Development and Internal Validation of an Interpretable Machine Learning Model for Predicting Atrial Fibrillation in Patients with Diabetic Kidney Disease: A Multicenter Study

Development and Internal Validation of an Interpretable Machine Learning Model for Predicting Atrial Fibrillation in Patients with Diabetic Kidney Disease: A Multicenter Study | 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 Development and Internal Validation of an Interpretable Machine Learning Model for Predicting Atrial Fibrillation in Patients with Diabetic Kidney Disease: A Multicenter Study Xiaoran Li, Xueying Wang, Shidong Wang, Xuebing Zhang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8146262/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Patients with diabetic kidney disease (DKD) have elevated atrial fibrillation (AF) risk, yet population-specific prediction tools are limited. We aimed to develop and internally validate an interpretable machine-learning (ML) model for AF risk in hospitalized DKD. Methods In this retrospective cohort from two hospitals (January 2021 to December 2024), 787 unique DKD admissions were randomly split into training (70%) and test (30%) sets. AF at index admission was ascertained from electrocardiograms, Holter monitoring when available, and ICD-10 codes with physician adjudication. Candidate predictors were routine clinical, laboratory, and echocardiographic variables. Least absolute shrinkage and selection operator (LASSO) selected features in the training set. Seven supervised models were trained; performance was assessed by area under the receiver-operating characteristic curve (AUC), calibration, and decision-curve analysis. SHAP quantified predictor contributions. Results LASSO retained 14 features, including 24-hour urine total protein (24UTP), serum creatinine (SCr), age, and left atrial diameters. In the test set, k-nearest neighbors (KNN) achieved AUC 0.927, accuracy 0.886, sensitivity 0.920, and specificity 0.856; calibration was good and decision curves showed net benefit across common thresholds. Five-fold cross-validation yielded mean AUC 0.90 ± 0.02. SHAP indicated proteinuria burden, renal dysfunction, age, and atrial size as leading contributors. The finalized model was deployed as a secure web calculator using routine inputs. Conclusions An interpretable ML-based model using standard clinical and echocardiographic data showed stable internal performance for AF risk estimation in DKD, with an accompanying web calculator for point-of-care use. Prospective multicenter studies are needed to confirm generalizability and clinical impact. Diabetic kidney disease (DKD) Atrial fibrillation (AF) Risk prediction Machine learning (ML) Shapley Additive Explanations (SHAP) Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction Diabetic kidney disease (DKD), one of the most common microvascular complications of diabetes, is a leading cause of end-stage renal disease and renal replacement therapy [ 1 ]. DKD not only markedly increases the risk of kidney failure but is also closely associated with thrombosis, stroke, myocardial infarction, and all-cause mortality [ 2 – 5 ]. The rising prevalence of DKD poses a major challenge to global health systems [ 6 ]. Atrial fibrillation (AF), one of the most frequent cardiac arrhythmias, has increased in prevalence by 33% over the past two decades [ 7 ]. Patients with AF have a 61% higher risk of all-cause death and face greater risks of cardiovascular mortality and stroke [ 8 ]. Studies have shown that individuals with type 2 diabetes have a 35% higher risk of developing AF than the general population, and the risk further increases among patients with DKD as estimated glomerular filtration rate (eGFR) declines and urinary protein excretion rises [ 9 , 10 ]. DKD and AF share several risk factors, including hypertension and coronary artery disease. Both conditions are influenced by thrombosis, systemic inflammation, and activation of the renin-angiotensin system (RAS), forming a bidirectional relationship [ 11 ]. On one hand, eGFR decline and proteinuria in DKD may promote AF onset [ 12 , 13 ]; on the other hand, diabetic patients with AF are more likely to develop DKD, and AF can accelerate renal function deterioration and proteinuria progression in DKD [ 14 , 15 ]. This mutual reinforcement creates a vicious cycle leading to poor clinical outcomes. In clinical practice, AF is often paroxysmal or asymptomatic in its early stage. Silent AF is frequently underdiagnosed or diagnosed late [ 16 , 17 ], especially in patients with DKD, who may lack overt arrhythmic symptoms. As a result, opportunities for early rhythm control and intervention are easily missed, which may accelerate the progression of both DKD and AF and increase the risk of serious cardiovascular and cerebrovascular events. These challenges highlight the need for effective risk assessment and early identification of AF in patients with DKD to improve clinical outcomes. Despite the growing clinical burden, no dedicated AF screening or prediction model has been developed for the DKD population. Existing tools show limited accuracy in identifying AF among these patients. Previous studies have mainly focused on the general population or isolated AF prediction. For example, some investigators combined clinical expertise with artificial intelligence (AI) analysis of 24-hour Holter recordings to improve AF detection rates [ 18 ]. However, such approaches are difficult to implement widely because DKD patients often lack routine Holter monitoring. Other models based on AI analysis of sinus-rhythm electrocardiograms can predict paroxysmal AF [ 19 ], but their performance remains limited without incorporating clinical and laboratory data. Given the increasing prevalence and clinical significance of AF in DKD and the diagnostic limitations of current methods, a reliable and interpretable risk prediction model is urgently needed. In this study, we developed and internally validated a machine learning (ML) based model integrating clinical, laboratory, and echocardiographic variables to predict AF in hospitalized patients with DKD. Our goal was to provide a practical and interpretable tool for early, accurate identification of high-risk patients and to support clinical decision making and management. 2 Material and methods 2.1 Study design and setting This retrospective cohort study included adult inpatients with DKD who were admitted to Dongzhimen Hospital, Beijing University of Chinese Medicine, and Beijing Electric Power Hospital between January 2021 and December 2024. A total of 814 admissions were screened, and 787 unique records met eligibility criteria and were included in the analysis. The dataset was randomly divided into a training set (70%) and an internal test set (30%) using stratified sampling to preserve the proportion of outcomes. Data were extracted by authorized personnel from identifiable electronic medical records to ensure accurate linkage, and all records were de-identified before analysis. Ethical approval was obtained from the institutional review boards of both hospitals (Dongzhimen Hospital and Beijing Electric Power Hospital; Approval Nos. 2025DZM-376). Since this study is a retrospective study, the need for consent to participate was waived by an Institutional Review Board. All procedures followed the principles of the Declaration of Helsinki and complied with applicable national regulations. An overview of the study workflow is shown in Fig. 1 . 2.2 Endpoint definition The primary endpoint was the presence of AF at the index admission. AF was identified using a combination of diagnostic sources, including a standard 12-lead electrocardiogram, 24-hour Holter monitoring when available, and ICD-10 diagnostic codes extracted from the electronic medical record. Each case was independently reviewed and confirmed by two physicians to ensure diagnostic accuracy. The prediction time point was defined as the earliest availability of all candidate predictors within 72 hours after hospital admission. 2.3 Participant selection Participants were identified according to the Chinese Clinical Guidelines for DKD[ 20 ] and the Chinese Guidelines for the Management of AF (2025 edition)[ 21 ]. Eligible patients were required to meet all of the following inclusion criteria: (1) age between 18 and 95 years (inclusive); (2) any sex; (3) diagnosis of diabetes mellitus consistent with Chinese and World Health Organization criteria; and (4) diagnosis of diabetic kidney disease confirmed by clinical and laboratory findings. Exclusion criteria included (1) type 1 diabetes, gestational diabetes, or other specific forms of diabetes; (2) coexisting primary renal diseases such as primary glomerulonephritis, lupus nephritis, or Henoch Schönlein nephritis; (3) ongoing glucocorticoid or immunosuppressive therapy; (4) concomitant malignant tumors, decompensated liver cirrhosis, active tuberculosis, infectious shock, disseminated intravascular coagulation, or other critical or terminal illnesses; and (5) major limb amputation within the preceding six months. Additional cardiac related exclusions were (1) pacemaker or implantable cardioverter defibrillator implantation; (2) other clinically significant tachyarrhythmias such as atrial flutter or sustained supraventricular tachycardia; and (3) a history of cardiac surgery including valve replacement or coronary artery bypass grafting. A previous diagnosis of atrial fibrillation was not an exclusion criterion because the endpoint focused on AF status at baseline rather than new onset cases. 2.4 Variable collection and measurements Information extracted from the electronic medical record included three domains: (1) demographics and clinical data, such as sex, age, height, weight, blood pressure, and diabetes duration; (2) laboratory tests, including complete blood count, urinalysis, biochemical profile, arterial blood gas, coagulation parameters, thyroid function, tumor markers, cardiac enzymes, and heart failure biomarkers; and (3) transthoracic echocardiography findings. The final predictor set used for model construction consisted of routinely available variables: HbA1c (glycated hemoglobin, %), BUN (blood urea nitrogen, mmol/L), SCr (serum creatinine, µmol/L), BNP (B-type natriuretic peptide, pg/mL), ALT (U/L), AST (U/L), CK-MB (U/L), Myo (myoglobin, ng/mL), WBC (×10⁹/L), RBC (×10¹²/L), 24UTP (24-hour urine total protein, g/24 h), INR, APTT (s), FIB (g/L), LAAPD (left atrial anterior–posterior diameter, mm), LAMLD (left atrial medial–lateral diameter, mm), LASID (left atrial superior–inferior diameter, mm), RAMLD (right atrial medial–lateral diameter, mm), RASID (right atrial superior–inferior diameter, mm), EF (%), and FS (%), plus sex. Echocardiography was performed by certified sonographers following a standardized institutional protocol using the same equipment platform (device and model to be specified). LAAPD was measured in the parasternal long-axis view at end-systole. LAMLD, LASID, and their right-atrial counterparts RAMLD and RASID were obtained from orthogonal apical views in medial lateral and superior inferior planes. All diameters were recorded from inner edge to inner edge and averaged over three cardiac cycles. A random 10% subset was re-measured to evaluate intra- and inter-observer consistency, with a target variability below 5%. 2.5 Data preprocessing and handling of missing data Variables with more than 20% missingness were excluded a priori. For the remaining predictors, for which overall missingness was typically less than 10%, missing values were imputed using the MissForest algorithm with default depth and 100 trees. Imputation was implemented within the training set only to prevent information leakage, and the trained imputer was then applied to the held-out test set. Categorical harmonization followed a unified data dictionary, and rare categories with a prevalence below 2% were merged. 2.6 Feature selection, model development, and hyper parameter tuning Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) regression in the training dataset. A ten-fold cross-validation procedure was applied with the glmnet package in R to identify a parsimonious subset of variables with the strongest predictive contribution. The optimal penalty parameter (λ) was chosen according to the minimum cross-validated deviance criterion, and the selected features were used for subsequent model construction. Seven supervised machine learning algorithms were developed on the selected predictors: gradient boosting decision tree (GBDT), k-nearest neighbors (KNN), linear discriminant analysis (LDA), Light Gradient Boosting Machine (LightGBM), random forest (RF), artificial neural network (ANN), and extreme gradient boosting (XGBoost). Each algorithm was trained separately to capture different model structures and learning strategies. Hyper parameters were tuned using stratified five-fold cross-validation within the training dataset through either grid search or randomized search. All preprocessing, feature selection, and tuning steps were restricted to the training data to prevent data leakage. Because the prevalence of atrial fibrillation was approximately balanced, no oversampling, undersampling, or class weighting was applied. Decision thresholds for secondary analyses were predefined before model evaluation. 