree-based models. The original SHAP [ 25 ] is presented in Eq. ( 5 ). $$\:{\:\phi\:}_{i\left(f\right)}=\sum\:_{S\subseteq\:\cup\:N\backslash\:\left\{i\right\}}\frac{\left|s\right|!\left(N-\left|s\right|-1\right)!}{N!}\left[{f}_{x}\left(s\cup\:\left\{i\right\}\right)-{f}_{x}\left(s\right)\right]$$ 5 Here, \(\:S\) is a subset of features excluding feature \(\:i\) , and \(\:N\) is the total number of features. \(\:{\varphi\:}_{i}\left(f\right)\) represents the contribution of feature \(\:i\) , calculated as the difference between the total prediction and the prediction without feature \(\:i\) . 2.7 Related work Recent research has increasingly focused on leveraging medical data and machine learning to address non-communicable diseases such as diabetes and its complications. In [ 26 ], XGBoost was used to predict diabetic nephropathy using 548 patient records from SAHDMU Hospital (2018–2019). After applying LASSO for feature selection, the model achieved an AUC of 0.966. Calibration curves and SHAP values were employed to enhance the clinical interpretability of the XGBoost model. In [ 27 ], a study on diabetic kidney disease (DKD) prediction was conducted using data from 1,177 diabetic patients at Beijing Pinggu Hospital (2013–2017), including 263 DKD cases and 914 non-DKD cases. The Random Forest classifier achieved an accuracy of 89.83%, suggesting its potential for initial DKD screening. Feature importance analysis identified creatinine and blood urea nitrogen as key predictors. CatBoost was employed in [ 28 ] to predict both chronic kidney disease (CKD) and diabetes. Using a public dataset comprising 202 samples and 29 features, the model achieved 95% accuracy in predicting CKD. For diabetes prediction, it attained 99% accuracy on a separate dataset containing 520 samples and 17 features. SHAP was used for feature effect analysis. The Random Forest model demonstrated superior performance in [ 29 ] at predicting chronic kidney disease (CKD), achieving 98.8% accuracy on a dataset comprising 455 samples and 25 features. Machine learning models were used in [ 30 ] to predict the five-year risk of developing diabetic kidney disease (DKD) following a diagnosis of type 2 diabetes (T2DM). Using a dataset of 87,973 records from USA diabetes patients (2007–2020), the Random Forest model achieved an AUC of 0.75. In [ 31 ], the Random Forest model was trained on a dataset comprising 10,064 samples extracted from Australian electronic health records (EHR) and validated on 597 samples from a Japanese dataset. It achieved a classification accuracy of 92.6% on the Australian data and 73.8% on the Japanese data. 3 Results 3.1 Experimental Setup The experiments were conducted using Python and relevant libraries, including Pandas, NumPy, Matplotlib, Seaborn, Scikit-learn, and FastShap for model explainability, as well as other required libraries. The workflow began with exploratory data analysis and feature engineering to prepare the final datasets. In consultation with a clinical nurse experienced in chronic kidney disease (CKD), relevant features were selected for model development. Extensive preprocessing was performed to ensure data quality suitable for training ensemble models to predict the risk of diabetic kidney disease (DKD). Model performance was evaluated using multiple metrics, and the best-performing model was interpreted using SHAP to provide both local and global insights into the decision-making process. 3.2 DKD Prediction The goal of this study is to predict the risk of diabetic kidney disease (DKD) by leveraging electronic health record (EHR) data and machine learning for the early identification of high-risk diabetic patients, thereby enabling timely interventions to prevent or delay severe complications and improve disease management. The predictive performance of all models on the MIMIC-IV dataset is summarized in Table 2 , where all models demonstrate good performance in predicting DKD risk, with an accuracy above 80% for each model. The best model is Random Forest and the model with the lowest performance is AdaBoost with 82.71% accuracy. Random Forest outperformed other models in three out of five metrics, achieving an accuracy of 87.97% and an AUC of 0.948. Table 2 Models evaluation on the test subset of the MIMIC-IV dataset Classifier Accuracy Sensitivity Specificity F1-score AUC-score RandomForest 87.97 75.32 93.12 78.38 0.948 XGBoost 85.34 77.92 88.36 75.47 0.929 CatBoost 86.09 77.92 89.42 76.43 0.945 LightGBM 87.22 76.62 91.53 77.63 0.931 AdaBoost 82.71 49.35 96.30 62.30 0.924 The same five ensemble models were applied to the Rwandan dataset for predicting the risk of diabetic kidney disease (DKD). All models showed slight improvements in performance, with Random Forest achieving the highest accuracy of 91.70%, and AdaBoost the lowest at 84.65%. Table 3 presents a summary of the performance of each model across all evaluation metrics. Table 3 Models evaluation on test subset of Rwanda DKD dataset Classifier Accuracy Sensitivity Specificity F1-score AUC-score RandomForest 91.70 88.16 93.33 87.01 0.960 XGBoost 77.18 77.63 76.97 68.21 0.861 CatBoost 89.21 84.21 91.52 83.12 0.951 LightGBM 86.32 80.26 89.09 78.71 0.931 AdaBoost 84.65 85.53 84.24 77.84 0.897 Random Forest achieved the highest performance across all five evaluation metrics for DKD risk prediction using the Rwandan dataset. In addition to these metrics, Fig. 10 presents the ROC-AUC curves to further illustrate model performance. Figure 10 (a) displays the ROC-AUC curves for the five ensemble models on the MIMIC-IV dataset, while Fig. 10 (b) shows the corresponding curves for the Rwandan dataset. The Random Forest model demonstrated superior performance on both datasets. Figure 11 presents its confusion matrix, which was used to compute the evaluation metrics. To further evaluate the strong predictive performance of the tree-based models on the Rwandan DKD dataset, calibration curves were plotted, as shown in Fig. 12 . The dashed black line represents perfect calibration, where the predicted probabilities align with the actual outcomes. Models below this line are overconfident, while those significantly above are underconfident. A well-calibrated model closely follows the diagonal. Beyond the 70% predicted probability threshold for the positive class (DKD), all models align closely with the diagonal except AdaBoost, which remains consistently below, indicating overconfidence. 3.3 SHAP Explanations The effectiveness of the tree-based ensemble models used in this study was further assessed using the XAI explainability method. The FastTreeSHAP approach was used to investigate how the best model on both datasets makes a decision on identifying the diabetic patients at risk of DKD development. Figure 13 provides a global view of feature contribution to the model decision using the MIMIC-IV dataset. According to the MIMIC-IV dataset, the presence of complications related to diabetes is the leading predictor of diabetes kidney disease risk, followed by creatinine. Gender is the least predictor of DKD risk. Figure 14 highlights the top 15 features contributing to the risk of DKD among diabetes patients in Rwanda. SHAP methods applied to a Random Forest trained on the Rwanda DKD dataset revealed the top 15 features highly contributing to the model's prediction of DKD risk. The blood urea nitrogen (Urea) is the leading predictor of DKD, followed by Hemoglobin. Among 15 features linked to DKD risk, diastolic blood pressure is ranked the lowest. SHAP methods were also used to assess the model's behavior for individual diabetic patients, providing a clear clinical interpretation of why a particular patient is predicted to be at risk or not at risk of DKD. Figure 15 illustrates the model interpretation at the patient level. Figure 15 (a) details the reason why the patient with index 19 in the test set of the Rwanda dataset was predicted at high risk of DKD, and Fig. 15 (b) shows that the patient with index 11 in the test set of the Rwanda dataset is free from DKD risk. 4 Discussion Experimental results demonstrated that the Random Forest model outperformed other models on both the MIMIC-IV and the dataset from three hospitals in Rwanda. Configured as RandomForestClassifier(random_state = 42, n_estimators = 300, max_depth = 6, bootstrap = True), it achieved an accuracy of 87.97% on the MIMIC-IV dataset, ranking highest in three out of five evaluation metrics. On the Rwanda dataset, it achieved 91.70% accuracy and led in all five metrics. These findings suggest that the model can serve as a reliable tool for the initial screening of diabetic kidney disease (DKD) at the onset of patient follow-up. The MIMIC-IV dataset exhibited higher missing value rates in key features such as HbA1c (28%), total cholesterol (TC) (38%), and BMI (18%) compared to the Rwanda dataset, which may have contributed to the slightly lower model performance. Missing values were imputed using the mean method. The Random Forest model achieved higher sensitivity and specificity on the Rwanda dataset, indicating a strong ability to correctly identify both DKD and non-DKD cases. These metrics are critical: failing to detect DKD may allow disease progression to advanced stages, while false positives could lead to unnecessary treatment and potential side effects. Model reliability was further confirmed through calibration analysis, with the Random Forest curve closely aligning with the diagonal, indicating good calibration. SHAP analysis of the Random Forest model provided valuable insights to support clinical interpretation and build trust in the model’s predictions. SHAP identified the most influential features contributing to DKD risk, with three of the top five features consistent across both datasets and 14 of the top 15 shared, though ranked differently, demonstrating the model’s generalizability across datasets with varying distributions. SHAP also enabled interpretation at the individual patient level. As illustrated in Fig. 15 (a), features highlighted in red contribute to increased DKD risk, while those in blue reduce it. For the patient examined, abnormal values in hemoglobin (11.8), creatinine (2.4), blood glucose (217.12), RBC (7.15), and neutrophils (64.5) support the high-risk prediction. Additionally, the patient’s eGFR was calculated using the method described in [ 26 ], yielding a value of 20. Since an eGFR ≥ 90 indicates normal kidney function, this low value supports the model’s predicted high risk of DKD, with a probability of 0.946, as indicated by f(x) in Fig. 15 (a). As shown in Fig. 3 , this corresponds to stage G4 DKD, indicating severely reduced kidney function. The SHAP interpretation in Fig. 15 (b) explains why the patient was predicted not to be at risk of DKD. Further analysis revealed that the patient described in Fig. 15 (b) is a 58-year-old male with a normal creatinine level of 0.69, as indicated in Table 1 . Additionally, the eGFR, calculated using the method in [ 26 ], was 104.64, confirming a normal kidney function and supporting the model’s prediction. 5 Conclusion and Future Work Early detection of diabetic kidney disease (DKD) benefits both patients and healthcare providers, as DKD is a progressive condition often driven by uncontrolled blood glucose and hypertension, which impair kidney function. This study demonstrates that explainable machine learning can effectively utilize electronic health record (EHR) data to identify individuals at high risk of DKD. Five tree-based ensemble models were developed and evaluated on both the MIMIC-IV and Rwanda datasets. Due to class imbalance, the SMOTE technique was applied to balance the data. Among the models, Random Forest achieved the highest performance, with accuracies of 87.97% on MIMIC-IV and 91.70% on the Rwanda dataset. All models were evaluated using multiple metrics, including ROC-AUC and calibration curves, to ensure robustness and interpretability. The best-performing model was further analyzed using SHAP methods to interpret its predictions. SHAP was used to visualize both global feature importance and individual-level explanations through waterfall plots. A ranked list of the top 15 features contributing to DKD risk in the Rwandan dataset was generated. At the individual level, predictions were validated using estimated glomerular filtration rate (eGFR) values, which aligned with the model’s outputs, as shown in Fig. 3 . This study had two key objectives: (1) to train and evaluate machine learning models on a publicly available dataset and DKD data from three Rwandan hospitals, and (2) to interpret the most effective model using SHAP for early DKD detection. These objectives were successfully achieved. The findings of this study can be integrated into clinical workflows to support early detection of DKD in diabetic patients, contributing to more effective and timely care and improving patients' quality of life. We recommend that Rwandan hospitals enhance the completeness of diabetes patient records, particularly for key variables such as BMI, total cholesterol (TC), blood urea nitrogen (Urea), HbA1c, systolic blood pressure (SBP), and diastolic blood pressure (DBP), which showed over 10% missing values but were among the top 15 predictors of DKD in the Rwandan dataset. Future work will focus on implementing these findings in clinical decision-making processes at CHUK, Ruhengeri, and Nyamata hospitals. Declarations Authors contribution All authors contributed equally to the manuscript. Silas led the research design, data collection, experimentation, and writing. Tony and Celestin supervised the study, contributed to manuscript refinement, and analyzed the results. Fiacre, drawing on extensive experience in nursing, particularly in chronic kidney disease at the nephrology department, extracted data from OpenMRS and OpenClinic systems in three participating hospitals in Rwanda and assessed the clinical relevance of our study. Acknowledgement The authors gratefully acknowledge the support provided by the University of Rwanda and Stockholm