Rapid decline in kidney function (RDKF) is an early and modifiable precursor of chronic kidney disease, yet proactive risk identification remains limited in community-based populations.
This study aimed to develop, validate, and deploy an interpretable machine learning model to predict 4-year RDKF risk among community-dwelling middle-aged and older adults in China.
Six machine learning models were developed, and hyperparameters were tuned using grid search with 5-fold cross-validation.
During 4 years of follow-up, 361 participants (7.20%) developed RDKF.
We developed an interpretable machine-learning model and implemented it as a web-based tool for predicting RDKF risk.
Rapid decline in kidney function (RDKF) is an early and modifiable precursor of chronic kidney disease, yet proactive risk identification remains limited in community-based populations. This study aimed to develop, validate, and deploy an interpretable machine learning model to predict 4-year RDKF risk among community-dwelling middle-aged and older adults in China. We used data from 5,012 participants in the China Health and Retirement Longitudinal Study (2011–2015). After preprocessing and imputation, data were randomly split into training and testing cohorts (7:3). Boruta was used for feature selection. Six machine learning models were developed, and hyperparameters were tuned using grid search with 5-fold cross-validation. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), Brier score, calibration, and decision curve analysis. SHapley Additive exPlanations (SHAP) was applied for model interpretability. The optimal model was deployed through a publicly available web-based platform. During 4 years of follow-up, 361 participants (7.20%) developed RDKF. Sixteen predictors were selected, including lipid profiles, glycemic markers, cardiometabolic conditions, medication use, and hematologic or renal-related biomarkers. In the testing cohort, the support vector machine model achieved the highest discrimination (AUC: 0.706, 95% CI: 0.652–0.760), with relatively good calibration and a positive net benefit across the evaluated threshold range. SHAP analysis identified TC, LDL-C, HDL-C, and hypertension as the most important predictors. The online platform supports individual and batch prediction with SHAP-based explanations. We developed an interpretable machine-learning model and implemented it as a web-based tool for predicting RDKF risk. The tool may support preliminary risk stratification in community settings, particularly where resources are limited. Independent external validation is warranted to establish its generalizability and practical utility.
