Machine learning algorithms to predict frailty in older adults in China: a cross-sectional study.

objectiveTo investigate the predictive value of Machine Learning (ML) for the occurrence of frailty in Chinese elderly and to identify significant factors associated with frailty.

methodsThis study utilized data from the 2020 China Health and Retirement Longitudinal Study (CHARLS), focusing on 7880 elderly individuals. Six machine learning (ML) algorithms were employed: Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree (DT), LightGBM, and Random Forest (RF). The performance of these models was assessed using several metrics including the Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, Negative Predictive Value (NPV), Positive Predictive Value (PPV), and Decision Curve Analysis (DCA). To compare the differences in ROC curves among different models, the Delong test was applied, followed by Bonferroni correction to adjust for multiple comparisons and control for Type I error. Additionally, the Brier score was used to evaluate the calibration of the prediction model.Permutation Importance and SHAP analysis were conducted to interpret feature contributions.

resultsAUC values ranged from 0.722 to 0.816, with LightGBM achieving the highest predictive performance and demonstrating the greatest net benefit across different threshold probability ranges. Feature importance analysis revealed Multimorbidity, Depression, Self-rated Health, Social Events, and Drinking as the top predictors. Permutation Importance indicated differences in ranking between LightGBM and Random Forest, with SHAP analysis showing LightGBM's ability to capture nonlinear effects, while Random Forest exhibited a more balanced global interpretation.

conclusionML algorithms can effectively predict the occurrence of frailty in Chinese elderly and identify significant factors associated with frailty. For datasets with complex feature interactions, LightGBM is preferable, while Random Forest offers improved stability and variance control. These findings support ML-based frailty screening and intervention strategies for aging populations.