University. Special thanks are due to the PhysioNet team, as well as the data managers from Ruhengeri, Nyamata, and CHUK hospitals, for their valuable support during the data collection process. Ethical Statement This study relied on secondary data derived from electronic health records (EHRs) and did not involve the use of any personally identifiable information. The MIMIC-IV dataset, which is publicly accessible, does not require ethical approval. The dataset from Rwanda consisted of secondary EHR data collected from three hospitals. Access to these data was granted under ethical approval reference CMHS/IRB/207/2024 by the Institutional Review Board of the College of Medicine and Health Sciences, University of Rwanda. All data processing and analysis were conducted in accordance with Rwandan personal data protection and privacy regulations, as well as the requirements of the Health Insurance Portability and Accountability Act (HIPAA) and the ethical principles outlined in the Declaration of Helsinki. Informed consent The study utilized secondary data extracted from electronic health record (EHR) systems to develop machine learning models for predicting diabetic kidney disease. As no data were collected directly from individual patients, informed consent was not required. Patient involvement was indirect, and the Ethics Review Boards granted a waiver of informed consent. The study was conducted in accordance with the principles outlined in the Declaration of Helsinki. Data availability statement The original contributions presented in the study are included in the article and supplementary materials. For further inquiries, please contact the corresponding authors. Conflict of interest The authors declare that there is no conflict of interest associated with this study. Funding This research was supported by the University of Rwanda and Stockholm University under the UR-Sweden Program for research and capacity building. Generative AI statement The authors declare that no Generative AI was used in the creation of this manuscript. Consent for publication Not applicable. References Ogurtsova K et al. IDF diabetes atlas: Global estimates of undiagnosed diabetes in adults for 2021. Diabetes Res Clin Pract, 183, 2022. Art 109118JD, Kabakambira P, Shumbusho G, Mujawamariya W, Rutagengwa M, Twagirumukiza. The Role of the Integrated District Hospital Based Non Communicable Diseases’ Clinics in Cardiovascular Disease Control: Preliminary Data from Rwanda. Dovepress. 2022;15:2107–15. Qiong B, Chunyan S, Wen T, Yike L. Machine learning to predict end stage kidney disease in chronic kidney disease. Sci Rep. 2022;12:8377. Levin A, Stevens PE. Early detection of CKD: the benefits, limitations and effects on prognosis. Nat Rev Nephrol. 2011;7:446–57. Alowais SA, Alghamdi SS, Alsuhebany N, Alqahtani T, Alshaya AI, Almohareb SN, et al. Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC Med Educ. 2023;23:689. Zhouyu G, Huating L, Ruhan L, Chun C, Yuexing L, Jiajia L, et al. Artificial intelligence in diabetes management: Advancements, opportunities, and challenges. Cell Rep Med. 2023;4:101213. Singh P, Goyal L, Mallick DC, Surani SR, Kaushik N, Chandramohan D, Simhadri PK. Artificial Intelligence in Nephrology: Clinical Applications, and Challenges. Kidney Med. 2025;7:100927. Preti LM, Ardito V, Compagni A, Petracca F, Cappellaro G. Implementation of Machine Learning Applications in Health Care Organizations: Systematic Review of Empirical Studies, JMIR, 2024, 26, 55897. Sridevi P, Meikandan PV, Upama PB, Rabbani M, Alam KS, Das D, et al. ML-Based Chronic Kidney Disease and Diabetes Prediction with Feature Effect Analysis Using SHAP. IEEE COMPSAC. 2024;48:843–50. Alistair EW, Lucas B, Lu S, Alvin G, Ayad S, Steven H, et al. Author Correction: MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10:1. Clinic M. The Essential Diabetes Book, 2nd Edition, ISBN-10: 0-8487-4339-3, 2014, p. 53. https:// . (Accessed on June 25, 2025). Clinic M. The Essential Diabetes Book, 2nd Edition, ISBN-10: 0-8487-4339-3, 2014, p. 42. Selden O, Rusingiza E, Balkrishnan R. Overview, Infrastructural Challenges, Barriers to Access, and Progress for Rwanda’s Healthcare System: A Review. Integr J Med Sci. 2025;12:745. Persson F, Rossing P. Diagnosis of diabetic kidney disease: state of the art and future perspective. Kidney Int Supplements. 2018;8:2–7. Dugan J, Shubrook J. International Classification of Diseases, 10th Revision, Coding for Diabetes, Clinical diabetes: a publication of the American Diabetes Association, 2017, 35, 232–238. Fernandez A, Garcia S, Herrera F. ISMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary. J Artif Intell Res. 2018;61:863–905. National Institute of Statistics of Rwanda (NISR), Fourth Population and Housing Census. 2012, 143–144. Yadu S, Chandra R, Sinha VK. Comparing Different Machine Learning Techniques in Predicting Diabetes on Early Stage, MDPI Engineering Proceedings, 2024, 62, 20. https://doi.org/10.3390/engproc2024062020 Ghosh SK, Khandoker AH. Investigation on explainable machine learning models to predict chronic kidney diseases. Nat Sci Rep. 2024. https://doi.org/10.1038/s41598-024-54375-4 . Liu XZ, Duan M, Huang HD, Zhang Y, Xiang TY, Niu WC, et al. Predicting diabetic kidney disease for type 2 diabetes mellitus by machine learning in the real world: a multicenter retrospective study. Front Endocrinol. 2023;14:1184190. https://doi.org/10.3389/fendo.2023.1184190 . Renuka P, Thilagavathi C, Sathiyanathan S, Murugesh L, Madhumitha M. Adaboost Machine Learning based Modelling to Predict Chronic Kidney Disease Staging. ICSTEM. 2024. 10.1109/ICSTEM61137.2024.1056088 . Dong Z, Wang Q, Yujing K, Zhang W, Hong Q, Liu C, Xiaomin L, et al. Prediction of 3-year risk of diabetic kidney disease using machine learning based on electronic medical records. J Translational Med. 2022;20:143. https://doi.org/10.1186/s12967-022-03339-1 . Muschalik M, Fumagalli F, Hammer B, Hullermeier E. Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles, 38th AAAI Conference on Artificial Intelligence, 2024, 1604, 14388–14396. Lundberg SM, Lee S. A Unified Approach to Interpreting Model Predictions. NIPS. 2017;31:4768–77. Jing-Mei Y, Yang L, Jun-Tang X, Jun-Tang Z, Guo-Wei Z, Zhong-Ze F, Lang Z. Explainable Machine Learning-Based Prediction Model for Diabetic Nephropathy. J Diabetes Res. 2024. 10.1155/2024/8857453 . Hongxia X, Yonghui K, Shaofeng T. Predictive Modeling of Diabetic Kidney Disease using Random Forest Algorithm along with Features Selection. ISAIMS. 2020;1:23. - https://doi.org/10.1145/3429889.3429894 . Sridevi P, Padmapriya M et al. ML-based Chronic Kidney Disease and Diabetes Prediction with Feature Effect Analysis Using SHAP, IEEE 48th Annual Computers, Software, and Applications Conference (COMPSAC), 2024, 48. Mohsin RS et al. Chronic Renal Disease Prediction using Clinical Data and Different Machine Learning Techniques, International Conference on Informatics and Software Engineering (IISEC), 2021, 2, 1–5. Allen A, et al. Prediction of diabetic kidney disease with machine learning algorithms, upon the initial diagnosis of type 2 diabetes mellitus. BMJ Open Diabetes Res Care. 2022;10. 10.1136/bmjdrc-2021-002560 . Reddy S, et al. Predicting chronic kidney disease progression using small pathology datasets and explainable machine learning models. Computer Methods and Programs in Biomedicine Update; 2024. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 26 Dec, 2025 Reviews received at journal 22 Dec, 2025 Reviewers agreed at journal 21 Dec, 2025 Reviews received at journal 08 Dec, 2025 Reviewers agreed at journal 06 Dec, 2025 Reviewers invited by journal 26 Nov, 2025 Editor invited by journal 21 Nov, 2025 Editor assigned by journal 19 Nov, 2025 Submission checks completed at journal 19 Nov, 2025 First submitted to journal 13 Nov, 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-8108979","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":552962959,"identity":"bb001d24-4160-447e-ae51-e75101de963a","order_by":0,"name":"Silas Majyambere","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA60lEQVRIie2OvQrCMBRGrwh2ic4p/j2BECnUoYKvkqJk0t3NzS6617cQhMyBDC4VVwsO7VIXBd3qIBjEwSnqJpgDF+5wz3c/AIPhN6FqhJqiwN8qJfqxAk8Fkc+UViAyfB3uG51gdomvU+gRYaWJTnEjyuw5z5xatFl55Sn4S4EcolUEZbjMpR/iEa9CBJQIVNI2dLcJs29cTsLmMbPz6FHskGuVHWVV9YVirMLRGApLAa7OUErS9+o8a4eIOR4aY38hkfOm2NCPT3zfxJZM45x0e5V1kJ61bwANXjPVXtTfKyz5JtNgMBj+njt0sE2JTunQCwAAAABJRU5ErkJggg==","orcid":"","institution":"Stockholm University","correspondingAuthor":true,"prefix":"","firstName":"Silas","middleName":"","lastName":"Majyambere","suffix":""},{"id":552962960,"identity":"10f8aebc-f52d-49c1-8bc3-af2714fda62f","order_by":1,"name":"Tony Lindgren","email":"","orcid":"","institution":"Stockholm University","correspondingAuthor":false,"prefix":"","firstName":"Tony","middleName":"","lastName":"Lindgren","suffix":""},{"id":552962961,"identity":"1caaefca-1c51-4fda-a923-f06512510f34","order_by":2,"name":"Celestin Twizere","email":"","orcid":"","institution":"University of Rwanda","correspondingAuthor":false,"prefix":"","firstName":"Celestin","middleName":"","lastName":"Twizere","suffix":""},{"id":552962962,"identity":"66bf28a8-9d12-4ec2-a739-72e8c85609c3","order_by":3,"name":"Fiacre Rugamba Rugero","email":"","orcid":"","institution":"University of Rwanda","correspondingAuthor":false,"prefix":"","firstName":"Fiacre","middleName":"Rugamba","lastName":"Rugero","suffix":""}],"badges":[],"createdAt":"2025-11-13 21:08:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8108979/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8108979/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97152081,"identity":"250c2afc-264d-4620-97bd-db74c82fea84","added_by":"auto","created_at":"2025-12-01 10:39:51","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5933939,"visible":true,"origin":"","legend":"","description":"","filename":"DKDPredictionSilasTonyCelestinFiacreBMC.docx","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/ae98a2455ad2b1d26fbb1a72.docx"},{"id":97152077,"identity":"d4e99f55-a061-42eb-8a3b-5bc23d1e0e81","added_by":"auto","created_at":"2025-12-01 10:39:51","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6153,"visible":true,"origin":"","legend":"","description":"","filename":"8086b7e8f7c845258b5caa77ec04619e.json","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/24b0c86cc31388795574d0eb.json"},{"id":97248447,"identity":"2520c3f1-dc10-4197-8f6c-e4ceed7dd8e6","added_by":"auto","created_at":"2025-12-02 12:58:46","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":123635,"visible":true,"origin":"","legend":"","description":"","filename":"8086b7e8f7c845258b5caa77ec04619e1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/d420edb919f424badd4da67d.xml"},{"id":97248631,"identity":"22b4d284-92d8-4641-930e-122a4ccfb44a","added_by":"auto","created_at":"2025-12-02 13:04:30","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2154756,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/8db881e034712141756c2195.png"},{"id":97248707,"identity":"01e7392d-dea0-4e3e-ad30-b9c250ce0a35","added_by":"auto","created_at":"2025-12-02 13:06:10","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":125244,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/911facee46c2b8ec686ed5c7.jpeg"},{"id":97152086,"identity":"8e63c002-9a38-4596-a151-b9dcc9b93598","added_by":"auto","created_at":"2025-12-01 10:39:51","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":525801,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage11.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/f1434fae8a62af7026b8ceba.jpeg"},{"id":97248497,"identity":"f489893a-a245-48be-96c1-e030cf02a76e","added_by":"auto","created_at":"2025-12-02 13:01:01","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":168461,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage12.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/a55581b4f8aa18759c1946da.jpeg"},{"id":97249472,"identity":"5a903fdc-622b-4e49-a63f-6de4631d2910","added_by":"auto","created_at":"2025-12-02 13:12:42","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":91719,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage13.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/0e96a46ad8fe067e8328ae5f.jpeg"},{"id":97152094,"identity":"89ba57f5-f0bb-42c9-81a0-a8c73ed776d5","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":361815,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/5b27ef5e6c105e8ef4ec2092.png"},{"id":97249327,"identity":"0056ae14-50b2-4f6f-9560-05345ef33a3f","added_by":"auto","created_at":"2025-12-02 13:12:11","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":156409,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/5198d6358036ee4d1859c591.png"},{"id":97249228,"identity":"a1a6b595-d54b-4a21-a851-5a76babc3cf7","added_by":"auto","created_at":"2025-12-02 13:11:31","extension":"jpeg","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":277595,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage16.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/a4c21961699622f935d8d85a.jpeg"},{"id":97152105,"identity":"3f717002-9c7f-4564-aba4-48571538978d","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":472959,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/5e02bb60f88987e65f801b42.png"},{"id":97152101,"identity":"acdc63d3-9865-482a-bbdb-2c28090cc2b7","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":667004,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/bda081cbd69fe6f93bd4950b.png"},{"id":97249183,"identity":"37f85f90-13c0-4c99-b31c-0947335d99bd","added_by":"auto","created_at":"2025-12-02 13:11:05","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":199549,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/7515c4402d29ca093e3e8d79.png"},{"id":97152120,"identity":"80cc689e-b4a8-41d8-ab06-5b5dbebc0b30","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"jpeg","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":218933,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/db23411adabf0e0ae8d8bfb6.jpeg"},{"id":97152107,"identity":"64d63152-ac9e-41a5-98ef-6fc8a33a1ac3","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"jpeg","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":93501,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/ad0180fa4ee74df65eb4e5de.jpeg"},{"id":97249574,"identity":"1b119227-6372-4abc-bd48-f7541d8c0b8c","added_by":"auto","created_at":"2025-12-02 13:12:56","extension":"jpeg","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":110423,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/a60e2a8654e1e0dd40ce18c5.jpeg"},{"id":97152115,"identity":"5fabb629-dd84-436d-9994-a0445439dce1","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"jpeg","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":502385,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage8.