2.7 Model evaluation, validation, interpretability, and software. Model performance was evaluated in both the training and test datasets. Discrimination was assessed using the area under the receiver-operating characteristic curve (AUC), with accuracy, sensitivity, specificity, precision, and F1-score as complementary metrics. Calibration was evaluated by the Brier score, calibration intercept and slope, and visual inspection of calibration plots generated with loess smoothing. Decision-curve analysis (DCA) was applied to estimate the net clinical benefit across prespecified probability thresholds ranging from 5% to 20%. Clinical-impact curves were used to illustrate the number of individuals classified as high risk and the corresponding number of true positives per 1,000 patients at representative thresholds. Model interpretability was examined for the k-nearest neighbors classifier using KernelSHAP, which quantified the contribution of each predictor to individual predictions. SHAP values were computed on standardized inputs with a background set of 100 training instances selected by k-medoids sampling to approximate the empirical data distribution. Descriptive statistics were performed in SPSS 25.0. Penalized regression for feature selection was conducted in R 4.4.1. All machine-learning models and visualizations, including receiver-operating characteristic, calibration, decision, clinical-impact, and SHAP plots, were implemented in Python 3.10.4 using standard scientific libraries. The finalized prediction model was further deployed as a secure web-based calculator. This tool allows clinicians to input routinely available variables and obtain individualized AF risk predictions in real time. The calculator operates within the institutional network under encrypted communication and restricted access, ensuring data confidentiality. It is intended to support, rather than replace, clinical decision making. 3 Results 3.1 Demographic and baseline characteristics A total of 787 patients with diabetic kidney disease were included and randomly divided in a 7:3 ratio into the training set (n = 550) and the test set (n = 237). No significant differences were observed between the two sets in age, diabetes duration, HbA1c, SCr, BNP, EF, or sex distribution (male 41.8% vs. 41.5%, P = 0.934), confirming good comparability between sets (Table 1). Within the training set, patients with AF differed markedly from those without AF (Table 2). The AF group was older (P < 0.001), had a longer diabetes duration (P = 0.013), and higher body weight (P < 0.001). Levels of BUN, SCr, CK-MB, and Myo were significantly higher in the AF group (all P < 0.01), whereas BNP was notably lower (P < 0.001). Echocardiographic measurements showed smaller atrial diameters (LAAPD, LAMLD, LASID, RAMLD, and RASID) but higher EF and FS values (all P < 0.001). Additionally, women were more prevalent in the AF group (69.2% vs. 48.8%, P < 0.001). Collectively, AF patients displayed distinct metabolic and cardiac profiles, suggesting links between atrial remodeling, renal dysfunction, and AF risk in DKD. 3.2 Feature selection Feature selection was performed using the LASSO regression with tenfold cross-validation in the training set. This method applies L1 regularization to shrink less informative coefficients toward zero, thereby retaining only the most predictive variables. Model performance was monitored across iterations to identify the subset with the best cross-validated performance. As shown in Figure 2A, the mean cross-validation curve indicated stable convergence, and the corresponding coefficient profiles are presented in Figure 2B. Fourteen variables with non-zero coefficients were ultimately retained: 24UTP, SCr, age, LAMLD, weight, LASID, LAAPD, RAMLD, FS, BUN, HbA1c, CK-MB, FIB, and INR. These features exhibited the strongest contribution to AF prediction and were subsequently used for model development. 3.3 Model performance comparison Seven supervised machine-learning algorithms were developed using the selected features: KNN, RF, XGBoost, LightGBM, GBDT, LDA, and ANN. In the training set, all models showed good discrimination (Figure 3A). The KNN model achieved the highest AUC (0.933) with an accuracy of 0.8745, sensitivity of 0.9077, specificity of 0.8448, and a Brier score of 0.100. RF, XGBoost, and LightGBM performed comparably, with AUCs of 0.9200, 0.9196, and 0.9144, respectively (Table 3). In the test set (Figure 3D), KNN maintained superior performance with an AUC of 0.9271, accuracy of 0.8861, sensitivity of 0.9196, and specificity of 0.8560. XGBoost achieved the lowest Brier score (0.094), while RF and LightGBM yielded AUCs of 0.8955 and 0.8791, respectively. Although LightGBM had a slightly higher Brier score (0.123), its calibration remained acceptable (Table 3). Calibration curves showed good agreement between predicted and observed probabilities for all models (Figures 3B and 3E). Decision-curve analysis indicated that KNN and XGBoost provided the greatest net clinical benefit across most threshold ranges (Figures 3C and 3F). Five-fold cross-validation further confirmed the robustness of model performance, with an average AUC of 0.90 ± 0.02 (Figure S1). Taken together, KNN demonstrated the most favorable balance of discrimination, calibration, and generalization, indicating a favorable balance of discrimination, calibration, and generalization. 3.4 Model interpretation and SHAP analysis To explore the interpretability of the KNN model and quantify the contribution of each feature, SHAP analysis was applied (Figure 4). The mean absolute SHAP value plot (Figure 4A) identified age, SCr, LAAPD, and 24UTP as the strongest predictors of AF. Among them, 24UTP showed the highest average SHAP value, indicating the greatest overall influence on model output. The SHAP summary plot (Figure 4B) illustrated the distribution of feature effects across all samples. Each point represents a single observation, with color indicating the feature value (red for high, blue for low) and position reflecting its positive or negative contribution to prediction. Features with low values exhibited stronger positive impacts on predicted AF risk, whereas higher values showed weaker effects. To visualize individual-level predictions, representative force and waterfall plots were generated (Figures 4C-F). These plots demonstrated how each feature shifted the prediction toward or away from AF, highlighting the cumulative contribution of multiple risk factors. Overall, the results suggested that increased proteinuria burden and impaired renal function were the primary drivers of AF risk in patients with DKD. 3.5 Clinical impact curve analysis To assess the clinical utility of the model, clinical impact curves (CICs) for the KNN classifier were generated in both the training and test sets (Figure 5A and 5B). As the decision threshold increased, the number of patients classified as high risk (blue line) gradually decreased. The number of true positives (red line) closely overlapped with the number of observed AF cases (green dashed line), indicating that the predicted probabilities were well aligned with actual outcomes. Across the common range of risk thresholds, the KNN model consistently demonstrated favorable screening efficiency and net clinical benefit. The similarity between the training and test set curves further supported robustness and generalizability, suggesting stable performance on new data. 3.6 Web calculator application The final KNN prediction model was integrated into a web-based calculator to facilitate clinical application (Figure 6). The interface is designed for straightforward clinical use: clinicians can input 14 routinely available clinical and laboratory variables and obtain an individualized AF risk prediction in real time. The output provides a visual estimate of predicted probability together with a categorical risk label to assist in patient stratification. The calculator operates on a secure institutional server with encrypted communication and restricted access. It is intended as a clinical support tool rather than a decision making substitute, aiming to improve the accessibility and timeliness of AF risk assessment in patients with DKD. The platform also allows for potential extension to multicenter validation and remote clinical use. The tool can be accessed at: https://endocrine-dzm-af-prediction-model.shinyapps.io/shiny/. 4 Discussion To our knowledge, this is the first clinical prediction model for atrial fibrillation (AF) in patients with diabetic kidney disease (DKD) developed using machine learning (ML). Using routinely collected information from electronic medical records, laboratory tests, and transthoracic echocardiography, we compared seven ML algorithms and identified k-nearest neighbors (KNN) as the best-performing approach in our dataset. KNN is a nonparametric, supervised method that classifies a query sample by the votes of its nearest neighbors, requires few hyperparameters, and is straightforward to implement [ 22 ]. Prior studies have used KNN for risk prediction in sleep apnea, coronary artery disease, and AF among intensive care unit patients [ 23 , 24 ]. To improve transparency, we applied SHAP to interpret model outputs and quantify feature contributions, consistent with earlier AF modeling studies [ 25 , 26 ]. Key predictors highlighted by SHAP were clinically plausible and measurable at the point of care. We also implemented a web calculator to support bedside use, given that the inputs are standard components of routine evaluation. Building on the model’s overall performance, the selected predictors reflected two major patterns: renal metabolic burden and cardiac structural functional remodeling, together outlining a biologically coherent framework for AF risk in DKD. Indicators of renal impairment and proteinuria emerged as dominant signals. Higher SCr and greater 24UTP were consistent with prior studies showing that kidney dysfunction and protein loss are linked to higher AF risk, often in a dose response pattern. [ 12 , 13 , 27 ]. BUN showed similar associations and has also been retained in critical care AF prediction models [ 25 , 28 ]. These findings support mechanisms involving systemic inflammation, activation of the renin–angiotensin–aldosterone system (RAAS), and disturbances in calcium homeostasis [ 29 ]. In parallel, aging and obesity were strong clinical correlates. The risk of AF increases steadily with advancing age, likely due to atrial remodeling, reduced conduction velocity, and sinoatrial node dysfunction [ 30 – 33 ], while excess body weight is associated with both incident AF and its progression from paroxysmal to persistent forms through mechanisms such as hemodynamic overload, hypertension, inflammation, and oxidative stress [ 34 – 37 ]. Obesity is also strongly linked to the onset and progression of DKD, highlighting a shared metabolic and structural risk pathway [ 38 – 40 ]. Meanwhile, cardiac structure and function further complemented these metabolic indicators. Enlarged atrial dimensions on echocardiography correlated with declining eGFR and higher cardiovascular risk among DKD patients [ 41 – 44 ]. Left atrial volume index was independently associated with new-onset AF and adverse outcomes regardless of ejection fraction or chamber geometry, consistent with chronic volume load, inflammation, and neurohormonal activation [ 45 – 48 ]. For the right atrium, genetic evidence from Mendelian randomization suggests that reduced eGFR may causally contribute to atrial enlargement, although DKD-specific clinical studies remain limited [ 49 ]. Reduced systolic performance, reflected by fractional shortening, was also consistent with the interlinked pathophysiology of AF, heart failure, and DKD [ 50 – 54 ]. Finally, other variables such as HbA1c, CK-MB, fibrinogen, and INR contributed additional clinical context by capturing glycemic exposure, ischemic injury, coagulation activity, and anticoagulation status, all of which have recognized relevance to AF risk and management [ 8 , 55 – 61 ]. Overall, we developed an interpretable ML-based model to estimate AF risk among patients with DKD using routinely available clinical, laboratory, and echocardiographic data. In our dataset, the model showed stable discrimination and calibration, and we implemented a web calculator to facilitate use at the point of care. However, this study also has the following limitations. First, the retrospective design introduces risks of selection and information bias, and reliance on existing records may lead to missing or misclassified data. Second, although the dataset was multicenter, all participants were from China; external validity in other populations remains to be established. Third, the model included coagulation indices such as FIB and INR, and we did not exclude patients receiving anticoagulants; although these features contributed modestly, we cannot determine whether anticoagulation affected calibration. Fourth, the sample size was moderate, which may limit statistical power and the precision of subgroup effects. Future work should expand the cohort, include additional centers, and undertake external and prospective validation. 