Machine learning prediction of CKD progression in hyperglycemic elderly adults: a prospective community cohort study

Hyperglycemia is a major risk factor for chronic kidney disease (CKD). This multicenter prospective study developed and validated a machine learning (ML) model to predict CKD risk in prediabetic and diabetic populations for early intervention, following TRIPOD+AI guidelines. Participants were enrolled from four communities, with three sites providing training (80%) and internal test (20%) datasets, and the fourth for external validation. Five ML algorithms were constructed, and SHapley Additive exPlanations (SHAP) was applied to interpret the optimal model. The XGBoost model showed excellent predictive performance, with AUCs of 0.905, 0.809, and 0.837 in training, internal test, and external validation sets, respectively. Serum creatinine (Scr), age, and hemoglobin (Hb) were the leading predictors, with higher Scr, older age, and lower Hb elevating CKD risk. Risk stratification (low: 0%–5%, medium: 5%–25%, high: 25%–100%) yielded distinct CKD incidences of 0.7%, 9.9%, and 55.5% (<i>p</i> &lt; 0.001). An online prediction tool was further established for community screening. This validated ML model enables accurate risk prediction and stratification in hyperglycemic individuals, providing a feasible approach for early CKD detection and targeted prevention in community healthcare. <b>Trial registration:</b> Not applicable. The study is not a clinical trial. Multi-panel poster summarizing machine learning methodology, results, and application for CKD risk prediction.A poster titled 'Machine Learning Prediction of CKD Risk in Hyperglycemic Populations,' is divided into two main vertical sections: 'Background and Methods' on the left and 'Results and Conclusions' on the right. Left Panel: Background and Methods This section introduces the challenge of CKD risk prediction in hyperglycemic patients, accompanied by IV bag and kidney icons. A flowchart details the data partitioning process: four 'Community' datasets merge and are then split into an 80% 'Training Set', a 20% 'Internal Test Set', and an 'External Test Set'. Text describes the evaluation of multiple machine learning algorithms, with XGBoost selected for its superior predictive performance and robustness, and SHAP values used for model interpretation. Right Panel: Results and Conclusions This section presents the study's findings. It reports the model's performance with high AUC values for training, internal testing, and external validation. Three small plots illustrate model evaluation: Panel A shows a calibration curve comparing predicted and true probabilities; Panel B is a Receiver Operating Characteristic (ROC) curve; and Panel C displays sensitivity versus false positive rate curves. Text identifies Serum Creatinine, Age, and Hemoglobin as the top predictors from SHAP analysis. Three subsequent plots further visualize feature importance: Panel A is a horizontal bar chart ranking features; Panel B is a pie chart showing SHAP value percentages; and Panel C displays violin plots illustrating the distribution of SHAP values for various features. 'Table 2 Model performance' provides a numerical comparison of AUC, Accuracy, Specificity, Sensitivity, Precision, Recall, F1-score, and G-mean across five machine learning models, highlighting XGBoost's strong AUC. Finally, three application interface mockups demonstrate an integrated online application. Panel A shows a probability density distribution with a threshold; Panel B displays a dashboard for comparing model performance; and Panel C illustrates a prediction results panel incorporating a SHAP feature contribution plot. Concluding text summarizes the validated ML model's ability to predict CKD risk and enable scalable community-based early screening. 1. Current Knowledge Hyperglycemic states accelerate chronic kidney disease (CKD) progression. While many CKD prediction models exist, data specifically for predicting future CKD onset in hyperglycemic individuals (prediabetes and diabetes) remain limited. 2. Novel Findings and Future Implications Routine physical examination indicators can predict CKD in hyperglycemic patients well, with the XGBoost model showing optimal performance. Individuals with high predicted CKD risk (actual incidence over 50%) require close monitoring. Our model provides valuable insights for assessing renal function and enabling risk stratification in this population, while the web application built on the optimized model enhances clinical utility.