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/cf934ab452ede34983f1d3c0.jpeg"},{"id":97249830,"identity":"d6bc6401-c3c5-491a-af17-edd61d28cf44","added_by":"auto","created_at":"2025-12-02 13:13:30","extension":"jpeg","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/73b49b5df415008ad732c2f6.jpeg"},{"id":97249543,"identity":"ffc100fe-f996-4bc1-96f1-f9dcb5409639","added_by":"auto","created_at":"2025-12-02 13:12:53","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":219191,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/aabfb52b7389f2e2901d5a11.png"},{"id":97248530,"identity":"304c0504-65b6-4e24-9180-7da2ea10e73d","added_by":"auto","created_at":"2025-12-02 13:02:43","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":30135,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/8cd06921400ef47ef6a4e16d.png"},{"id":97248455,"identity":"29006bb1-8d19-46d2-9894-087ee2b87f79","added_by":"auto","created_at":"2025-12-02 12:58:52","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":93232,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/d97c52255186d8ca883b17f9.png"},{"id":97249826,"identity":"8b5c96f7-d143-47d8-8ca8-40ce2a27bf4a","added_by":"auto","created_at":"2025-12-02 13:13:30","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37036,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/1883ec8c74ad3f3ec162ea7c.png"},{"id":97152108,"identity":"c1155ef1-9930-446c-a41e-d946db100cf3","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24252,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/6e7cfb2d30b99a696e2d0af2.png"},{"id":97152128,"identity":"9f812dcc-91da-4692-8cf4-e940a5b37fee","added_by":"auto","created_at":"2025-12-01 10:39:53","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":57120,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/4cfd7aa9bbe46f02f6a03ff7.png"},{"id":97152110,"identity":"80538f77-eab5-4274-b0b2-a20ff8d39ad3","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":32140,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/2505c14e344b642384a3596f.png"},{"id":97152111,"identity":"10bbcb93-fa40-4156-aab2-802aee2a8ba7","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":72106,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/f10f4533f7e2db5f7923a9af.png"},{"id":97152126,"identity":"4957dfb3-fd3b-4fe3-80af-f17d33af098d","added_by":"auto","created_at":"2025-12-01 10:39:53","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40714,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/b655b6fe916a47eda1c9672f.png"},{"id":97152122,"identity":"b1a7918d-c705-4de4-85a2-3081c1b00019","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111779,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/aff3ad1b2d24c05e8851e50d.png"},{"id":97152125,"identity":"4eba0915-c1f8-4829-a85e-dfdf8530b3d5","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44446,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/84c1e676424ecec1fe65937d.png"},{"id":97152112,"identity":"33f91322-710a-47be-84f8-99872182b588","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":73789,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/50dd9a79eeb9052ec493c4fb.png"},{"id":97152127,"identity":"7f83001b-84ba-4261-9ce8-f21216f7eac1","added_by":"auto","created_at":"2025-12-01 10:39:53","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":19520,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/aae99cec60d52e450adc009f.png"},{"id":97248782,"identity":"08fb1fd4-6291-4216-933c-4b8bc3df8dbd","added_by":"auto","created_at":"2025-12-02 13:07:10","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":29077,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/7aab6080fa506fc3ecb8efc9.png"},{"id":97249806,"identity":"23789807-ae9e-40cd-8fd0-f07221dedcf0","added_by":"auto","created_at":"2025-12-02 13:13:28","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":85163,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/f3f25a22e384cf9c2a390220.png"},{"id":97152114,"identity":"dfb0adcc-ed0b-46c7-a88b-d75bb1f719b3","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/ae4ecc1cb802903014da8238.png"},{"id":97249599,"identity":"736904fc-787e-4ff2-927b-f556fe8c85be","added_by":"auto","created_at":"2025-12-02 13:13:00","extension":"xml","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":120969,"visible":true,"origin":"","legend":"","description":"","filename":"8086b7e8f7c845258b5caa77ec04619e1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/5a6f602c58c3afb745eb59ef.xml"},{"id":97152117,"identity":"eb5c216d-19b0-4380-b6b3-a508d462f64b","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"html","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":134282,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/db5fc993f7cdd6b03f99830e.html"},{"id":97152078,"identity":"bd5779c5-7362-4ad3-a1f1-a17ee0086e60","added_by":"auto","created_at":"2025-12-01 10:39:51","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2154756,"visible":true,"origin":"","legend":"\u003cp\u003eIllustration of key functions of the kidney [12]\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/b2bcb2e50aba18a6903ee63d.png"},{"id":97152079,"identity":"d8a62d8b-322f-46b1-b5af-f06a14ad0c6a","added_by":"auto","created_at":"2025-12-01 10:39:51","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":472959,"visible":true,"origin":"","legend":"\u003cp\u003eLevels of Healthcare System in Rwanda [14]\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/1ff0fcc8f2d31943e53825ce.png"},{"id":97248383,"identity":"81dd20a7-b1b8-41d4-998e-03c77c295820","added_by":"auto","created_at":"2025-12-02 12:56:05","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":667004,"visible":true,"origin":"","legend":"\u003cp\u003eDiagnosis of Diabetes Kidney Disease [15].\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/1bdd07dc4d32c18cb12c107f.png"},{"id":97248382,"identity":"71ac1147-adf2-4260-8947-f62bba7a224e","added_by":"auto","created_at":"2025-12-02 12:56:05","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":199549,"visible":true,"origin":"","legend":"\u003cp\u003eStudy Design\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/9283e4f900939c6088a18da7.png"},{"id":97249224,"identity":"b5845c87-13fe-41ed-afcb-369028c739b0","added_by":"auto","created_at":"2025-12-02 13:11:29","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":218933,"visible":true,"origin":"","legend":"\u003cp\u003eFeature correlation with DKD for the MIMIC-IV dataset.\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/789b6147752159a05a0cf5b5.jpeg"},{"id":97152090,"identity":"4bde1bc5-11df-476b-b0e0-e052452fa271","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":93501,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of categorical features in MIMIC-IV DKD dataset\u003c/p\u003e","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/b06e20550c6b15281ebb8774.jpeg"},{"id":97152087,"identity":"43f44237-c6a4-4ba5-8283-6a3366706dc8","added_by":"auto","created_at":"2025-12-01 10:39:51","extension":"jpeg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":110423,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of categorical features for the Dataset from Rwanda\u003c/p\u003e","description":"","filename":"floatimage7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/0f8a5df682856f51dc8a2fdf.jpeg"},{"id":97152088,"identity":"1e1a6a09-20a8-472e-bc41-4b3bcaeb9613","added_by":"auto","created_at":"2025-12-01 10:39:51","extension":"jpeg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":75761,"visible":true,"origin":"","legend":"\u003cp\u003eDiabetes Kidney Disease (DKD) distribution by Age group. (a) MIMI-IV dataset. (b) Dataset from Rwanda\u003c/p\u003e","description":"","filename":"floatimage8.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/5ada75697201a038f2f07931.jpeg"},{"id":97152092,"identity":"62d308f0-e2ec-4b1b-8608-24d10120de13","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"jpeg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":125244,"visible":true,"origin":"","legend":"\u003cp\u003eVisualizing numerical data using a box plot\u003c/p\u003e","description":"","filename":"floatimage9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/9b5d2aa386791eb3bb52c348.jpeg"},{"id":97248518,"identity":"8164b150-e842-4a87-aeb5-833791c494dd","added_by":"auto","created_at":"2025-12-02 13:02:18","extension":"jpeg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":84562,"visible":true,"origin":"","legend":"\u003cp\u003eReceiver operating characteristic curves for all machine learning models on DKD prediction: (a) MIMIC-IV dataset, (b) DKD dataset from Rwanda. AUC: Area Under the Curve.\u003c/p\u003e","description":"","filename":"floatimage10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/ef698b2ae09647f7fdfbf84d.jpeg"},{"id":97249307,"identity":"3f469ef4-8e0b-458b-8916-07fd33f7a50f","added_by":"auto","created_at":"2025-12-02 13:12:05","extension":"jpeg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":168461,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion Matrix of Random Forest as the best model on both datasets: (a) MIMIC-IV dataset, (b) DKD dataset from Rwanda.\u003c/p\u003e","description":"","filename":"floatimage11.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/ad1a2961cdc0b0762b903d4c.jpeg"},{"id":97248528,"identity":"2ac76cb4-3071-4bb9-b878-47a53fccccc6","added_by":"auto","created_at":"2025-12-02 13:02:41","extension":"jpeg","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":91719,"visible":true,"origin":"","legend":"\u003cp\u003eCalibration curve of five tree-based ensemble models on the Rwanda DKD dataset.\u003c/p\u003e","description":"","filename":"floatimage12.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/df31c0c91fbe724723cf01a4.jpeg"},{"id":97152097,"identity":"5f094997-51aa-4b64-ae08-3ff36aeb9780","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":361815,"visible":true,"origin":"","legend":"\u003cp\u003eGlobal model explanation using SHAP summary plot (MIMIC-IV Dataset).\u003c/p\u003e","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/63b0519e149b381aa273f8d4.png"},{"id":97152113,"identity":"c8aee741-c734-427e-bcef-cad3870fefff","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":156409,"visible":true,"origin":"","legend":"\u003cp\u003eTop 15 features contributing to DKD risk using the Rwanda dataset.\u003c/p\u003e","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/c2d73a2542a2c1644551d1c2.png"},{"id":97152099,"identity":"8aef0774-c73a-4139-a012-8241230c0f50","added_by":"auto","created_at":"2025-12-01 10:39:52","extension":"jpeg","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":277595,"visible":true,"origin":"","legend":"\u003cp\u003eModel interpretation at patient level: (a) SHAP explanations DKD positive case using the dataset from Rwanda, (b) SHAP explanations of a negative DKD case using the dataset from Rwanda.\u003c/p\u003e","description":"","filename":"floatimage15.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/f1f77e5758c0daac3b62e8ff.jpeg"},{"id":97664508,"identity":"44bc398c-1065-4713-bde1-ef79a64998e0","added_by":"auto","created_at":"2025-12-08 09:07:19","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5713346,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8108979/v1/a5cb8ce9-d162-4e8d-a1c2-03da485b6408.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Using Explainable Machine Learning for Early Detection of Diabetic Kidney Disease in Rwandan Diabetic Patients","fulltext":[{"header":"1 INTRODUCTION","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eDiabetes Mellitus (DM) a chronic non-communicable disease resulting from the metabolic disorder characterized by high blood sugar levels, remains the all-time global health concern. According to the International Diabetes Federation (IDF), an estimated 537\u0026nbsp;million people were living with diabetes in 2021, and the highest proportions of undiagnosed diabetes (53.6%) were found in Africa, and diabetes cases are expected to rise to 783\u0026nbsp;million by 2045 [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Uncontrolled diabetes has no limit in causing burdens to diabetic patients. The progression of diabetes brings new complications, including Diabetes Kidney Disease (DKD), also known as diabetes nephropathy. DKD is a progressive condition that impairs the kidneys' ability to filter blood, primarily due to prolonged high blood glucose and hypertension, both of which are prevalent in diabetic patients, as reported in [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. It is a leading cause of Chronic Kidney Diseases (CKD), including renal failure and the End Stage Renal Disease (ESRD) [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. In its early stages, DKD is typically asymptomatic, making early detection difficult. As the disease progresses, symptoms such as frequent urination, foamy urine, itching, and general discomfort may appear. DKD also imposes a significant financial burden, especially in advanced stages where treatment may involve dialysis or kidney transplantation [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. These interventions are not only costly but also place a considerable strain on healthcare systems and patients, particularly in low-resource settings. Machine learning (ML) has demonstrated strong potential in uncovering complex, non-linear patterns within data [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Its application in disease detection is particularly promising, offering valuable support for the early diagnosis of multifactorial conditions such as diabetes and its associated complications.