5 Conclusions This study developed and internally validated an interpretable machine-learning model for predicting atrial fibrillation among patients with diabetic kidney disease. By integrating routine clinical, laboratory, and echocardiographic variables, the model achieved stable discrimination and calibration within the study cohort. The use of SHAP improved interpretability, and a web-based calculator was implemented to facilitate clinical application. These findings suggest that machine learning may assist in early AF risk stratification in DKD and support individualized management. Future prospective, multicenter validation studies are warranted to confirm generalizability and assess clinical impact. Abbreviations DKD: Diabetic kidney disease; AF: Atrial fibrillation; ML: Machine learning; eGFR: Estimated glomerular filtration rate; SCr: Serum creatinine; BUN: Blood urea nitrogen; 24UTP: 24-hour urine total protein; HbA1c: Glycated hemoglobin; CK-MB: Creatine kinase-MB; Myo: Myoglobin; WBC: White blood cell; RBC: Red blood cell; INR: International normalized ratio; APTT: Activated partial thromboplastin time; FIB: Fibrinogen; LAAPD: Left atrial anterior-posterior diameter; LAMLD: Left atrial medial-lateral diameter; LASID: Left atrial superior-inferior diameter; RAMLD: Right atrial medial-lateral diameter; RASID: Right atrial superior-inferior diameter; EF: Ejection fraction; FS: Fractional shortening; KNN: K-nearest neighbors; RF: Random forest; XGBoost: Extreme gradient boosting; LightGBM: Light gradient boosting machine; GBDT: Gradient boosting decision tree; LDA: Linear discriminant analysis; ANN: Artificial neural network; LASSO: Least absolute shrinkage and selection operator; SHAP: Shapley additive explanations; DCA: Decision-curve analysis; AUC: Area under the curve; ICD-10: International Classification of Diseases, 10th Revision; RAAS: Renin-angiotensin-aldosterone system; ROC: Receiver operating characteristic. Declarations Competing interests Xiaoran Li, Xueying Wang, Xuebing Zhang, and Shidong Wang declare that they have no conflicts of interest. The funding bodies had no role in the design, conduct, or interpretation of the study. Funding This research was supported by Funding for Clinical and Scientific Research Business Expenses of High-level Central Traditional Chinese Medicine Hospitals [DZMG-QNZX-24018]. Authors' contributions Xiaoran Li and Xueying Wang and Xuebing Zhang and Shidong Wang devised the research plan. Xiaoran Li and Xueying Wang performed rigorous data analysis using advanced statistical methods, enhancing the reliability of the findings. The manuscript was skillfully composed by Xiaoran Li and Xueying Wang and Xuebing Zhang and Shidong Wang diligently revised the manuscript to improve its clarity and coherence. All authors have thoroughly reviewed and given their approval to the final version of this scholarly work. Consent for publication Not applicable Acknowledgments Not applicable. Availability of data and materials The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request. Ethics approval and consent to participate This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of Dongzhimen Hospital (Approval Nos. 2025DZM-376). Since this study is a retrospective study, the need for consent to participate was waived by an Institutional Review Board. References USRDS: the United States Renal Data System. Am J Kidney Dis. 2003;42(6 Suppl 5):1–230. Afkarian M, Sachs MC, Kestenbaum B, Hirsch IB, Tuttle KR, Himmelfarb J, de Boer IH. Kidney disease and increased mortality risk in type 2 diabetes. J Am Soc Nephrol. 2013;24(2):302–8. Christiansen CF, Schmidt M, Lamberg AL, Horvath-Puho E, Baron JA, Jespersen B, Sorensen HT. Kidney disease and risk of venous thromboembolism: a nationwide population-based case-control study. J Thromb Haemost. 2014;12(9):1449–54. Wang CY, Chen JF, Tu ST, Lee CC, Ou HY. Microvascular disease burden and macrovascular outcomes in type 2 diabetes: Risk calculator in Taiwan. Prim Care Diabetes 2025. Kaze AD, Jaar BG, Fonarow GC, Echouffo-Tcheugui JB. Diabetic kidney disease and risk of incident stroke among adults with type 2 diabetes. BMC Med. 2022;20(1):127. Corwin TR, Ozieh MN, Garacci E, Palatnik A, Egede LE. The relationship between financial hardship and incident diabetic kidney disease in older US adults - a longitudinal study. BMC Nephrol. 2021;22(1):167. Lippi G, Sanchis-Gomar F, Cervellin G. Global epidemiology of atrial fibrillation: An increasing epidemic and public health challenge. Int J Stroke. 2021;16(2):217–21. Du X, Ninomiya T, de Galan B, Abadir E, Chalmers J, Pillai A, Woodward M, Cooper M, Harrap S, Hamet P, et al. Risks of cardiovascular events and effects of routine blood pressure lowering among patients with type 2 diabetes and atrial fibrillation: results of the ADVANCE study. Eur Heart J. 2009;30(9):1128–35. Seyed Ahmadi S, Svensson AM, Pivodic A, Rosengren A, Lind M. Risk of atrial fibrillation in persons with type 2 diabetes and the excess risk in relation to glycaemic control and renal function: a Swedish cohort study. Cardiovasc Diabetol. 2020;19(1):9. Kaze AD, Yuyun MF, Fonarow GC, Echouffo-Tcheugui JB. Burden of Microvascular Disease and Risk of Atrial Fibrillation in Adults with Type 2 Diabetes. Am J Med. 2022;135(9):1093–100. e1092. Gadde S, Kalluru R, Cherukuri SP, Chikatimalla R, Dasaradhan T, Koneti J. Atrial Fibrillation in Chronic Kidney Disease: An Overview. Cureus. 2022;14(8):e27753. Carrero JJ, Trevisan M, Sood MM, Barany P, Xu H, Evans M, Friberg L, Szummer K. Incident Atrial Fibrillation and the Risk of Stroke in Adults with Chronic Kidney Disease: The Stockholm CREAtinine Measurements (SCREAM) Project. Clin J Am Soc Nephrol. 2018;13(9):1314–20. Alonso A, Lopez FL, Matsushita K, Loehr LR, Agarwal SK, Chen LY, Soliman EZ, Astor BC, Coresh J. Chronic kidney disease is associated with the incidence of atrial fibrillation: the Atherosclerosis Risk in Communities (ARIC) study. Circulation. 2011;123(25):2946–53. Watanabe H, Watanabe T, Sasaki S, Nagai K, Roden DM, Aizawa Y. Close bidirectional relationship between chronic kidney disease and atrial fibrillation: the Niigata preventive medicine study. Am Heart J. 2009;158(4):629–36. Kwon S, Lee SR, Choi EK, Ahn HJ, Lee SW, Jung JH, Han KD, Oh S, Lip GYH. Association Between Atrial Fibrillation and Diabetes-Related Complications: A Nationwide Cohort Study. Diabetes Care. 2023;46(12):2240–8. Richard Espiga F, Almendro Delia M, Caballero Martinez F, Monge Martin D, Neria Serrano F, Quiros Lopez R. Delayed diagnosis and missed opportunities in the early detection of atrial fibrillation: a cross-sectional study. Rev Clin Esp (Barc). 2024;224(9):560–8. Kaufman ES, Waldo AL. The impact of asymptomatic atrial fibrillation. J Am Coll Cardiol. 2004;43(1):53–4. Zhang P, Lin F, Ma F, Chen Y, Liu Y, Feng X, Fang S, Zhang H, Xiao S, Yang X, et al. Clinician-artificial intelligence collaboration: A win-win solution for efficiency and reliability in atrial fibrillation diagnosis. Med. 2025;6(7):100668. Hygrell T, Viberg F, Dahlberg E, Charlton PH, Kemp Gudmundsdottir K, Mant J, Hornlund JL, Svennberg E. An artificial intelligence-based model for prediction of atrial fibrillation from single-lead sinus rhythm electrocardiograms facilitating screening. Europace. 2023;25(4):1332–8. Nephrology, EGoCSo. Chinese guidelines for diagnosis and treatment of diabetic kidney disease. Chin J Nephrol. 2021;37:255–304. Electrophysiology and Pacing Branch of the Chinese Medical Association CRPCotCMDA, Huang CX. Chinese Guidelines for the management of atrial fibrillation(2025). Chin J Cardiac Pacing Electrophysiol. 2025;39(4):273–343. Jeng FC, Jeng YS. Implementation of Machine Learning on Human Frequency-Following Responses: A Tutorial. Semin Hear. 2022;43(3):251–74. Silva CAO, Morillo CA, Leite-Castro C, Gonzalez-Otero R, Bessani M, Gonzalez R, Castellanos JC, Otero L. Machine learning for atrial fibrillation risk prediction in patients with sleep apnea and coronary artery disease. Front Cardiovasc Med. 2022;9:1050409. Bashar SK, Ding E, Albuquerque D, Winter M, Binici S, Walkey AJ, McManus DD, Chon KH. Atrial Fibrillation Detection in ICU Patients: A Pilot Study on MIMIC III Data(). Annu Int Conf IEEE Eng Med Biol Soc. 2019;2019:298–301. Guan C, Gong A, Zhao Y, Yin C, Geng L, Liu L, Yang X, Lu J, Xiao B. Interpretable machine learning model for new-onset atrial fibrillation prediction in critically ill patients: a multi-center study. Crit Care. 2024;28(1):349. Papadopoulou A, Harding D, Slabaugh G, Marouli E, Deloukas P. Prediction of atrial fibrillation and stroke using machine learning models in UK Biobank. Heliyon. 2024;10(7):e28034. Laukkanen JA, Zaccardi F, Karppi J, Ronkainen K, Kurl S. Reduced kidney function is a risk factor for atrial fibrillation. Nephrol (Carlton). 2016;21(8):717–20. Lan Q, Zheng L, Zhou X, Wu H, Buys N, Liu Z, Sun J, Fan H. The Value of Blood Urea Nitrogen in the Prediction of Risks of Cardiovascular Disease in an Older Population. Front Cardiovasc Med. 2021;8:614117. Ding WY, Gupta D, Wong CF, Lip GYH. Pathophysiology of atrial fibrillation and chronic kidney disease. Cardiovasc Res. 2021;117(4):1046–59. Schnabel RB, Yin X, Gona P, Larson MG, Beiser AS, McManus DD, Newton-Cheh C, Lubitz SA, Magnani JW, Ellinor PT, et al. 50 year trends in atrial fibrillation prevalence, incidence, risk factors, and mortality in the Framingham Heart Study: a cohort study. Lancet. 2015;386(9989):154–62. Kistler PM, Sanders P, Fynn SP, Stevenson IH, Spence SJ, Vohra JK, Sparks PB, Kalman JM. Electrophysiologic and electroanatomic changes in the human atrium associated with age. J Am Coll Cardiol. 2004;44(1):109–16. Morillo CA. Age and Atrial Fibrillation Outcomes: Myth or Muzak, Scandinavian Lessons. JACC Clin Electrophysiol. 2025;11(1):95–7. Gao P, Gao X, Xie B, Tse G, Liu T. Aging and atrial fibrillation: A vicious circle. Int J Cardiol. 2024;395:131445. Wong CX, Sullivan T, Sun MT, Mahajan R, Pathak RK, Middeldorp M, Twomey D, Ganesan AN, Rangnekar G, Roberts-Thomson KC, et al. Obesity and the Risk of Incident, Post-Operative, and Post-Ablation Atrial Fibrillation: A Meta-Analysis of 626,603 Individuals in 51 Studies. JACC Clin Electrophysiol. 2015;1(3):139–52. Jones NR, Taylor KS, Taylor CJ, Aveyard P. Weight change and the risk of incident atrial fibrillation: a systematic review and meta-analysis. Heart. 2019;105(23):1799–805. Vyas V, Lambiase P. Obesity and Atrial Fibrillation: Epidemiology, Pathophysiology and Novel Therapeutic Opportunities. Arrhythm Electrophysiol Rev. 2019;8(1):28–36. Shu H, Cheng J, Li N, Zhang Z, Nie J, Peng Y, Wang Y, Wang DW, Zhou N. Obesity and atrial fibrillation: a narrative review from arrhythmogenic mechanisms to clinical significance. Cardiovasc Diabetol. 2023;22(1):192. Jiang W, Wang J, Shen X, Lu W, Wang Y, Li W, Gao Z, Xu J, Li X, Liu R, et al. Establishment and Validation of a Risk Prediction Model for Early Diabetic Kidney Disease Based on a Systematic Review and Meta-Analysis of 20 Cohorts. Diabetes Care. 2020;43(4):925–33. Man REK, Gan ATL, Fenwick EK, Gupta P, Wong MYZ, Wong TY, Tan GSW, Teo BW, Sabanayagam C, Lamoureux EL. The Relationship between Generalized and Abdominal Obesity with Diabetic Kidney Disease in Type 2 Diabetes: A Multiethnic Asian Study and Meta-Analysis. Nutrients 2018, 10(11). Keane WF, Brenner BM, de Zeeuw D, Grunfeld JP, McGill J, Mitch WE, Ribeiro AB, Shahinfar S, Simpson RL, Snapinn SM, et al. The risk of developing end-stage renal disease in patients with type 2 diabetes and nephropathy: the RENAAL study. Kidney Int. 2003;63(4):1499–507. Gong M, Xu M, Pan S, Jiang S, Jiang X. Association of Left Atrium Remodeling With Major Adverse Cardiovascular Events in Asymptomatic Type 2 Diabetes Patients With Early Chronic Kidney Disease. Rev Cardiovasc Med. 2025;26(5):27247. Kadappu KK, Abhayaratna K, Boyd A, French JK, Xuan W, Abhayaratna W, Thomas L. Independent Echocardiographic Markers of Cardiovascular Involvement in Chronic Kidney Disease: The Value of Left Atrial Function and Volume. J Am Soc Echocardiogr. 2016;29(4):359–67. Gong M, Xu M, Pan S, Jiang X. Association between left atrial remodeling and early renal impairment in asymptomatic patients with type 2 diabetes. BMC Cardiovasc Disord. 2025;25(1):436. Li X, Dong Y, Zheng C, Wang P, Xu M, Zou C, Wang L. Assessment of real-time three-dimensional echocardiography as a tool for evaluating left atrial volume and function in patients with type 2 diabetes mellitus. Aging. 2020;13(1):991–1000. Tripepi G, Benedetto FA, Mallamaci F, Tripepi R, Malatino L, Zoccali C. Left atrial volume in end-stage renal disease: a prospective cohort study. J Hypertens. 2006;24(6):1173–80. Tripepi G, Benedetto FA, Mallamaci F, Tripepi R, Malatino L, Zoccali C. Left atrial volume monitoring and cardiovascular risk in patients with end-stage renal disease: a prospective cohort study. J Am Soc Nephrol. 2007;18(4):1316–22. Paoletti E, Zoccali C. A look at the upper heart chamber: the left atrium in chronic kidney disease. Nephrol Dial Transpl. 2014;29(10):1847–53. Hakamaki M, Hellman T, Lankinen R, Koivuviita N, Parkka J, Kallio P, Kiviniemi T, Airaksinen KEJ, Jarvisalo MJ, Metsarinne K. Elevated Troponin T and Enlarged Left Atrium Are Associated with the Incidence of Atrial Fibrillation in Patients with CKD Stage 4–5. Nephron. 2021;145(1):71–7. Zhou X, Ruan W, Zhao L, Lin K, Li J, Liu H, Wang T, Zhang G. Causal Links Between Renal Function and Cardiac Structure, Function, and Disease Risk. Glob Heart. 2024;19(1):83. Santhanakrishnan R, Wang N, Larson MG, Magnani JW, McManus DD, Lubitz SA, Ellinor PT, Cheng S, Vasan RS, Lee DS, et al. Atrial Fibrillation Begets Heart Failure and Vice Versa: Temporal Associations and Differences in Preserved Versus Reduced Ejection Fraction. Circulation. 2016;133(5):484–92. Bell DSH, McGill JB, Jerkins T. Management of the 'wicked' combination of heart failure and chronic kidney disease in the patient with diabetes. Diabetes Obes Metab. 2023;25(10):2795–804. Bell DS. Heart failure: the frequent, forgotten, and often fatal complication of diabetes. Diabetes Care. 2003;26(8):2433–41. Vaziri SM, Larson MG, Benjamin EJ, Levy D. Echocardiographic predictors of nonrheumatic atrial fibrillation. Framingham Heart Study Circulation. 1994;89(2):724–30. Leclercq ADES, Halimi JF, Fiorello F, Bertrand P, Attuel C. Evaluation of time course and predicting factors of progression of paroxysmal or persistent atrial fibrillation to permanent atrial fibrillation. Pacing Clin Electrophysiol. 2014;37(3):345–55. Qi W, Zhang N, Korantzopoulos P, Letsas KP, Cheng M, Di F, Tse G, Liu T, Li G. Serum glycated hemoglobin level as a predictor of atrial fibrillation: A systematic review with meta-analysis and meta-regression. PLoS ONE. 2017;12(3):e0170955. Berton G, Cordiano R, Cucchini F, Cavuto F, Pellegrinet M, Palatini P. Atrial fibrillation during acute myocardial infarction: association with all-cause mortality and sudden death after 7-year of follow-up. Int J Clin Pract. 2009;63(5):712–21. Mukamal KJ, Tolstrup JS, Friberg J, Gronbaek M, Jensen G. Fibrinogen and albumin levels and risk of atrial fibrillation in men and women (the Copenhagen City Heart Study). Am J Cardiol. 2006;98(1):75–81. Pisters R, Lane DA, Nieuwlaat R, de Vos CB, Crijns HJ, Lip GY. A novel user-friendly score (HAS-BLED) to assess 1-year risk of major bleeding in patients with atrial fibrillation: the Euro Heart Survey. Chest. 2010;138(5):1093–100. Li-Saw-Hee FL, Blann AD, Gurney D, Lip GY. Plasma von Willebrand factor, fibrinogen and soluble P-selectin levels in paroxysmal, persistent and permanent atrial fibrillation. Effects of cardioversion and return of left atrial function. Eur Heart J. 2001;22(18):1741–7. Li Z, Pang M, Li Y, Yu Y, Peng T, Hu Z, Niu R, Li J, Wang X. Development and validation of a predictive model for new-onset atrial fibrillation in sepsis based on clinical risk factors. Front Cardiovasc Med. 2022;9:968615. Kannel WB, Wolf PA, Benjamin EJ, Levy D. Prevalence, incidence, prognosis, and predisposing conditions for atrial fibrillation: population-based estimates. Am J Cardiol. 1998;82(8A):N2–9. Tables Table 1 to 3 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files FigS1.pdf Figure S1. ROC curve and average AUC under five-fold cross-validation. Displays the ROC curve from five-fold cross-validation, with lighter colors representing individual folds and the dark blue line representing the average ROC. The average AUC is approximately 0.90 ± 0.02, indicating stable discriminatory performance of the model on training data. Table1.docx Table2.docx Table3.docx Cite Share Download PDF Status: Posted Version 1 posted 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-8146262","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":554407514,"identity":"b8bf9a44-33ba-4efc-a983-cc761c9ac27b","order_by":0,"name":"Xiaoran Li","email":"","orcid":"","institution":"University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Xiaoran","middleName":"","lastName":"Li","suffix":""},{"id":554407515,"identity":"e44914c1-0fff-41ea-baf2-c746af548731","order_by":1,"name":"Xueying Wang","email":"","orcid":"","institution":"Beijing Electric Power Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xueying","middleName":"","lastName":"Wang","suffix":""},{"id":554407516,"identity":"49b4f99e-5968-4ddc-bd36-5a5ebe994e4c","order_by":2,"name":"Shidong Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAApUlEQVRIiWNgGAWjYHACw8dQhgHRWoyNSdZiJk2aFoPzh7dVF7bZRTOwN2+TYKi5Q4SWG2llt2e2Jec28Bwrk2A49owIV93gMbvN23Ygt0Eix0yCseEwEVrOnzErBmuRf0OslgM5ZswQW3iI1GJ/I61Yesa55Nw2nrRii4RjRGiR7D+88XNBmV1uP/vhjTc+1BChBQ7YQEQCCRpGwSgYBaNgFOABADWXNxbfvzWEAAAAAElFTkSuQmCC","orcid":"","institution":"University of Chinese Medicine","correspondingAuthor":true,"prefix":"","firstName":"Shidong","middleName":"","lastName":"Wang","suffix":""},{"id":554407517,"identity":"55586f6c-8486-4f75-8e37-f33f267cb9e6","order_by":3,"name":"Xuebing Zhang","email":"","orcid":"","institution":"University of Chinese Medicine","correspondingAuthor":false,"prefix":"","firstName":"Xuebing","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-11-18 13:53:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8146262/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8146262/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97395880,"identity":"96aa7800-1eb9-4798-9d98-be9a9f75d3f5","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":178671,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/a672223916f6885ba7b0caed.docx"},{"id":97395877,"identity":"ceb744af-8b98-4689-b766-4273e7a4322f","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"json","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6511,"visible":true,"origin":"","legend":"","description":"","filename":"196a9641d3ac4680ae17a7ade3f64b13.json","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/0ec914fa2632953fde3a0072.json"},{"id":97666022,"identity":"e1a30c4d-3d6a-4981-b9c0-1279c70b19b8","added_by":"auto","created_at":"2025-12-08 09:20:20","extension":"pdf","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15159,"visible":true,"origin":"","legend":"","description":"","filename":"FigS1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/2c7123c0e198745c04bb3c63.pdf"},{"id":97395882,"identity":"67878805-43da-459d-8a90-99a20f570960","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"xml","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":192998,"visible":true,"origin":"","legend":"","description":"","filename":"196a9641d3ac4680ae17a7ade3f64b131enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/1de1e0029bb53af680fe8f5b.xml"},{"id":97665374,"identity":"e3dc6fc0-3e12-4715-9cab-95d7cb20e042","added_by":"auto","created_at":"2025-12-08 09:18:05","extension":"pdf","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":207360,"visible":true,"origin":"","legend":"","description":"","filename":"Fig1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/0325843b6a315f5c0fd82fd7.pdf"},{"id":97395889,"identity":"daf869ba-4159-4170-8a57-696336563419","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"pdf","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":936771,"visible":true,"origin":"","legend":"","description":"","filename":"Fig2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/5e18b5261c7512f29b01e467.pdf"},{"id":97395890,"identity":"9081204b-b333-41bc-b5fe-cfe688ed5174","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"pdf","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3034344,"visible":true,"origin":"","legend":"","description":"","filename":"Fig3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/5b1ecdaadfbe3d6d3f615954.pdf"},{"id":97395888,"identity":"328802dc-7e1e-4899-86c3-bb8a079c2345","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"pdf","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1663664,"visible":true,"origin":"","legend":"","description":"","filename":"Fig4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/ae7d427546d2e5904828c288.pdf"},{"id":97395886,"identity":"972ba318-a041-4014-a4c7-c516359b75ca","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"pdf","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":791625,"visible":true,"origin":"","legend":"","description":"","filename":"Fig5.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/d2a3b8f3a5a10ce381159934.pdf"},{"id":97665569,"identity":"7d73c293-b466-4e2e-9ea2-f69c5afc135c","added_by":"auto","created_at":"2025-12-08 09:19:04","extension":"pdf","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":700639,"visible":true,"origin":"","legend":"","description":"","filename":"Fig6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/1f625873379399353906718c.pdf"},{"id":97666386,"identity":"e78d9271-61b8-4f8b-a884-b2370c08e2e0","added_by":"auto","created_at":"2025-12-08 09:21:05","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":191085,"visible":true,"origin":"","legend":"","description":"","filename":"196a9641d3ac4680ae17a7ade3f64b131structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/3c364e04ffd1ea6b3098b78d.xml"},{"id":97395885,"identity":"ad83c136-a7e8-4fc4-a715-fc80706eef8b","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":201791,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/edabce2116162c277e2aefb3.html"},{"id":97395868,"identity":"32fb0832-c137-4391-a845-604c4eb1a6b1","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":39341,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy workflow and model development validation and framework.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe figure illustrates the overall process of case selection, data preprocessing, feature selection, model construction and evaluation, and result interpretation. A total of 787 cases and 26 variables were included and randomly divided into a training set (n=550) and a test set (n=237) in a 7:3 ratio.\u003c/p\u003e","description":"","filename":"Fig181.png","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/90a39ff8838127b77ae3d675.png"},{"id":97665326,"identity":"551cc952-3c35-48bc-bd24-f6b072b2ba8c","added_by":"auto","created_at":"2025-12-08 09:17:51","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":164485,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eLASSO feature selection.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) The curve showing the change in binomial deviance with log(λ) under 10-fold cross-validation, with vertical dashed lines indicating λ_min and λ_1se, and the top axis showing the number of non-zero coefficients at different penalty strengths. (B) Coefficient path plot: The regression coefficients of each candidate variable gradually shrink towards zero as log(λ) increases, with the final non-zero coefficients forming the selected feature set for the model.\u003c/p\u003e","description":"","filename":"Fig182.png","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/4e15de73b80231c30f7b0a35.png"},{"id":97666365,"identity":"a9c84034-a75f-46c9-ae46-dd61866ec208","added_by":"auto","created_at":"2025-12-08 09:21:04","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":696351,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparison of model performance across multiple algorithms in training/test sets.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) ROC curve comparison in the training set (7 algorithms, with dashed line representing the chance line). (B) Calibration curve in the training set (the diagonal dashed line represents ideal calibration, with Brier score shown). (C) Decision-curve analysis (DCA) in the training set: Net benefit at various risk thresholds, with \"treat-all/treat-none\" as the reference. (D) ROC curve comparison in the test set. (E) Calibration curve in the test set (Brier score annotated). (F) DCA in the test set.\u003c/p\u003e","description":"","filename":"Fig183.png","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/f699fca06936f7e65493b638.png"},{"id":97666238,"identity":"5c5305ff-e45c-4ef8-b878-59f4592801ec","added_by":"auto","created_at":"2025-12-08 09:20:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":241924,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP interpretability analysis of the optimal KNN model.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) Global feature importance, showing the major contributions of 24UTP, SCr, and age-related indicators to prediction. (B) SHAP summary plot: Horizontal position represents the magnitude of positive/negative influence of each feature on individual predictions, with color indicating high/low feature values. (C-D) Decision/Force plot: Exemplifies the cumulative contribution paths of features in low-risk and high-risk individuals, where red increases and blue decreases the predicted risk, with the endpoint representing the final predicted probability for each individual. (E-F) Waterfall plot: Shows the stepwise contribution or reduction of each feature to the prediction probability for an individual, with the final value being the model output.\u003c/p\u003e","description":"","filename":"Fig184.png","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/3c29260453b93e930f4cffb4.png"},{"id":97665446,"identity":"f0f456fd-d7db-44fd-86a2-1f88b2166595","added_by":"auto","created_at":"2025-12-08 09:18:29","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":183374,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eClinical impact curve analysis (CIC) of the optimal KNN model.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eClinical impact curves for the training set (A) and test set (B) at different risk thresholds: The blue line represents the number of individuals classified as high risk, the red line represents the true positives among them, and the green dashed line represents the observed number of events.