\u003c/p\u003e\u003cp\u003eThe growing availability of patient health data provides an opportunity to leverage ML for the early identification of DKD, which can significantly contribute to timely prevention and intervention strategies in diabetes management [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, the adoption of AI-based tools in diabetes management has been relatively slow, largely due to the black-box nature of many ML models, which rely on complex algorithms that lack transparency [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Although these models often deliver high predictive performance, their limited interpretability presents challenges for clinical integration and decision-making [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo address this, our study employs SHAP (SHapley Additive exPlanations) [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], a post-hoc model interpretation method, to enhance the transparency of the best-performing ML model. By providing clear insights into the risk factors influencing predictions, SHAP helps bridge the gap between predictive performance and clinical interpretability, enabling clinical decisions supported by evidence from the data and thereby improving trust and facilitating medical adoption. Patients with either type 1 or type 2 diabetes who have had the condition for five years or longer are considered at elevated risk for developing Diabetic Kidney Disease (DKD) and are typically recommended to undergo regular screening. However, because the precise onset of diabetes is not known and early stages of DKD are asymptomatic, there is a critical need for explainable machine learning approaches to support early detection, particularly among high-risk diabetic populations.\u003c/p\u003e\u003cp\u003eMost diabetes-related complications, including DKD, are preventable through early detection, timely intervention, and effective management of blood glucose and blood pressure. This study proposes an explainable machine learning-based predictive tool for the early identification of diabetic patients at risk of developing DKD. The models were trained on the publicly available Medical Information Mart for Intensive Care (MIMIC)-IV dataset [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] and clinical datasets from three hospitals in Rwanda. Among five trained models, the best-performing one was further interpreted using the SHAP algorithm, which provides transparency into how individual predictions are made. This level of explainability enables clinicians to evaluate the model's outputs in the context of clinical judgment, thereby enhancing trust and potential adoption in real-world healthcare settings. In the context of Rwandan healthcare, explainable ML offers a dual advantage. First, it enhances the accuracy of early DKD detection by identifying individuals at risk before clinical symptoms manifest. Second, it supports data-driven decision-making by highlighting the key risk factors, thus allowing for targeted preventive strategies that preserve healthcare expenditure and maintain optimal care for diabetes patients.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e\u003ch2\u003e1.1 Kidney Functions\u003c/h2\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the four key stages of healthy kidney function [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. In stage 1, blood containing various waste products enters the kidney, where the glomerular filtration system selectively filters out waste while retaining essential proteins and cells. Stage 2 involves the reabsorption of vital substances, including glucose, amino acids, sodium, potassium, and water, which helps maintain fluid and electrolyte balance in the body. In stage 3, the kidneys secrete additional waste and regulate acid-base balance, directing the final waste products to the bladder for excretion. In the final stage, the filtered, clean blood is returned to the bloodstream via the renal veins.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eElevated blood glucose and hypertension damage the small blood vessels in the glomerular filtration membrane, compromising its selective permeability and allowing large molecules, such as proteins, to pass through. This progressively reduces filtration efficiency, leading to waste accumulation in the body. As a result, diabetic patients may experience symptoms such as itchiness, Frequent urination, fatigue, dry mouth, loss of appetite, and foamy urine [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e1.2 Rwanda Healthcare System\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eRwanda has undergone a significant transformation in its healthcare sector, marked by substantial improvements in infrastructure, including the adoption of information and communication technology (ICT) to support digital health and enhance the delivery of health services. Key reforms have included the expansion of the healthcare workforce, the decentralization of service provision, and the implementation of universal health coverage through the Community-Based Health Insurance (CBHI) program, which enables citizens to access healthcare services at a small cost. The national healthcare system is structured across five hierarchical levels, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. At the top are the National Referral Hospitals (NRH) and University Teaching Hospitals (UTH), while the bottom level comprises basic healthcare services delivered by Community Health Workers (CHWs).\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e1.3 DKD Diagnosis\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eDiabetic Kidney Disease (DKD) is a form of Chronic Kidney Disease (CKD) that results from diabetes. Its diagnosis follows similar procedures to those used for CKD, beginning with the evaluation of clinical symptoms and laboratory assessments, including urine analysis and measurement of Glomerular Filtration Rate (GFR). The presence of macromolecules such as albumin and proteins in the urine serves as a primary indicator of DKD. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the measurements of Albuminuria and GFR, along with their diagnostic interpretation. Three stages of Albuminuria (A1-A3) and DKD stages based on GFR (G1-G5).\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eClinically, Diabetic Kidney Disease (DKD) is characterized by a decline in the kidneys\u0026rsquo; ability to filter waste products from the bloodstream, which is reflected in a reduced estimated Glomerular Filtration Rate (eGFR) and persistently elevated levels of albuminuria (\u0026gt;\u0026thinsp;300 mg/g of creatinine).\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e1.4 DKD Treatment and Prevention Strategies\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe management of Diabetic Kidney Disease (DKD) typically involves lifestyle modifications, as well as the regulation of blood glucose and blood pressure levels. Effective prevention relies on the early detection of DKD in high-risk diabetic individuals. This study aims to predict the risk of DKD among diabetic patients using machine learning by leveraging demographic information, vital signs, and laboratory results extracted from Electronic Health Records (EHRs).\u003c/p\u003e\u003cp\u003eThis study is guided by two primary objectives: (1) to train and evaluate machine learning models using a publicly available dataset (MIMIC-IV) and diabetic kidney disease (DKD) dataset obtained from three Rwandan hospitals, and (2) to interpret the most effective model for early DKD detection using SHAP (SHapley Additive exPlanations) techniques. The rest of the paper is structured as follows: Section 2 details the methodology, including data sources, preprocessing, model development, and evaluation techniques. Section 3 presents the experimental results. Section 4 discusses the implications of the findings in the context of DKD prediction and healthcare in Rwanda. Finally, Section 5 concludes the study and outlines directions for future research.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"2 METHODOLOGY","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThis study was conducted in five phases: the study design began with data extraction and dataset preprocessing, followed by the training of five tree-based ensemble models, evaluation of model performance, and interpretation of model predictions using SHAP methods on both the public dataset and the dataset from Rwanda. The primary purpose of training and testing machine learning models on datasets from two sources is to analyze how the models generalize to data recorded in different contexts and distributions. The overall study design, including all research phases, is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Study Design\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThis study utilized two datasets: one extracted from the Medical Information Mart for Intensive Care (MIMIC-IV) database provided by the Beth Israel Deaconess Medical Center, and another comprising aggregated data from three hospitals in Rwanda. Data preprocessing techniques were applied to prepare both datasets for machine learning applications. Following preprocessing, each dataset was partitioned into training and testing subsets. Five tree-based ensemble machine learning models were developed, comprising one bagged ensemble model and four boosted ensemble models. These models were trained on the training data and evaluated on the test data using five performance metrics. The best-performing model was subsequently interpreted using SHAP methods.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Data Extraction\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eData were extracted from the Electronic Health Record (EHR) systems. Patient data were identified using the 10th Revision of the International Classification of Diseases (ICD-10) codes for type 1 and type 2 diabetes mellitus. Specifically, the format E10.XXX was used for Type 1 Diabetes Mellitus (T1DM), and E11.XXX for Type 2 Diabetes Mellitus (T2DM), where X denotes a digit from 0 to 9 indicating the stage or complication of the disease [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Diabetic Kidney Disease (DKD) cases were identified using codes E10.2X for T1DM patients with kidney complications and E11.2X for T2DM patients with kidney disease. For the Rwandan dataset, the same data extraction method was applied. In instances where the relevant ICD-10 code was not available, clinical notes accompanying the diagnosis were reviewed, and in most cases, they provided sufficient information to confirm DKD status. The MIMIC-IV dataset contains data of diabetic patients recorded between 2008 and 2019.\u003c/p\u003e\u003cp\u003eDiabetic Kidney Disease (DKD) is a progressive complication that typically affects both type 1 and type 2 diabetic patients after at least five years following a confirmed diabetes diagnosis. However, due to the undiagnosed diabetes, the precise onset dates are not known; some patients may develop DKD earlier within one year of diagnosis or even at the time of initial diabetes screening. To address this uncertainty and improve the reliability of the analysis, only patients with a minimum of one year of follow-up were included in the study. The final datasets consisted of 949 samples from the MIMIC-IV database and 810 samples from the Rwandan dataset. The Rwandan data were collected from three hospitals: 383 samples from the National Referral Hospital (CHUK), 204 from a Teaching University Hospital (Ruhengeri Hospital), and 223 from a District Hospital (Nyamata Hospital). Both datasets included 20 predictive features for DKD, comprising 16 numerical and 4 categorical variables, along with a single class label indicating the presence or absence of DKD. The dataset from Rwanda contains data of diabetes patients recorded between 2012 and 2024\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Data Preprocessing\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eData preprocessing is a crucial step in any machine learning pipeline, as the performance of predictive models, such as those used to detect DKD risk, largely depends on the quality and relevance of the data used during training and evaluation. Figure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e illustrates the correlation between various features and the presence of DKD in the MIMIC-IV dataset. In Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, lighter orange shades represent features with weaker correlations to DKD risk, while darker orange shades indicate stronger correlations. Among the features, Complication, Urea (Blood Urea Nitrogen), and Creatinine exhibit the strongest associations with the risk of developing DKD in this dataset.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e illustrate the class imbalance present in both datasets, where the minority class corresponds to patients at risk of developing Diabetic Kidney Disease (DKD). This imbalance reflects a common clinical reality, as DKD typically affects approximately one-third of the diabetic population. To address this issue, the Synthetic Minority Oversampling Technique (SMOTE) [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] was applied. SMOTE generates new synthetic instances of the minority class through interpolation between existing samples, thereby increasing the representation of the minority class and enhancing the model's ability to learn from imbalanced data.\u003c/p\u003e\u003cp\u003eThe dataset contained missing values across all numerical features except for the Age feature. The highest proportion of missing data was observed in Total Cholesterol (TC) at 38.04%, followed by Glycated Hemoglobin (HbA1c) at 28.33%. To address this, missing values were imputed using the mean of each respective feature. Additionally, samples containing extreme outlier data points that significantly deviated from the overall distribution were considered irrelevant and subsequently removed. For instance, records with a Body Mass Index (BMI) of 3658.5 and a Blood Glucose (BG) level of 1120 mg/dL were excluded from the analysis.\u003c/p\u003e\u003cp\u003eAll numerical features were normalized using the Min-Max scaling technique to ensure that all variables were on the same scale, thereby preventing the models from under learning features with smaller value ranges. To prevent data leakage and ensure a fair evaluation of model performance, the Synthetic Minority Oversampling Technique (SMOTE) was applied exclusively to the training dataset, while the test dataset remained in its original form.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe MIMIC-IV DKD dataset comprises 274 patients diagnosed with Diabetic Kidney Disease (DKD) and 675 patients without DKD. Among the total cohort, 577 patients present with at least one diabetes-related complication other than nephropathy, while 372 have no recorded complications. In terms of diabetes type, 603 patients have Type 2 Diabetes Mellitus (T2DM) and 346 have Type 1 Diabetes Mellitus (T1DM). Additionally, 306 patients are diagnosed with chronic hypertension, whereas 643 have no history of hypertension. The dataset also includes 482 male and 467 female patients.\u003c/p\u003e\u003cp\u003eThe dataset from Rwanda exhibits an imbalanced distribution of class labels, comprising 256 patients diagnosed with Diabetic Kidney Disease (DKD) and 554 patients without DKD. The dataset counts 164 patients with diabetes-related complications and 646 without complications. Additionally, 312 patients present with hypertension (HT), whereas 498 are without this condition. In terms of diabetes type, there are 558 patients with Type 2 Diabetes Mellitus (T2DM) and 252 with Type 1 Diabetes Mellitus (T1DM). The gender distribution includes 533 female and 277 male patients. Figure\u0026nbsp;8 illustrates the distribution of DKD cases by age group. Figure\u0026nbsp;8 (a) shows that DKD steadily increases with age. Figure\u0026nbsp;8 (b) reveals a peak incidence between the ages of 30 and 69. The number of DKD cases slightly declines after the age of 70, which corresponds with Rwanda\u0026rsquo;s life expectancy of 69.1 years as reported by the National Institute of Statistics of Rwanda (NISR) in 2022 [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTable 1 presents the descriptive analysis of the numerical variables in the Rwandan DKD dataset. The analysis reveals that several variables have a significant proportion of missing values, exceeding 10%. Specifically, Body Mass Index (BMI) has 16.54% missing data, glycated hemoglobin (HbA1c) 12.96%, urea 14.2%, systolic blood pressure (SBP) 11.85%, diastolic blood pressure (DBP) 11.98%, and total cholesterol (TC) 19.88%. The missing values were handled using mean imputation methods. Each missing value was replaced by the column mean value.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eDescriptive analysis of the numerical variables in the Rwandan DKD dataset\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"12\"\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\u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eFeature Name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCounts\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMean\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eStd\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003eMin\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003eMax\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003eMissing Values (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c12\"\u003e\u003cp\u003eNormal Range\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eAge\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e810\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e56.51\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e16.01\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003eN/A\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eBMI\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e676\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e27.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5.07\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e15.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e48.49\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e16.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e18.5\u0026ndash;24.9\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eHbA1C\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e705\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e10.15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2.65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e5.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e19.10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e12.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;5.7\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eBG\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e805\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e229.45\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e106.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e53.69\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e833.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e0.62\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e70\u0026ndash;100\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eCreatinine\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e766\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e7.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e5.43\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e0.6\u0026ndash;1.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eUrea\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e695\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e10.14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4.50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e3.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e37.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e14.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e10\u0026ndash;20\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eSodium\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e789\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e138.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4.64\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e121.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e155.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e2.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e135\u0026ndash;145\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003ePotassium\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e749\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e4.41\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.78\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e2.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e8.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e2.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e3.6\u0026ndash;5.2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eHemoglobin\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e749\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e13.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2.52\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e6.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e21.40\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e7.57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e13.2\u0026ndash;16.6\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eRBC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e758\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e4.74\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.26\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e1.51\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e10.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e6.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e4.2\u0026ndash;6.1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eWBC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e757\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e7.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3.87\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e2.36\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e30.66\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e6.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e4.5\u0026ndash;11.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eNeutrophils\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e758\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e54.83\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e15.35\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e16.80\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e92.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e6.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e40\u0026ndash;60\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eLymphocytes\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e753\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e32.79\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e13.94\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e10.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e69.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e7.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e20\u0026ndash;40\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eSBP\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e714\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e137.56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e25.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e92.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e210.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e11.85\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;120\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eDBP\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e713\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e80.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e14.70\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e38.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e145.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e11.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;80\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eTC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e649\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e5.13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1.58\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e\u003cp\u003e0.98\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e\u003cp\u003e11.30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c11\" namest=\"c10\"\u003e\u003cp\u003e19.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u003cp\u003e3.9\u0026ndash;5.5\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=\"BlockQuote\"\u003e\u003cp\u003eThe descriptive analysis reveals that some diabetic patients exhibit vital signs and laboratory results that deviate significantly from clinically accepted normal ranges. Since age does not have a standardized normal range, it was marked as not applicable (N/A) in this context. To further explore these deviations, a box plot of the numerical features is presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e9\u003c/span\u003e, providing a visual representation for identifying potential outliers within the dataset.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e9\u003c/span\u003e displays box plots that highlight data points lying above or below the whiskers, suggesting the presence of potential outliers. Upon further analysis, it was observed that patients diagnosed with DKD often exhibited abnormal clinical values. While these abnormal values were retained due to their clinical relevance, extreme outliers, which fall far outside the overall data distribution, were excluded. For instance, blood glucose levels exceeding 600 mg/dL were considered extreme outliers and were subsequently removed from the dataset. The final Rwandan dataset comprises 801 samples. Prior to model training, the dataset was normalized using the Min-Max scaling technique. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training subset, while the test subset remained unaltered to ensure an unbiased evaluation.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Machine Learning Models\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eDiabetes Kidney Disease (DKD) risk prediction is a classification task that uses supervised machine learning technique. In this study, ensemble machine learning models were employed to predict the risk of DKD among diabetes patients, given their demonstrated effectiveness in handling complex datasets, particularly those comprising both numerical and categorical data. These models are widely recognized for their strong performance in medical applications. Specifically, we utilized one bagging-based model (Random Forest) and four boosting-based models: Extreme Gradient Boosting (XGBoost), CatBoost, AdaBoost, and LightGBM.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eRandom Forest.