\u003c/p\u003e","description":"","filename":"Fig185.png","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/367b1bb65a3859f4073fd6aa.png"},{"id":97666030,"identity":"ed48943a-63ba-4eac-890a-b0912819417d","added_by":"auto","created_at":"2025-12-08 09:20:20","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":741677,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eUser interface of the web-based risk calculator.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe figure shows the online prediction platform developed from the final KNN model using the Shiny framework. Users can input 14 key variables to obtain personalized predictions for whether DKD have atrial fibrillation. The calculator is publicly accessible and designed to support bedside risk assessment and clinical decision-making.\u003c/p\u003e","description":"","filename":"Fig186.png","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/096d1f91f667ae645589f5b8.png"},{"id":98622881,"identity":"a7e880e6-a0ba-43c6-b9c3-6de29d66cb42","added_by":"auto","created_at":"2025-12-19 17:03:18","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2806059,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/45fbae53-82c4-40df-a5d0-98eadfccda2d.pdf"},{"id":97395873,"identity":"08a04401-4f1c-4584-9698-4208f37d69a6","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":15159,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure S1. ROC curve and average AUC under five-fold cross-validation.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDisplays the ROC curve from five-fold cross-validation, with lighter colors representing individual folds and the dark blue line representing the average ROC. The average AUC is approximately 0.90 ± 0.02, indicating stable discriminatory performance of the model on training data.\u003c/p\u003e","description":"","filename":"FigS1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/c84ec73cc0b186ed5cf42d55.pdf"},{"id":97395870,"identity":"f2eb7685-74ad-45fb-b6c4-13c4ebaed3ac","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":27782,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/f6b068499aed83df1be4854d.docx"},{"id":97395871,"identity":"6444e558-5942-46fd-8ccb-0fff65d2adf3","added_by":"auto","created_at":"2025-12-04 00:13:57","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":28162,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/ade13f6f705533ab0d561077.docx"},{"id":97665476,"identity":"d1192836-047f-4e75-ab44-821136e36811","added_by":"auto","created_at":"2025-12-08 09:18:36","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":17173,"visible":true,"origin":"","legend":"","description":"","filename":"Table3.docx","url":"https://assets-eu.researchsquare.com/files/rs-8146262/v1/48168c17ca08d450d862da42.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and Internal Validation of an Interpretable Machine Learning Model for Predicting Atrial Fibrillation in Patients with Diabetic Kidney Disease: A Multicenter Study","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eDiabetic kidney disease (DKD), one of the most common microvascular complications of diabetes, is a leading cause of end-stage renal disease and renal replacement therapy [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. DKD not only markedly increases the risk of kidney failure but is also closely associated with thrombosis, stroke, myocardial infarction, and all-cause mortality [\u003cspan additionalcitationids=\"CR3 CR4\" citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. The rising prevalence of DKD poses a major challenge to global health systems [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Atrial fibrillation (AF), one of the most frequent cardiac arrhythmias, has increased in prevalence by 33% over the past two decades [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Patients with AF have a 61% higher risk of all-cause death and face greater risks of cardiovascular mortality and stroke [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Studies have shown that individuals with type 2 diabetes have a 35% higher risk of developing AF than the general population, and the risk further increases among patients with DKD as estimated glomerular filtration rate (eGFR) declines and urinary protein excretion rises [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. DKD and AF share several risk factors, including hypertension and coronary artery disease. Both conditions are influenced by thrombosis, systemic inflammation, and activation of the renin-angiotensin system (RAS), forming a bidirectional relationship [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. On one hand, eGFR decline and proteinuria in DKD may promote AF onset [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]; on the other hand, diabetic patients with AF are more likely to develop DKD, and AF can accelerate renal function deterioration and proteinuria progression in DKD [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. This mutual reinforcement creates a vicious cycle leading to poor clinical outcomes.\u003c/p\u003e\u003cp\u003eIn clinical practice, AF is often paroxysmal or asymptomatic in its early stage. Silent AF is frequently underdiagnosed or diagnosed late [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], especially in patients with DKD, who may lack overt arrhythmic symptoms. As a result, opportunities for early rhythm control and intervention are easily missed, which may accelerate the progression of both DKD and AF and increase the risk of serious cardiovascular and cerebrovascular events. These challenges highlight the need for effective risk assessment and early identification of AF in patients with DKD to improve clinical outcomes. Despite the growing clinical burden, no dedicated AF screening or prediction model has been developed for the DKD population. Existing tools show limited accuracy in identifying AF among these patients. Previous studies have mainly focused on the general population or isolated AF prediction. For example, some investigators combined clinical expertise with artificial intelligence (AI) analysis of 24-hour Holter recordings to improve AF detection rates [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. However, such approaches are difficult to implement widely because DKD patients often lack routine Holter monitoring. Other models based on AI analysis of sinus-rhythm electrocardiograms can predict paroxysmal AF [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], but their performance remains limited without incorporating clinical and laboratory data.\u003c/p\u003e\u003cp\u003eGiven the increasing prevalence and clinical significance of AF in DKD and the diagnostic limitations of current methods, a reliable and interpretable risk prediction model is urgently needed. In this study, we developed and internally validated a machine learning (ML) based model integrating clinical, laboratory, and echocardiographic variables to predict AF in hospitalized patients with DKD. Our goal was to provide a practical and interpretable tool for early, accurate identification of high-risk patients and to support clinical decision making and management.\u003c/p\u003e"},{"header":"2 Material and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e\u003cb\u003e2.1 Study design and setting\u003c/b\u003e\u003c/h2\u003e\u003cp\u003eThis retrospective cohort study included adult inpatients with DKD who were admitted to Dongzhimen Hospital, Beijing University of Chinese Medicine, and Beijing Electric Power Hospital between January 2021 and December 2024. A total of 814 admissions were screened, and 787 unique records met eligibility criteria and were included in the analysis. The dataset was randomly divided into a training set (70%) and an internal test set (30%) using stratified sampling to preserve the proportion of outcomes. Data were extracted by authorized personnel from identifiable electronic medical records to ensure accurate linkage, and all records were de-identified before analysis.\u003c/p\u003e\u003cp\u003eEthical approval\u003c/strong\u003e was obtained from the institutional review boards of both hospitals (Dongzhimen Hospital and Beijing Electric Power Hospital; Approval Nos. 2025DZM-376). Since this study is a retrospective study, the need for consent to participate was waived by an Institutional Review Board. All procedures followed the principles of the Declaration of Helsinki and complied with applicable national regulations. An overview of the study workflow is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Endpoint definition\u003c/h2\u003e\u003cp\u003eThe primary endpoint was the presence of AF at the index admission. AF was identified using a combination of diagnostic sources, including a standard 12-lead electrocardiogram, 24-hour Holter monitoring when available, and ICD-10 diagnostic codes extracted from the electronic medical record. Each case was independently reviewed and confirmed by two physicians to ensure diagnostic accuracy.\u003c/p\u003e\u003cp\u003eThe prediction time point was defined as the earliest availability of all candidate predictors within 72 hours after hospital admission.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Participant selection\u003c/h2\u003e\u003cp\u003eParticipants were identified according to the Chinese Clinical Guidelines for DKD[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] and the Chinese Guidelines for the Management of AF (2025 edition)[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Eligible patients were required to meet all of the following inclusion criteria: (1) age between 18 and 95 years (inclusive); (2) any sex; (3) diagnosis of diabetes mellitus consistent with Chinese and World Health Organization criteria; and (4) diagnosis of diabetic kidney disease confirmed by clinical and laboratory findings.\u003c/p\u003e\u003cp\u003eExclusion criteria included (1) type 1 diabetes, gestational diabetes, or other specific forms of diabetes; (2) coexisting primary renal diseases such as primary glomerulonephritis, lupus nephritis, or Henoch Sch\u0026ouml;nlein nephritis; (3) ongoing glucocorticoid or immunosuppressive therapy; (4) concomitant malignant tumors, decompensated liver cirrhosis, active tuberculosis, infectious shock, disseminated intravascular coagulation, or other critical or terminal illnesses; and (5) major limb amputation within the preceding six months.\u003c/p\u003e\u003cp\u003eAdditional cardiac related exclusions were (1) pacemaker or implantable cardioverter defibrillator implantation; (2) other clinically significant tachyarrhythmias such as atrial flutter or sustained supraventricular tachycardia; and (3) a history of cardiac surgery including valve replacement or coronary artery bypass grafting. A previous diagnosis of atrial fibrillation was not an exclusion criterion because the endpoint focused on AF status at baseline rather than new onset cases.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Variable collection and measurements\u003c/h2\u003e\u003cp\u003eInformation extracted from the electronic medical record included three domains: (1) demographics and clinical data, such as sex, age, height, weight, blood pressure, and diabetes duration; (2) laboratory tests, including complete blood count, urinalysis, biochemical profile, arterial blood gas, coagulation parameters, thyroid function, tumor markers, cardiac enzymes, and heart failure biomarkers; and (3) transthoracic echocardiography findings.\u003c/p\u003e\u003cp\u003eThe final predictor set used for model construction consisted of routinely available variables: HbA1c (glycated hemoglobin, %), BUN (blood urea nitrogen, mmol/L), SCr (serum creatinine, \u0026micro;mol/L), BNP (B-type natriuretic peptide, pg/mL), ALT (U/L), AST (U/L), CK-MB (U/L), Myo (myoglobin, ng/mL), WBC (\u0026times;10⁹/L), RBC (\u0026times;10\u0026sup1;\u0026sup2;/L), 24UTP (24-hour urine total protein, g/24 h), INR, APTT (s), FIB (g/L), LAAPD (left atrial anterior\u0026ndash;posterior diameter, mm), LAMLD (left atrial medial\u0026ndash;lateral diameter, mm), LASID (left atrial superior\u0026ndash;inferior diameter, mm), RAMLD (right atrial medial\u0026ndash;lateral diameter, mm), RASID (right atrial superior\u0026ndash;inferior diameter, mm), EF (%), and FS (%), plus sex.\u003c/p\u003e\u003cp\u003eEchocardiography was performed by certified sonographers following a standardized institutional protocol using the same equipment platform (device and model to be specified). LAAPD was measured in the parasternal long-axis view at end-systole. LAMLD, LASID, and their right-atrial counterparts RAMLD and RASID were obtained from orthogonal apical views in medial lateral and superior inferior planes. All diameters were recorded from inner edge to inner edge and averaged over three cardiac cycles. A random 10% subset was re-measured to evaluate intra- and inter-observer consistency, with a target variability below 5%.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5 Data preprocessing and handling of missing data\u003c/h2\u003e\u003cp\u003eVariables with more than 20% missingness were excluded a priori. For the remaining predictors, for which overall missingness was typically less than 10%, missing values were imputed using the MissForest algorithm with default depth and 100 trees. Imputation was implemented within the training set only to prevent information leakage, and the trained imputer was then applied to the held-out test set. Categorical harmonization followed a unified data dictionary, and rare categories with a prevalence below 2% were merged.