\u003c/b\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eRandom Forest is a bagging-based ensemble model that leverages bootstrapping and aggregation techniques for learning. It constructs multiple decision trees that are trained in parallel, each trained on randomly selected subsets of the original dataset, using either instance-based or feature-based sampling. This allows individual samples or features to appear multiple times within a training subset. The model aggregates the predictions from all individual trees to produce a final output by averaging in regression tasks and majority voting in classification tasks. Random Forest has been widely adopted in medical research, particularly for disease detection. For example, it was utilized in [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e] for predictive modeling of early-stage diabetes, achieving a classification accuracy of 97.03%.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eExtreme Gradient Boosting (XGBoost)\u003c/b\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eExtreme Gradient Boosting (XGBoost) is a sequential ensemble learning method that builds decision trees iteratively, with each subsequent tree aiming to correct the prediction errors of its predecessors through gradient-based optimization. As an open-source library, XGBoost implements a distributed gradient-boosted decision tree algorithm designed for both classification and regression tasks. It integrates regularization techniques with advanced optimization strategies to enhance predictive accuracy while reducing training time. The core learner in XGBoost is the Classification and Regression Tree (CART). Due to its high accuracy and computational efficiency, XGBoost is widely applied in solving problems involving structured data. For instance, in [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], XGBoost was used to predict chronic kidney disease, achieving a classification accuracy of 93.29% and ranking it as the best model.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eCatBoost Model\u003c/b\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eCatBoost is an open-source gradient boosting library specifically designed to handle categorical features effectively without extensive preprocessing or transformation. Developed by Yandex, it is particularly well-suited for machine learning tasks involving heterogeneous data types. CatBoost constructs its model iteratively using decision trees as weak learners, optimizing performance through gradient boosting. A key advantage of CatBoost lies in its use of ordered boosting and random permutations, which help prevent overfitting and improve generalization. Additionally, it incorporates gradient-based optimization to enhance predictive accuracy on complex datasets. In [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], CatBoost was applied to predict diabetic kidney disease in patients with type 2 diabetes, achieving a classification accuracy of 75.5%.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eAdaBoost Model\u003c/b\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAdaptive Boosting (AdaBoost) is a boosting ensemble algorithm that combines multiple weak decision tree classifiers to form a single, strong predictive model. It operates by training weak learners sequentially, starting from one decision tree, also known as a decision stump, with each iteration placing increased emphasis on the instances that were previously misclassified. This adaptive weighting mechanism enables the model to focus on difficult cases, thereby improving overall performance. AdaBoost is particularly effective for binary classification and regression tasks, with its strength lying in its iterative learning approach. While it performs well on smaller datasets, the sequential nature of its training process can lead to longer training times when applied to larger datasets. In [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], AdaBoost was utilized for predicting diabetic kidney disease staging, achieving a classification accuracy of 83.5%.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eLightGBM Algorithm\u003c/b\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eLight Gradient Boosting Machine (LightGBM) is an open-source gradient boosting framework developed by Microsoft for classification and regression tasks. It constructs decision trees using a leaf-wise strategy, which enhances training speed and accuracy. LightGBM\u0026rsquo;s efficiency is largely attributed to two key techniques: Gradient-Based One-Side Sampling (GOSS), which improves computational efficiency by prioritizing instances with high gradients, and Exclusive Feature Bundling (EFB), which reduces feature dimensionality by combining mutually exclusive features. These innovations enable LightGBM to train effectively on large datasets with reduced resource demands. In [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], LightGBM achieved an area under the curve (AUC) score of 0.815 in predicting diabetic kidney disease, outperforming other models used in the study.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e2.5 Models Performance Evaluation\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe selected ensemble models were trained on 70% of each dataset, with the remaining 30% used for testing. Model performance was evaluated using five metrics: accuracy, sensitivity, specificity, F1-score, and area under the curve (AUC). For each dataset, the confusion matrix of the best-performing model was used to visualize its ability to identify the risk of diabetic kidney disease (DKD). Additionally, ROC-AUC curves were plotted to illustrate the trade-off between the true positive rate (TPR) and false positive rate (FPR), and a calibration curve was used to visualize the reliability of models\u0026rsquo; confidence in predicting DKD risk. The equations (1) to (4) explain how accuracy, sensitivity, specificity, and F1-score are calculated.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:Accuracy=\\:\\:\\frac{TP\\:+\\:TN}{TP\\:+\\:TN\\:+FP\\:+\\:FN}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:Sensitivity=\\:\\:\\frac{TP}{TP\\:+\\:FN}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(2\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:Specificity=\\frac{TN}{TN\\:+\\:FP}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(3\\right)$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:F1-score\\:\\:=\\frac{2*TP}{2*TP\\:+\\:FP\\:+\\:FN}\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\left(4\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eHere, TP (true positives) denotes correctly identified positive cases, TN (true negatives) indicates correctly identified negative cases, FP (false positives) refers to negative cases incorrectly classified as positive, and FN (false negatives) represents positive cases incorrectly classified as negative. Sensitivity reflects the model\u0026rsquo;s ability to detect positive cases, while specificity measures its ability to correctly identify negatives.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e2.6 Explainable Artificial Intelligence (XAI) Methods\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eEnsemble models offer an improved predictive performance by combining multiple learners, but often result in increased complexity, making them difficult to interpret. In critical domains such as medicine, the lack of transparency in black-box models can hinder their adoption, as a model that does not offer details of its predictions may lead to incorrect clinical decisions and compromise patient safety. Explainable Artificial Intelligence (XAI) addresses this challenge by developing methods to interpret model behavior. In this study, a model-agnostic approach based on feature attribution is employed. Specifically, Shapley Additive exPlanations (SHAP), which use the principles of game theory to compute the contribution of each feature to the prediction. While SHAP can be computationally intensive, optimized variants have been developed to improve efficiency. We utilize FastTreeSHAP [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e], an advanced version of TreeSHAP, to compute SHAP values for tree-based models. The original SHAP [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e] is presented in Eq.\u0026nbsp;(\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:{\\:\\phi\\:}_{i\\left(f\\right)}=\\sum\\:_{S\\subseteq\\:\\cup\\:N\\backslash\\:\\left\\{i\\right\\}}\\frac{\\left|s\\right|!\\left(N-\\left|s\\right|-1\\right)!}{N!}\\left[{f}_{x}\\left(s\\cup\\:\\left\\{i\\right\\}\\right)-{f}_{x}\\left(s\\right)\\right]$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:S\\)\u003c/span\u003e\u003c/span\u003e is a subset of features excluding feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:N\\)\u003c/span\u003e\u003c/span\u003e is the total number of features. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\varphi\\:}_{i}\\left(f\\right)\\)\u003c/span\u003e\u003c/span\u003e represents the contribution of feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e, calculated as the difference between the total prediction and the prediction without feature \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e2.7 Related work\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eRecent research has increasingly focused on leveraging medical data and machine learning to address non-communicable diseases such as diabetes and its complications. In [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], XGBoost was used to predict diabetic nephropathy using 548 patient records from SAHDMU Hospital (2018\u0026ndash;2019). After applying LASSO for feature selection, the model achieved an AUC of 0.966. Calibration curves and SHAP values were employed to enhance the clinical interpretability of the XGBoost model. In [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], a study on diabetic kidney disease (DKD) prediction was conducted using data from 1,177 diabetic patients at Beijing Pinggu Hospital (2013\u0026ndash;2017), including 263 DKD cases and 914 non-DKD cases. The Random Forest classifier achieved an accuracy of 89.83%, suggesting its potential for initial DKD screening. Feature importance analysis identified creatinine and blood urea nitrogen as key predictors.\u003c/p\u003e\u003cp\u003eCatBoost was employed in [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] to predict both chronic kidney disease (CKD) and diabetes. Using a public dataset comprising 202 samples and 29 features, the model achieved 95% accuracy in predicting CKD. For diabetes prediction, it attained 99% accuracy on a separate dataset containing 520 samples and 17 features. SHAP was used for feature effect analysis. The Random Forest model demonstrated superior performance in [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] at predicting chronic kidney disease (CKD), achieving 98.8% accuracy on a dataset comprising 455 samples and 25 features.\u003c/p\u003e\u003cp\u003eMachine learning models were used in [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e] to predict the five-year risk of developing diabetic kidney disease (DKD) following a diagnosis of type 2 diabetes (T2DM). Using a dataset of 87,973 records from USA diabetes patients (2007\u0026ndash;2020), the Random Forest model achieved an AUC of 0.75. In [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], the Random Forest model was trained on a dataset comprising 10,064 samples extracted from Australian electronic health records (EHR) and validated on 597 samples from a Japanese dataset. It achieved a classification accuracy of 92.6% on the Australian data and 73.8% on the Japanese data.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Experimental Setup\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe experiments were conducted using Python and relevant libraries, including Pandas, NumPy, Matplotlib, Seaborn, Scikit-learn, and FastShap for model explainability, as well as other required libraries. The workflow began with exploratory data analysis and feature engineering to prepare the final datasets. In consultation with a clinical nurse experienced in chronic kidney disease (CKD), relevant features were selected for model development. Extensive preprocessing was performed to ensure data quality suitable for training ensemble models to predict the risk of diabetic kidney disease (DKD). Model performance was evaluated using multiple metrics, and the best-performing model was interpreted using SHAP to provide both local and global insights into the decision-making process.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e3.2 DKD Prediction\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe goal of this study is to predict the risk of diabetic kidney disease (DKD) by leveraging electronic health record (EHR) data and machine learning for the early identification of high-risk diabetic patients, thereby enabling timely interventions to prevent or delay severe complications and improve disease management. The predictive performance of all models on the MIMIC-IV dataset is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e2\u003c/span\u003e, where all models demonstrate good performance in predicting DKD risk, with an accuracy above 80% for each model. The best model is Random Forest and the model with the lowest performance is AdaBoost with 82.71% accuracy. Random Forest outperformed other models in three out of five metrics, achieving an accuracy of 87.97% and an AUC of 0.948.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eModels evaluation on the test subset of the MIMIC-IV dataset\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\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eClassifier\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eAUC-score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eRandomForest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e87.97\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e75.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e\u003cb\u003e93.12\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e\u003cb\u003e78.38\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e\u003cb\u003e0.948\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e85.34\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e77.