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e2.6 Feature selection, model development, and hyper parameter tuning\u003c/h2\u003e\u003cp\u003eFeature selection was performed using the least absolute shrinkage and selection operator (LASSO) regression in the training dataset. A ten-fold cross-validation procedure was applied with the \u003cem\u003eglmnet\u003c/em\u003e package in R to identify a parsimonious subset of variables with the strongest predictive contribution. The optimal penalty parameter (λ) was chosen according to the minimum cross-validated deviance criterion, and the selected features were used for subsequent model construction.\u003c/p\u003e\u003cp\u003eSeven supervised machine learning algorithms were developed on the selected predictors: gradient boosting decision tree (GBDT), k-nearest neighbors (KNN), linear discriminant analysis (LDA), Light Gradient Boosting Machine (LightGBM), random forest (RF), artificial neural network (ANN), and extreme gradient boosting (XGBoost). Each algorithm was trained separately to capture different model structures and learning strategies.\u003c/p\u003e\u003cp\u003eHyper parameters were tuned using stratified five-fold cross-validation within the training dataset through either grid search or randomized search. All preprocessing, feature selection, and tuning steps were restricted to the training data to prevent data leakage. Because the prevalence of atrial fibrillation was approximately balanced, no oversampling, undersampling, or class weighting was applied. Decision thresholds for secondary analyses were predefined before model evaluation.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e2.7 Model evaluation, validation, interpretability, and software.\u003c/h2\u003e\u003cp\u003eModel performance was evaluated in both the training and test datasets. Discrimination was assessed using the area under the receiver-operating characteristic curve (AUC), with accuracy, sensitivity, specificity, precision, and F1-score as complementary metrics. Calibration was evaluated by the Brier score, calibration intercept and slope, and visual inspection of calibration plots generated with loess smoothing. Decision-curve analysis (DCA) was applied to estimate the net clinical benefit across prespecified probability thresholds ranging from 5% to 20%. Clinical-impact curves were used to illustrate the number of individuals classified as high risk and the corresponding number of true positives per 1,000 patients at representative thresholds. Model interpretability was examined for the k-nearest neighbors classifier using KernelSHAP, which quantified the contribution of each predictor to individual predictions. SHAP values were computed on standardized inputs with a background set of 100 training instances selected by k-medoids sampling to approximate the empirical data distribution.\u003c/p\u003e\u003cp\u003eDescriptive statistics were performed in SPSS 25.0. Penalized regression for feature selection was conducted in R 4.4.1. All machine-learning models and visualizations, including receiver-operating characteristic, calibration, decision, clinical-impact, and SHAP plots, were implemented in Python 3.10.4 using standard scientific libraries. The finalized prediction model was further deployed as a secure web-based calculator. This tool allows clinicians to input routinely available variables and obtain individualized AF risk predictions in real time. The calculator operates within the institutional network under encrypted communication and restricted access, ensuring data confidentiality. It is intended to support, rather than replace, clinical decision making.\u003c/p\u003e\u003c/div\u003e"},{"header":"3 Results","content":"\u003cp\u003e3.1 Demographic and baseline characteristics\u003c/p\u003e\n\u003cp\u003eA total of 787 patients with diabetic kidney disease were included and randomly divided in a 7:3 ratio into the training set (n = 550) and the test set (n = 237). No significant differences were observed between the two sets in age, diabetes duration, HbA1c, SCr, BNP, EF, or sex distribution (male 41.8% vs. 41.5%, P = 0.934), confirming good comparability between sets (Table 1).\u003c/p\u003e\n\u003cp\u003eWithin the training set, patients with AF differed markedly from those without AF (Table 2). The AF group was older (P \u0026lt; 0.001), had a longer diabetes duration (P = 0.013), and higher body weight (P \u0026lt; 0.001). Levels of BUN, SCr, CK-MB, and Myo were significantly higher in the AF group (all P \u0026lt; 0.01), whereas BNP was notably lower (P \u0026lt; 0.001). Echocardiographic measurements showed smaller atrial diameters (LAAPD, LAMLD, LASID, RAMLD, and RASID) but higher EF and FS values (all P \u0026lt; 0.001). Additionally, women were more prevalent in the AF group (69.2% vs. 48.8%, P \u0026lt; 0.001).\u003c/p\u003e\n\u003cp\u003eCollectively, AF patients displayed distinct metabolic and cardiac profiles, suggesting links between atrial remodeling, renal dysfunction, and AF risk in DKD.\u003c/p\u003e\n\u003cp\u003e3.2 Feature selection\u003c/p\u003e\n\u003cp\u003eFeature selection was performed using the LASSO regression with tenfold cross-validation in the training set. This method applies L1 regularization to shrink less informative coefficients toward zero, thereby retaining only the most predictive variables. Model performance was monitored across iterations to identify the subset with the best cross-validated performance.\u003c/p\u003e\n\u003cp\u003eAs shown in Figure 2A, the mean cross-validation curve indicated stable convergence, and the corresponding coefficient profiles are presented in Figure 2B. Fourteen variables with non-zero coefficients were ultimately retained: 24UTP, SCr, age, LAMLD, weight, LASID, LAAPD, RAMLD, FS, BUN, HbA1c, CK-MB, FIB, and INR. These features exhibited the strongest contribution to AF prediction and were subsequently used for model development.\u003c/p\u003e\n\u003cp\u003e3.3 \u003cstrong\u003eModel performance comparison\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSeven supervised machine-learning algorithms were developed using the selected features: KNN, RF, XGBoost, LightGBM, GBDT, LDA, and ANN. In the training set, all models showed good discrimination (Figure 3A). The KNN model achieved the highest AUC (0.933) with an accuracy of 0.8745, sensitivity of 0.9077, specificity of 0.8448, and a Brier score of 0.100. RF, XGBoost, and LightGBM performed comparably, with AUCs of 0.9200, 0.9196, and 0.9144, respectively (Table 3).\u003c/p\u003e\n\u003cp\u003eIn the test set (Figure 3D), KNN maintained superior performance with an AUC of 0.9271, accuracy of 0.8861, sensitivity of 0.9196, and specificity of 0.8560. XGBoost achieved the lowest Brier score (0.094), while RF and LightGBM yielded AUCs of 0.8955 and 0.8791, respectively. Although LightGBM had a slightly higher Brier score (0.123), its calibration remained acceptable (Table 3).\u003c/p\u003e\n\u003cp\u003eCalibration curves showed good agreement between predicted and observed probabilities for all models (Figures 3B and 3E). Decision-curve analysis indicated that KNN and XGBoost provided the greatest net clinical benefit across most threshold ranges (Figures 3C and 3F). Five-fold cross-validation further confirmed the robustness of model performance, with an average AUC of 0.90 \u0026plusmn; 0.02 (Figure S1). Taken together, KNN demonstrated the most favorable balance of discrimination, calibration, and generalization, indicating a favorable balance of discrimination, calibration, and generalization.\u003c/p\u003e\n\u003cp\u003e3.4 \u003cstrong\u003eModel interpretation and SHAP analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo explore the interpretability of the KNN model and quantify the contribution of each feature, SHAP analysis was applied (Figure 4). The mean absolute SHAP value plot (Figure 4A) identified age, SCr, LAAPD, and 24UTP as the strongest predictors of AF. Among them, 24UTP showed the highest average SHAP value, indicating the greatest overall influence on model output.\u003c/p\u003e\n\u003cp\u003eThe SHAP summary plot (Figure 4B) illustrated the distribution of feature effects across all samples. Each point represents a single observation, with color indicating the feature value (red for high, blue for low) and position reflecting its positive or negative contribution to prediction. Features with low values exhibited stronger positive impacts on predicted AF risk, whereas higher values showed weaker effects.\u003c/p\u003e\n\u003cp\u003eTo visualize individual-level predictions, representative force and waterfall plots were generated (Figures 4C-F). These plots demonstrated how each feature shifted the prediction toward or away from AF, highlighting the cumulative contribution of multiple risk factors. Overall, the results suggested that increased proteinuria burden and impaired renal function were the primary drivers of AF risk in patients with DKD.\u003c/p\u003e\n\u003cp\u003e3.5 Clinical impact curve analysis\u003c/p\u003e\n\u003cp\u003eTo assess the clinical utility of the model, clinical impact curves (CICs) for the KNN classifier were generated in both the training and test sets (Figure 5A and 5B). As the decision threshold increased, the number of patients classified as high risk (blue line) gradually decreased. The number of true positives (red line) closely overlapped with the number of observed AF cases (green dashed line), indicating that the predicted probabilities were well aligned with actual outcomes.\u003c/p\u003e\n\u003cp\u003eAcross the common range of risk thresholds, the KNN model consistently demonstrated favorable screening efficiency and net clinical benefit. The similarity between the training and test set curves further supported robustness and generalizability, suggesting stable performance on new data.\u003c/p\u003e\n\u003cp\u003e3.6 Web calculator application\u003c/p\u003e\n\u003cp\u003eThe final KNN prediction model was integrated into a web-based calculator to facilitate clinical application (Figure 6). The interface is designed for straightforward clinical use: clinicians can input 14 routinely available clinical and laboratory variables and obtain an individualized AF risk prediction in real time. The output provides a visual estimate of predicted probability together with a categorical risk label to assist in patient stratification.\u003c/p\u003e\n\u003cp\u003eThe calculator operates on a secure institutional server with encrypted communication and restricted access. It is intended as a clinical support tool rather than a decision making substitute, aiming to improve the accessibility and timeliness of AF risk assessment in patients with DKD. The platform also allows for potential extension to multicenter validation and remote clinical use. The tool can be accessed at: https://endocrine-dzm-af-prediction-model.shinyapps.io/shiny/.\u0026nbsp;\u003c/p\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eTo our knowledge, this is the first clinical prediction model for atrial fibrillation (AF) in patients with diabetic kidney disease (DKD) developed using machine learning (ML). Using routinely collected information from electronic medical records, laboratory tests, and transthoracic echocardiography, we compared seven ML algorithms and identified k-nearest neighbors (KNN) as the best-performing approach in our dataset. KNN is a nonparametric, supervised method that classifies a query sample by the votes of its nearest neighbors, requires few hyperparameters, and is straightforward to implement [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Prior studies have used KNN for risk prediction in sleep apnea, coronary artery disease, and AF among intensive care unit patients [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. To improve transparency, we applied SHAP to interpret model outputs and quantify feature contributions, consistent with earlier AF modeling studies [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Key predictors highlighted by SHAP were clinically plausible and measurable at the point of care. We also implemented a web calculator to support bedside use, given that the inputs are standard components of routine evaluation.\u003c/p\u003e\u003cp\u003eBuilding on the model\u0026rsquo;s overall performance, the selected predictors reflected two major patterns: renal metabolic burden and cardiac structural functional remodeling, together outlining a biologically coherent framework for AF risk in DKD. Indicators of renal impairment and proteinuria emerged as dominant signals. Higher SCr and greater 24UTP were consistent with prior studies showing that kidney dysfunction and protein loss are linked to higher AF risk, often in a dose response pattern. [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. BUN showed similar associations and has also been retained in critical care AF prediction models [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. These findings support mechanisms involving systemic inflammation, activation of the renin\u0026ndash;angiotensin\u0026ndash;aldosterone system (RAAS), and disturbances in calcium homeostasis [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. In parallel, aging and obesity were strong clinical correlates. The risk of AF increases steadily with advancing age, likely due to atrial remodeling, reduced conduction velocity, and sinoatrial node dysfunction [\u003cspan additionalcitationids=\"CR31 CR32\" citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], while excess body weight is associated with both incident AF and its progression from paroxysmal to persistent forms through mechanisms such as hemodynamic overload, hypertension, inflammation, and oxidative stress [\u003cspan additionalcitationids=\"CR35 CR36\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. Obesity is also strongly linked to the onset and progression of DKD, highlighting a shared metabolic and structural risk pathway [\u003cspan additionalcitationids=\"CR39\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eMeanwhile, cardiac structure and function further complemented these metabolic indicators. Enlarged atrial dimensions on echocardiography correlated with declining eGFR and higher cardiovascular risk among DKD patients [\u003cspan additionalcitationids=\"CR42 CR43\" citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Left atrial volume index was independently associated with new-onset AF and adverse outcomes regardless of ejection fraction or chamber geometry, consistent with chronic volume load, inflammation, and neurohormonal activation [\u003cspan additionalcitationids=\"CR46 CR47\" citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. For the right atrium, genetic evidence from Mendelian randomization suggests that reduced eGFR may causally contribute to atrial enlargement, although DKD-specific clinical studies remain limited [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Reduced systolic performance, reflected by fractional shortening, was also consistent with the interlinked pathophysiology of AF, heart failure, and DKD [\u003cspan additionalcitationids=\"CR51 CR52 CR53\" citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. Finally, other variables such as HbA1c, CK-MB, fibrinogen, and INR contributed additional clinical context by capturing glycemic exposure, ischemic injury, coagulation activity, and anticoagulation status, all of which have recognized relevance to AF risk and management [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan additionalcitationids=\"CR56 CR57 CR58 CR59 CR60\" citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eOverall, we developed an interpretable ML-based model to estimate AF risk among patients with DKD using routinely available clinical, laboratory, and echocardiographic data. In our dataset, the model showed stable discrimination and calibration, and we implemented a web calculator to facilitate use at the point of care. However, this study also has the following limitations. First, the retrospective design introduces risks of selection and information bias, and reliance on existing records may lead to missing or misclassified data. Second, although the dataset was multicenter, all participants were from China; external validity in other populations remains to be established. Third, the model included coagulation indices such as FIB and INR, and we did not exclude patients receiving anticoagulants; although these features contributed modestly, we cannot determine whether anticoagulation affected calibration. Fourth, the sample size was moderate, which may limit statistical power and the precision of subgroup effects. Future work should expand the cohort, include additional centers, and undertake external and prospective validation.\u003c/p\u003e"},{"header":"5 Conclusions","content":"\u003cp\u003eThis study developed and internally validated an interpretable machine-learning model for predicting atrial fibrillation among patients with diabetic kidney disease. By integrating routine clinical, laboratory, and echocardiographic variables, the model achieved stable discrimination and calibration within the study cohort. The use of SHAP improved interpretability, and a web-based calculator was implemented to facilitate clinical application. These findings suggest that machine learning may assist in early AF risk stratification in DKD and support individualized management. Future prospective, multicenter validation studies are warranted to confirm generalizability and assess clinical impact.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eDKD: Diabetic kidney disease; AF: Atrial fibrillation; ML: Machine learning; eGFR: Estimated glomerular filtration rate; SCr: Serum creatinine; BUN: Blood urea nitrogen; 24UTP: 24-hour urine total protein; HbA1c: Glycated hemoglobin; CK-MB: Creatine kinase-MB; Myo: Myoglobin; WBC: White blood cell; RBC: Red blood cell; INR: International normalized ratio; APTT: Activated partial thromboplastin time; FIB: Fibrinogen; LAAPD: Left atrial anterior-posterior diameter; LAMLD: Left atrial medial-lateral diameter; LASID: Left atrial superior-inferior diameter; RAMLD: Right atrial medial-lateral diameter; RASID: Right atrial superior-inferior diameter; EF: Ejection fraction; FS: Fractional shortening; KNN: K-nearest neighbors; RF: Random forest; XGBoost: Extreme gradient boosting; LightGBM: Light gradient boosting machine; GBDT: Gradient boosting decision tree; LDA: Linear discriminant analysis; ANN: Artificial neural network; LASSO: Least absolute shrinkage and selection operator; SHAP: Shapley additive explanations; DCA: Decision-curve analysis; AUC: Area under the curve; ICD-10: International Classification of Diseases, 10th Revision; RAAS: Renin-angiotensin-aldosterone system; ROC: Receiver operating characteristic.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eXiaoran Li, Xueying Wang, Xuebing Zhang, and Shidong Wang declare that they have no conflicts of interest. The funding bodies had no role in the design, conduct, or interpretation of the study.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eThis research was supported by Funding for Clinical and Scientific Research Business Expenses of High-level Central Traditional Chinese Medicine Hospitals [DZMG-QNZX-24018].\u003c/p\u003e\n\u003cp\u003eAuthors\u0026apos; contributions\u003c/p\u003e\n\u003cp\u003eXiaoran Li and Xueying Wang and Xuebing Zhang and Shidong Wang devised the research plan. Xiaoran Li and Xueying Wang performed rigorous data analysis using advanced statistical methods, enhancing the reliability of the findings. The manuscript was skillfully composed by Xiaoran Li and Xueying Wang and Xuebing Zhang and Shidong Wang diligently revised the manuscript to improve its clarity and coherence. All authors have thoroughly reviewed and given their approval to the final version of this scholarly work.\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eAcknowledgments\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003eAvailability of data and materials\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThis study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of Dongzhimen Hospital (Approval Nos. 2025DZM-376). Since this study is a retrospective study, the need for consent to participate was waived by an Institutional Review Board.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eUSRDS: the United States Renal Data System. Am J Kidney Dis. 2003;42(6 Suppl 5):1\u0026ndash;230.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAfkarian M, Sachs MC, Kestenbaum B, Hirsch IB, Tuttle KR, Himmelfarb J, de Boer IH. Kidney disease and increased mortality risk in type 2 diabetes. J Am Soc Nephrol. 2013;24(2):302\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChristiansen CF, Schmidt M, Lamberg AL, Horvath-Puho E, Baron JA, Jespersen B, Sorensen HT. Kidney disease and risk of venous thromboembolism: a nationwide population-based case-control study. J Thromb Haemost. 2014;12(9):1449\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang CY, Chen JF, Tu ST, Lee CC, Ou HY. Microvascular disease burden and macrovascular outcomes in type 2 diabetes: Risk calculator in Taiwan. Prim Care Diabetes 2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaze AD, Jaar BG, Fonarow GC, Echouffo-Tcheugui JB. Diabetic kidney disease and risk of incident stroke among adults with type 2 diabetes. BMC Med. 2022;20(1):127.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCorwin TR, Ozieh MN, Garacci E, Palatnik A, Egede LE. The relationship between financial hardship and incident diabetic kidney disease in older US adults - a longitudinal study. BMC Nephrol. 2021;22(1):167.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLippi G, Sanchis-Gomar F, Cervellin G. Global epidemiology of atrial fibrillation: An increasing epidemic and public health challenge. Int J Stroke. 2021;16(2):217\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDu X, Ninomiya T, de Galan B, Abadir E, Chalmers J, Pillai A, Woodward M, Cooper M, Harrap S, Hamet P, et al. Risks of cardiovascular events and effects of routine blood pressure lowering among patients with type 2 diabetes and atrial fibrillation: results of the ADVANCE study. Eur Heart J. 2009;30(9):1128\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSeyed Ahmadi S, Svensson AM, Pivodic A, Rosengren A, Lind M. Risk of atrial fibrillation in persons with type 2 diabetes and the excess risk in relation to glycaemic control and renal function: a Swedish cohort study. Cardiovasc Diabetol. 2020;19(1):9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaze AD, Yuyun MF, Fonarow GC, Echouffo-Tcheugui JB. Burden of Microvascular Disease and Risk of Atrial Fibrillation in Adults with Type 2 Diabetes. Am J Med. 2022;135(9):1093\u0026ndash;100. e1092.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGadde S, Kalluru R, Cherukuri SP, Chikatimalla R, Dasaradhan T, Koneti J. Atrial Fibrillation in Chronic Kidney Disease: An Overview. Cureus. 2022;14(8):e27753.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCarrero JJ, Trevisan M, Sood MM, Barany P, Xu H, Evans M, Friberg L, Szummer K. Incident Atrial Fibrillation and the Risk of Stroke in Adults with Chronic Kidney Disease: The Stockholm CREAtinine Measurements (SCREAM) Project. Clin J Am Soc Nephrol. 2018;13(9):1314\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlonso A, Lopez FL, Matsushita K, Loehr LR, Agarwal SK, Chen LY, Soliman EZ, Astor BC, Coresh J. Chronic kidney disease is associated with the incidence of atrial fibrillation: the Atherosclerosis Risk in Communities (ARIC) study. Circulation. 2011;123(25):2946\u0026ndash;53.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWatanabe H, Watanabe T, Sasaki S, Nagai K, Roden DM, Aizawa Y. Close bidirectional relationship between chronic kidney disease and atrial fibrillation: the Niigata preventive medicine study. Am Heart J. 2009;158(4):629\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKwon S, Lee SR, Choi EK, Ahn HJ, Lee SW, Jung JH, Han KD, Oh S, Lip GYH. Association Between Atrial Fibrillation and Diabetes-Related Complications: A Nationwide Cohort Study. Diabetes Care. 2023;46(12):2240\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRichard Espiga F, Almendro Delia M, Caballero Martinez F, Monge Martin D, Neria Serrano F, Quiros Lopez R. Delayed diagnosis and missed opportunities in the early detection of atrial fibrillation: a cross-sectional study. Rev Clin Esp (Barc). 2024;224(9):560\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKaufman ES, Waldo AL. The impact of asymptomatic atrial fibrillation. J Am Coll Cardiol. 2004;43(1):53\u0026ndash;4.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang P, Lin F, Ma F, Chen Y, Liu Y, Feng X, Fang S, Zhang H, Xiao S, Yang X, et al. Clinician-artificial intelligence collaboration: A win-win solution for efficiency and reliability in atrial fibrillation diagnosis. Med. 2025;6(7):100668.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHygrell T, Viberg F, Dahlberg E, Charlton PH, Kemp Gudmundsdottir K, Mant J, Hornlund JL, Svennberg E. An artificial intelligence-based model for prediction of atrial fibrillation from single-lead sinus rhythm electrocardiograms facilitating screening. Europace. 2023;25(4):1332\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNephrology, EGoCSo. Chinese guidelines for diagnosis and treatment of diabetic kidney disease. Chin J Nephrol. 2021;37:255\u0026ndash;304.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eElectrophysiology and Pacing Branch of the Chinese Medical Association CRPCotCMDA, Huang CX. Chinese Guidelines for the management of atrial fibrillation(2025). Chin J Cardiac Pacing Electrophysiol. 2025;39(4):273\u0026ndash;343.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJeng FC, Jeng YS. Implementation of Machine Learning on Human Frequency-Following Responses: A Tutorial. Semin Hear. 2022;43(3):251\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSilva CAO, Morillo CA, Leite-Castro C, Gonzalez-Otero R, Bessani M, Gonzalez R, Castellanos JC, Otero L. Machine learning for atrial fibrillation risk prediction in patients with sleep apnea and coronary artery disease. Front Cardiovasc Med. 2022;9:1050409.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBashar SK, Ding E, Albuquerque D, Winter M, Binici S, Walkey AJ, McManus DD, Chon KH. Atrial Fibrillation Detection in ICU Patients: A Pilot Study on MIMIC III Data(). Annu Int Conf IEEE Eng Med Biol Soc. 2019;2019:298\u0026ndash;301.