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e88.36\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e75.47\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.929\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eCatBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e86.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e77.92\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e89.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e76.43\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.945\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eLightGBM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e87.22\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e76.62\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e91.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e77.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.931\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eAdaBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e82.71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e49.35\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e96.30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e62.30\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.924\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=\"BlockQuote\"\u003e\u003cp\u003eThe same five ensemble models were applied to the Rwandan dataset for predicting the risk of diabetic kidney disease (DKD). All models showed slight improvements in performance, with Random Forest achieving the highest accuracy of 91.70%, and AdaBoost the lowest at 84.65%. Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents a summary of the performance of each model across all evaluation metrics.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eModels evaluation on test subset of Rwanda DKD dataset\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\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eClassifier\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eAUC-score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eRandomForest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e91.70\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e88.16\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e\u003cb\u003e93.33\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e\u003cb\u003e87.01\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e\u003cb\u003e0.960\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e77.18\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e77.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e76.97\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e68.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.861\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eCatBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e89.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e84.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e91.52\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e83.12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.951\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eLightGBM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e86.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e80.26\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e89.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e78.71\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.931\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eAdaBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e84.65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e85.53\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u003cp\u003e84.24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colspan=\"3\" nameend=\"c9\" namest=\"c7\"\u003e\u003cp\u003e77.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e0.897\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=\"BlockQuote\"\u003e\u003cp\u003eRandom Forest achieved the highest performance across all five evaluation metrics for DKD risk prediction using the Rwandan dataset. In addition to these metrics, Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e10\u003c/span\u003e presents the ROC-AUC curves to further illustrate model performance. Figure\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e10\u003c/span\u003e (a) displays the ROC-AUC curves for the five ensemble models on the MIMIC-IV dataset, while Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e10\u003c/span\u003e (b) shows the corresponding curves for the Rwandan dataset.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe Random Forest model demonstrated superior performance on both datasets. Figure\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e11\u003c/span\u003e presents its confusion matrix, which was used to compute the evaluation metrics.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eTo further evaluate the strong predictive performance of the tree-based models on the Rwandan DKD dataset, calibration curves were plotted, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e12\u003c/span\u003e. The dashed black line represents perfect calibration, where the predicted probabilities align with the actual outcomes. Models below this line are overconfident, while those significantly above are underconfident. A well-calibrated model closely follows the diagonal. Beyond the 70% predicted probability threshold for the positive class (DKD), all models align closely with the diagonal except AdaBoost, which remains consistently below, indicating overconfidence.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003e3.3 SHAP Explanations\u003c/h2\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eThe effectiveness of the tree-based ensemble models used in this study was further assessed using the XAI explainability method. The FastTreeSHAP approach was used to investigate how the best model on both datasets makes a decision on identifying the diabetic patients at risk of DKD development. Figure\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e13\u003c/span\u003e provides a global view of feature contribution to the model decision using the MIMIC-IV dataset.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eAccording to the MIMIC-IV dataset, the presence of complications related to diabetes is the leading predictor of diabetes kidney disease risk, followed by creatinine. Gender is the least predictor of DKD risk. Figure\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e14\u003c/span\u003e highlights the top 15 features contributing to the risk of DKD among diabetes patients in Rwanda.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eSHAP methods applied to a Random Forest trained on the Rwanda DKD dataset revealed the top 15 features highly contributing to the model's prediction of DKD risk. The blood urea nitrogen (Urea) is the leading predictor of DKD, followed by Hemoglobin. Among 15 features linked to DKD risk, diastolic blood pressure is ranked the lowest. SHAP methods were also used to assess the model's behavior for individual diabetic patients, providing a clear clinical interpretation of why a particular patient is predicted to be at risk or not at risk of DKD. Figure\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e illustrates the model interpretation at the patient level. Figure\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e (a) details the reason why the patient with index 19 in the test set of the Rwanda dataset was predicted at high risk of DKD, and Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e (b) shows that the patient with index 11 in the test set of the Rwanda dataset is free from DKD risk.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eExperimental results demonstrated that the Random Forest model outperformed other models on both the MIMIC-IV and the dataset from three hospitals in Rwanda. Configured as RandomForestClassifier(random_state\u0026thinsp;=\u0026thinsp;42, n_estimators\u0026thinsp;=\u0026thinsp;300, max_depth\u0026thinsp;=\u0026thinsp;6, bootstrap\u0026thinsp;=\u0026thinsp;True), it achieved an accuracy of 87.97% on the MIMIC-IV dataset, ranking highest in three out of five evaluation metrics. On the Rwanda dataset, it achieved 91.70% accuracy and led in all five metrics. These findings suggest that the model can serve as a reliable tool for the initial screening of diabetic kidney disease (DKD) at the onset of patient follow-up.\u003c/p\u003e\u003cp\u003eThe MIMIC-IV dataset exhibited higher missing value rates in key features such as HbA1c (28%), total cholesterol (TC) (38%), and BMI (18%) compared to the Rwanda dataset, which may have contributed to the slightly lower model performance. Missing values were imputed using the mean method. The Random Forest model achieved higher sensitivity and specificity on the Rwanda dataset, indicating a strong ability to correctly identify both DKD and non-DKD cases. These metrics are critical: failing to detect DKD may allow disease progression to advanced stages, while false positives could lead to unnecessary treatment and potential side effects. Model reliability was further confirmed through calibration analysis, with the Random Forest curve closely aligning with the diagonal, indicating good calibration.\u003c/p\u003e\u003cp\u003eSHAP analysis of the Random Forest model provided valuable insights to support clinical interpretation and build trust in the model\u0026rsquo;s predictions. SHAP identified the most influential features contributing to DKD risk, with three of the top five features consistent across both datasets and 14 of the top 15 shared, though ranked differently, demonstrating the model\u0026rsquo;s generalizability across datasets with varying distributions. SHAP also enabled interpretation at the individual patient level. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e(a), features highlighted in red contribute to increased DKD risk, while those in blue reduce it. For the patient examined, abnormal values in hemoglobin (11.8), creatinine (2.4), blood glucose (217.12), RBC (7.15), and neutrophils (64.5) support the high-risk prediction.\u003c/p\u003e\u003cp\u003eAdditionally, the patient\u0026rsquo;s eGFR was calculated using the method described in [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], yielding a value of 20. Since an eGFR\u0026thinsp;\u0026ge;\u0026thinsp;90 indicates normal kidney function, this low value supports the model\u0026rsquo;s predicted high risk of DKD, with a probability of 0.946, as indicated by f(x) in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e(a). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, this corresponds to stage G4 DKD, indicating severely reduced kidney function. The SHAP interpretation in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e(b) explains why the patient was predicted not to be at risk of DKD. Further analysis revealed that the patient described in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e15\u003c/span\u003e (b) is a 58-year-old male with a normal creatinine level of 0.69, as indicated in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Additionally, the eGFR, calculated using the method in [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], was 104.64, confirming a normal kidney function and supporting the model\u0026rsquo;s prediction.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"5 Conclusion and Future Work","content":"\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eEarly detection of diabetic kidney disease (DKD) benefits both patients and healthcare providers, as DKD is a progressive condition often driven by uncontrolled blood glucose and hypertension, which impair kidney function. This study demonstrates that explainable machine learning can effectively utilize electronic health record (EHR) data to identify individuals at high risk of DKD. Five tree-based ensemble models were developed and evaluated on both the MIMIC-IV and Rwanda datasets. Due to class imbalance, the SMOTE technique was applied to balance the data. Among the models, Random Forest achieved the highest performance, with accuracies of 87.97% on MIMIC-IV and 91.70% on the Rwanda dataset. All models were evaluated using multiple metrics, including ROC-AUC and calibration curves, to ensure robustness and interpretability.\u003c/p\u003e\u003cp\u003eThe best-performing model was further analyzed using SHAP methods to interpret its predictions. SHAP was used to visualize both global feature importance and individual-level explanations through waterfall plots. A ranked list of the top 15 features contributing to DKD risk in the Rwandan dataset was generated. At the individual level, predictions were validated using estimated glomerular filtration rate (eGFR) values, which aligned with the model\u0026rsquo;s outputs, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. This study had two key objectives: (1) to train and evaluate machine learning models on a publicly available dataset and DKD data from three Rwandan hospitals, and (2) to interpret the most effective model using SHAP for early DKD detection. These objectives were successfully achieved.\u003c/p\u003e\u003cp\u003eThe findings of this study can be integrated into clinical workflows to support early detection of DKD in diabetic patients, contributing to more effective and timely care and improving patients' quality of life. We recommend that Rwandan hospitals enhance the completeness of diabetes patient records, particularly for key variables such as BMI, total cholesterol (TC), blood urea nitrogen (Urea), HbA1c, systolic blood pressure (SBP), and diastolic blood pressure (DBP), which showed over 10% missing values but were among the top 15 predictors of DKD in the Rwandan dataset. Future work will focus on implementing these findings in clinical decision-making processes at CHUK, Ruhengeri, and Nyamata hospitals.