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuan C, Gong A, Zhao Y, Yin C, Geng L, Liu L, Yang X, Lu J, Xiao B. Interpretable machine learning model for new-onset atrial fibrillation prediction in critically ill patients: a multi-center study. Crit Care. 2024;28(1):349.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePapadopoulou A, Harding D, Slabaugh G, Marouli E, Deloukas P. Prediction of atrial fibrillation and stroke using machine learning models in UK Biobank. Heliyon. 2024;10(7):e28034.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLaukkanen JA, Zaccardi F, Karppi J, Ronkainen K, Kurl S. Reduced kidney function is a risk factor for atrial fibrillation. Nephrol (Carlton). 2016;21(8):717\u0026ndash;20.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLan Q, Zheng L, Zhou X, Wu H, Buys N, Liu Z, Sun J, Fan H. The Value of Blood Urea Nitrogen in the Prediction of Risks of Cardiovascular Disease in an Older Population. Front Cardiovasc Med. 2021;8:614117.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDing WY, Gupta D, Wong CF, Lip GYH. Pathophysiology of atrial fibrillation and chronic kidney disease. Cardiovasc Res. 2021;117(4):1046\u0026ndash;59.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchnabel RB, Yin X, Gona P, Larson MG, Beiser AS, McManus DD, Newton-Cheh C, Lubitz SA, Magnani JW, Ellinor PT, et al. 50 year trends in atrial fibrillation prevalence, incidence, risk factors, and mortality in the Framingham Heart Study: a cohort study. Lancet. 2015;386(9989):154\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKistler PM, Sanders P, Fynn SP, Stevenson IH, Spence SJ, Vohra JK, Sparks PB, Kalman JM. Electrophysiologic and electroanatomic changes in the human atrium associated with age. J Am Coll Cardiol. 2004;44(1):109\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMorillo CA. Age and Atrial Fibrillation Outcomes: Myth or Muzak, Scandinavian Lessons. JACC Clin Electrophysiol. 2025;11(1):95\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGao P, Gao X, Xie B, Tse G, Liu T. Aging and atrial fibrillation: A vicious circle. Int J Cardiol. 2024;395:131445.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWong CX, Sullivan T, Sun MT, Mahajan R, Pathak RK, Middeldorp M, Twomey D, Ganesan AN, Rangnekar G, Roberts-Thomson KC, et al. Obesity and the Risk of Incident, Post-Operative, and Post-Ablation Atrial Fibrillation: A Meta-Analysis of 626,603 Individuals in 51 Studies. JACC Clin Electrophysiol. 2015;1(3):139\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJones NR, Taylor KS, Taylor CJ, Aveyard P. Weight change and the risk of incident atrial fibrillation: a systematic review and meta-analysis. Heart. 2019;105(23):1799\u0026ndash;805.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVyas V, Lambiase P. Obesity and Atrial Fibrillation: Epidemiology, Pathophysiology and Novel Therapeutic Opportunities. Arrhythm Electrophysiol Rev. 2019;8(1):28\u0026ndash;36.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShu H, Cheng J, Li N, Zhang Z, Nie J, Peng Y, Wang Y, Wang DW, Zhou N. Obesity and atrial fibrillation: a narrative review from arrhythmogenic mechanisms to clinical significance. Cardiovasc Diabetol. 2023;22(1):192.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJiang W, Wang J, Shen X, Lu W, Wang Y, Li W, Gao Z, Xu J, Li X, Liu R, et al. Establishment and Validation of a Risk Prediction Model for Early Diabetic Kidney Disease Based on a Systematic Review and Meta-Analysis of 20 Cohorts. Diabetes Care. 2020;43(4):925\u0026ndash;33.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMan REK, Gan ATL, Fenwick EK, Gupta P, Wong MYZ, Wong TY, Tan GSW, Teo BW, Sabanayagam C, Lamoureux EL. The Relationship between Generalized and Abdominal Obesity with Diabetic Kidney Disease in Type 2 Diabetes: A Multiethnic Asian Study and Meta-Analysis. Nutrients 2018, 10(11).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKeane WF, Brenner BM, de Zeeuw D, Grunfeld JP, McGill J, Mitch WE, Ribeiro AB, Shahinfar S, Simpson RL, Snapinn SM, et al. The risk of developing end-stage renal disease in patients with type 2 diabetes and nephropathy: the RENAAL study. Kidney Int. 2003;63(4):1499\u0026ndash;507.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGong M, Xu M, Pan S, Jiang S, Jiang X. Association of Left Atrium Remodeling With Major Adverse Cardiovascular Events in Asymptomatic Type 2 Diabetes Patients With Early Chronic Kidney Disease. Rev Cardiovasc Med. 2025;26(5):27247.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKadappu KK, Abhayaratna K, Boyd A, French JK, Xuan W, Abhayaratna W, Thomas L. Independent Echocardiographic Markers of Cardiovascular Involvement in Chronic Kidney Disease: The Value of Left Atrial Function and Volume. J Am Soc Echocardiogr. 2016;29(4):359\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGong M, Xu M, Pan S, Jiang X. Association between left atrial remodeling and early renal impairment in asymptomatic patients with type 2 diabetes. BMC Cardiovasc Disord. 2025;25(1):436.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi X, Dong Y, Zheng C, Wang P, Xu M, Zou C, Wang L. Assessment of real-time three-dimensional echocardiography as a tool for evaluating left atrial volume and function in patients with type 2 diabetes mellitus. Aging. 2020;13(1):991\u0026ndash;1000.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTripepi G, Benedetto FA, Mallamaci F, Tripepi R, Malatino L, Zoccali C. Left atrial volume in end-stage renal disease: a prospective cohort study. J Hypertens. 2006;24(6):1173\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTripepi G, Benedetto FA, Mallamaci F, Tripepi R, Malatino L, Zoccali C. Left atrial volume monitoring and cardiovascular risk in patients with end-stage renal disease: a prospective cohort study. J Am Soc Nephrol. 2007;18(4):1316\u0026ndash;22.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePaoletti E, Zoccali C. A look at the upper heart chamber: the left atrium in chronic kidney disease. Nephrol Dial Transpl. 2014;29(10):1847\u0026ndash;53.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHakamaki M, Hellman T, Lankinen R, Koivuviita N, Parkka J, Kallio P, Kiviniemi T, Airaksinen KEJ, Jarvisalo MJ, Metsarinne K. Elevated Troponin T and Enlarged Left Atrium Are Associated with the Incidence of Atrial Fibrillation in Patients with CKD Stage 4\u0026ndash;5. Nephron. 2021;145(1):71\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou X, Ruan W, Zhao L, Lin K, Li J, Liu H, Wang T, Zhang G. Causal Links Between Renal Function and Cardiac Structure, Function, and Disease Risk. Glob Heart. 2024;19(1):83.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSanthanakrishnan R, Wang N, Larson MG, Magnani JW, McManus DD, Lubitz SA, Ellinor PT, Cheng S, Vasan RS, Lee DS, et al. Atrial Fibrillation Begets Heart Failure and Vice Versa: Temporal Associations and Differences in Preserved Versus Reduced Ejection Fraction. Circulation. 2016;133(5):484\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBell DSH, McGill JB, Jerkins T. Management of the 'wicked' combination of heart failure and chronic kidney disease in the patient with diabetes. Diabetes Obes Metab. 2023;25(10):2795\u0026ndash;804.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBell DS. Heart failure: the frequent, forgotten, and often fatal complication of diabetes. Diabetes Care. 2003;26(8):2433\u0026ndash;41.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVaziri SM, Larson MG, Benjamin EJ, Levy D. Echocardiographic predictors of nonrheumatic atrial fibrillation. Framingham Heart Study Circulation. 1994;89(2):724\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLeclercq ADES, Halimi JF, Fiorello F, Bertrand P, Attuel C. Evaluation of time course and predicting factors of progression of paroxysmal or persistent atrial fibrillation to permanent atrial fibrillation. Pacing Clin Electrophysiol. 2014;37(3):345\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQi W, Zhang N, Korantzopoulos P, Letsas KP, Cheng M, Di F, Tse G, Liu T, Li G. Serum glycated hemoglobin level as a predictor of atrial fibrillation: A systematic review with meta-analysis and meta-regression. PLoS ONE. 2017;12(3):e0170955.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBerton G, Cordiano R, Cucchini F, Cavuto F, Pellegrinet M, Palatini P. Atrial fibrillation during acute myocardial infarction: association with all-cause mortality and sudden death after 7-year of follow-up. Int J Clin Pract. 2009;63(5):712\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMukamal KJ, Tolstrup JS, Friberg J, Gronbaek M, Jensen G. Fibrinogen and albumin levels and risk of atrial fibrillation in men and women (the Copenhagen City Heart Study). Am J Cardiol. 2006;98(1):75\u0026ndash;81.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePisters R, Lane DA, Nieuwlaat R, de Vos CB, Crijns HJ, Lip GY. A novel user-friendly score (HAS-BLED) to assess 1-year risk of major bleeding in patients with atrial fibrillation: the Euro Heart Survey. Chest. 2010;138(5):1093\u0026ndash;100.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi-Saw-Hee FL, Blann AD, Gurney D, Lip GY. Plasma von Willebrand factor, fibrinogen and soluble P-selectin levels in paroxysmal, persistent and permanent atrial fibrillation. Effects of cardioversion and return of left atrial function. Eur Heart J. 2001;22(18):1741\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi Z, Pang M, Li Y, Yu Y, Peng T, Hu Z, Niu R, Li J, Wang X. Development and validation of a predictive model for new-onset atrial fibrillation in sepsis based on clinical risk factors. Front Cardiovasc Med. 2022;9:968615.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKannel WB, Wolf PA, Benjamin EJ, Levy D. Prevalence, incidence, prognosis, and predisposing conditions for atrial fibrillation: population-based estimates. Am J Cardiol. 1998;82(8A):N2\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 to 3 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"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":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Diabetic kidney disease (DKD), Atrial fibrillation (AF), Risk prediction, Machine learning (ML), Shapley Additive Explanations (SHAP)","lastPublishedDoi":"10.21203/rs.3.rs-8146262/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8146262/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003ePatients with diabetic kidney disease (DKD) have elevated atrial fibrillation (AF) risk, yet population-specific prediction tools are limited. We aimed to develop and internally validate an interpretable machine-learning (ML) model for AF risk in hospitalized DKD.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eIn this retrospective cohort from two hospitals (January 2021 to December 2024), 787 unique DKD admissions were randomly split into training (70%) and test (30%) sets. AF at index admission was ascertained from electrocardiograms, Holter monitoring when available, and ICD-10 codes with physician adjudication. Candidate predictors were routine clinical, laboratory, and echocardiographic variables. Least absolute shrinkage and selection operator (LASSO) selected features in the training set. Seven supervised models were trained; performance was assessed by area under the receiver-operating characteristic curve (AUC), calibration, and decision-curve analysis. SHAP quantified predictor contributions.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eLASSO retained 14 features, including 24-hour urine total protein (24UTP), serum creatinine (SCr), age, and left atrial diameters. In the test set, k-nearest neighbors (KNN) achieved AUC 0.927, accuracy 0.886, sensitivity 0.920, and specificity 0.856; calibration was good and decision curves showed net benefit across common thresholds. Five-fold cross-validation yielded mean AUC 0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02. SHAP indicated proteinuria burden, renal dysfunction, age, and atrial size as leading contributors. The finalized model was deployed as a secure web calculator using routine inputs.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eAn interpretable ML-based model using standard clinical and echocardiographic data showed stable internal performance for AF risk estimation in DKD, with an accompanying web calculator for point-of-care use. Prospective multicenter studies are needed to confirm generalizability and clinical impact.\u003c/p\u003e","manuscriptTitle":"Development and Internal Validation of an Interpretable Machine Learning Model for Predicting Atrial Fibrillation in Patients with Diabetic Kidney Disease: A Multicenter Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-04 00:13:52","doi":"10.21203/rs.3.rs-8146262/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e2a013c6-c3a1-4fa8-978b-415cbb4b0f78","owner":[],"postedDate":"December 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-10T16:01:37+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-04 00:13:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8146262","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8146262","identity":"rs-8146262","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}