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eAuthors contribution\u003c/p\u003e\n\u003cp\u003eAll authors contributed equally to the manuscript. Silas led the research design, data collection, experimentation, and writing. Tony and Celestin supervised the study, contributed to manuscript refinement, and analyzed the results. Fiacre, drawing on extensive experience in nursing, particularly in chronic kidney disease at the nephrology department, extracted data from OpenMRS and OpenClinic systems in three participating hospitals in Rwanda and assessed the clinical relevance of our study.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAcknowledgement\u003c/p\u003e\n\u003cp\u003eThe authors gratefully acknowledge the support provided by the University of Rwanda and Stockholm University. Special thanks are due to the PhysioNet team, as well as the data managers from Ruhengeri, Nyamata, and CHUK hospitals, for their valuable support during the data collection process. \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eEthical Statement\u003c/p\u003e\n\u003cp\u003eThis study relied on secondary data derived from electronic health records (EHRs) and did not involve the use of any personally identifiable information. The MIMIC-IV dataset, which is publicly accessible, does not require ethical approval. The dataset from Rwanda consisted of secondary EHR data collected from three hospitals. Access to these data was granted under ethical approval reference CMHS/IRB/207/2024 by the Institutional Review Board of the College of Medicine and Health Sciences, University of Rwanda. All data processing and analysis were conducted in accordance with Rwandan personal data protection and privacy regulations, as well as the requirements of the Health Insurance Portability and Accountability Act (HIPAA) and the ethical principles outlined in the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003eInformed consent\u003c/p\u003e\n\u003cp\u003eThe study utilized secondary data extracted from electronic health record (EHR) systems to develop machine learning models for predicting diabetic kidney disease. As no data were collected directly from individual patients, informed consent was not required. Patient involvement was indirect, and the Ethics Review Boards granted a waiver of informed consent. The study was conducted in accordance with the principles outlined in the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003eData availability statement\u003c/p\u003e\n\u003cp\u003eThe original contributions presented in the study are included in the article and supplementary materials. For further inquiries, please contact the corresponding authors.\u003c/p\u003e\n\u003cp\u003eConflict of interest\u003c/p\u003e\n\u003cp\u003eThe authors declare that there is no conflict of interest associated with this study.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis research was supported by the University of Rwanda and Stockholm University under the UR-Sweden Program for research and capacity building.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eGenerative AI statement\u003c/p\u003e\n\u003cp\u003eThe authors declare that no Generative AI was used in the creation of this manuscript.\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eOgurtsova K et al. IDF diabetes atlas: Global estimates of undiagnosed diabetes in adults for 2021. Diabetes Res Clin Pract, 183, 2022.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eArt 109118JD, Kabakambira P, Shumbusho G, Mujawamariya W, Rutagengwa M, Twagirumukiza. The Role of the Integrated District Hospital Based Non Communicable Diseases\u0026rsquo; Clinics in Cardiovascular Disease Control: Preliminary Data from Rwanda. Dovepress. 2022;15:2107\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQiong B, Chunyan S, Wen T, Yike L. Machine learning to predict end stage kidney disease in chronic kidney disease. Sci Rep. 2022;12:8377.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLevin A, Stevens PE. Early detection of CKD: the benefits, limitations and effects on prognosis. Nat Rev Nephrol. 2011;7:446\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlowais SA, Alghamdi SS, Alsuhebany N, Alqahtani T, Alshaya AI, Almohareb SN, et al. Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC Med Educ. 2023;23:689.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhouyu G, Huating L, Ruhan L, Chun C, Yuexing L, Jiajia L, et al. Artificial intelligence in diabetes management: Advancements, opportunities, and challenges. Cell Rep Med. 2023;4:101213.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSingh P, Goyal L, Mallick DC, Surani SR, Kaushik N, Chandramohan D, Simhadri PK. Artificial Intelligence in Nephrology: Clinical Applications, and Challenges. Kidney Med. 2025;7:100927.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePreti LM, Ardito V, Compagni A, Petracca F, Cappellaro G. Implementation of Machine Learning Applications in Health Care Organizations: Systematic Review of Empirical Studies, JMIR, 2024, 26, 55897.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSridevi P, Meikandan PV, Upama PB, Rabbani M, Alam KS, Das D, et al. ML-Based Chronic Kidney Disease and Diabetes Prediction with Feature Effect Analysis Using SHAP. IEEE COMPSAC. 2024;48:843\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlistair EW, Lucas B, Lu S, Alvin G, Ayad S, Steven H, et al. Author Correction: MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10:1.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eClinic M. The Essential Diabetes Book, 2nd Edition, ISBN-10: 0-8487-4339-3, 2014, p. 53.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ehttps://\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e\u003c/span\u003e\u003cspan address=\"http://www.kidney.org/kidney-topics/kidney-function\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. (Accessed on June 25, 2025).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eClinic M. The Essential Diabetes Book, 2nd Edition, ISBN-10: 0-8487-4339-3, 2014, p. 42.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSelden O, Rusingiza E, Balkrishnan R. Overview, Infrastructural Challenges, Barriers to Access, and Progress for Rwanda\u0026rsquo;s Healthcare System: A Review. Integr J Med Sci. 2025;12:745.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePersson F, Rossing P. Diagnosis of diabetic kidney disease: state of the art and future perspective. Kidney Int Supplements. 2018;8:2\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDugan J, Shubrook J. International Classification of Diseases, 10th Revision, Coding for Diabetes, Clinical diabetes: a publication of the American Diabetes Association, 2017, 35, 232\u0026ndash;238.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFernandez A, Garcia S, Herrera F. ISMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary. J Artif Intell Res. 2018;61:863\u0026ndash;905.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNational Institute of Statistics of Rwanda (NISR), Fourth Population and Housing Census. 2012, 143\u0026ndash;144.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYadu S, Chandra R, Sinha VK. Comparing Different Machine Learning Techniques in Predicting Diabetes on Early Stage, MDPI Engineering Proceedings, 2024, 62, 20. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/engproc2024062020\u003c/span\u003e\u003cspan address=\"10.3390/engproc2024062020\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGhosh SK, Khandoker AH. Investigation on explainable machine learning models to predict chronic kidney diseases. Nat Sci Rep. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-024-54375-4\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-54375-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu XZ, Duan M, Huang HD, Zhang Y, Xiang TY, Niu WC, et al. Predicting diabetic kidney disease for type 2 diabetes mellitus by machine learning in the real world: a multicenter retrospective study. Front Endocrinol. 2023;14:1184190. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fendo.2023.1184190\u003c/span\u003e\u003cspan address=\"10.3389/fendo.2023.1184190\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRenuka P, Thilagavathi C, Sathiyanathan S, Murugesh L, Madhumitha M. Adaboost Machine Learning based Modelling to Predict Chronic Kidney Disease Staging. ICSTEM. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ICSTEM61137.2024.1056088\u003c/span\u003e\u003cspan address=\"10.1109/ICSTEM61137.2024.1056088\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDong Z, Wang Q, Yujing K, Zhang W, Hong Q, Liu C, Xiaomin L, et al. Prediction of 3-year risk of diabetic kidney disease using machine learning based on electronic medical records. J Translational Med. 2022;20:143. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s12967-022-03339-1\u003c/span\u003e\u003cspan address=\"10.1186/s12967-022-03339-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMuschalik M, Fumagalli F, Hammer B, Hullermeier E. Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles, 38th AAAI Conference on Artificial Intelligence, 2024, 1604, 14388\u0026ndash;14396.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLundberg SM, Lee S. A Unified Approach to Interpreting Model Predictions. NIPS. 2017;31:4768\u0026ndash;77.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJing-Mei Y, Yang L, Jun-Tang X, Jun-Tang Z, Guo-Wei Z, Zhong-Ze F, Lang Z. Explainable Machine Learning-Based Prediction Model for Diabetic Nephropathy. J Diabetes Res. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1155/2024/8857453\u003c/span\u003e\u003cspan address=\"10.1155/2024/8857453\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHongxia X, Yonghui K, Shaofeng T. Predictive Modeling of Diabetic Kidney Disease using Random Forest Algorithm along with Features Selection. ISAIMS. 2020;1:23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e- https://doi.org/10.1145/3429889.3429894\u003c/span\u003e\u003cspan address=\"- 10.1145/3429889.3429894\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSridevi P, Padmapriya M et al. ML-based Chronic Kidney Disease and Diabetes Prediction with Feature Effect Analysis Using SHAP, IEEE 48th Annual Computers, Software, and Applications Conference (COMPSAC), 2024, 48.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMohsin RS et al. Chronic Renal Disease Prediction using Clinical Data and Different Machine Learning Techniques, International Conference on Informatics and Software Engineering (IISEC), 2021, 2, 1\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAllen A, et al. Prediction of diabetic kidney disease with machine learning algorithms, upon the initial diagnosis of type 2 diabetes mellitus. BMJ Open Diabetes Res Care. 2022;10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmjdrc-2021-002560\u003c/span\u003e\u003cspan address=\"10.1136/bmjdrc-2021-002560\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eReddy S, et al. Predicting chronic kidney disease progression using small pathology datasets and explainable machine learning models. Computer Methods and Programs in Biomedicine Update; 2024.\u003c/span\u003e\u003c/li\u003e\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-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Diabetic Kidney Disease, Diabetes Management, Ensemble Models, Explainable Artificial Intelligence, Machine Learning, SHAP","lastPublishedDoi":"10.21203/rs.3.rs-8108979/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8108979/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe global prevalence of diabetes is increasing, often leading to complications such as diabetic kidney disease (DKD). Uncontrolled blood glucose and hypertension are key risk factors that progressively impair kidney function, potentially resulting in kidney failure or end-stage renal disease (ESRD). Early detection of DKD is crucial but challenging due to its asymptomatic onset. This study employs explainable artificial intelligence (XAI) to predict DKD risk in diabetic patients using tree-based ensemble models and SHAP (Shapley Additive exPlanations), leveraging the MIMIC-IV dataset and a dataset from hospitals in Rwanda. Among the models used, Random Forest demonstrated superior performance, achieving accuracies of 87.97% on MIMIC-IV and 91.70% on the Rwandan dataset. Models were evaluated using multiple metrics, including ROC-AUC and calibration curves. SHAP provided both global and individual-level explanations, with predictions validated using estimated glomerular filtration rate (eGFR) values. Our findings highlight the promising potential of integrating predictive modeling with explainability to develop transparent and trustworthy tools for early detection of DKD, with potential applications in clinical workflows.\u003c/p\u003e","manuscriptTitle":"Using Explainable Machine Learning for Early Detection of Diabetic Kidney Disease in Rwandan Diabetic Patients","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-01 10:39:47","doi":"10.21203/rs.3.rs-8108979/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-12-26T11:23:22+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-22T17:32:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"264422805205987033181895088115900979508","date":"2025-12-21T13:28:43+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-12-08T13:06:40+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"60837422618625309102854515519747550343","date":"2025-12-06T07:54:49+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-26T09:52:53+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-11-21T06:34:40+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-11-19T08:49:11+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-11-19T08:48:50+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2025-11-13T21:06:53+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d36f4259-b560-4808-ae5b-d8bf7643c417","owner":[],"postedDate":"December 1st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-06T21:38:41+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-01 10:39:47","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8108979","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8108979","identity":"rs-8108979","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}