Development and multi-cohort validation of a machine learning-based simplified frailty assessment tool for clinical risk prediction

Abstract

Background: Frailty significantly impacts health outcomes in aging populations, yet its routine assessment remains challenging due to the complexity and time-consuming nature of existing tools. This study aimed to develop and validate a clinically feasible, machine learning-based frailty assessment tool that balances predictive accuracy with implementation simplicity in real-world clinical settings.

Methods: We conducted a multi-cohort study leveraging data from the National Health and Nutrition Examination Survey (NHANES, n = 3,480), China Health and Retirement Longitudinal Study (CHARLS, n = 16,792), China Health and Nutrition Survey (CHNS, n = 6,035), and Sun Yat-sen University Third Affiliated Hospital CKD cohort (SYSU3 CKD, n = 2,264). Through systematic application of five complementary feature selection algorithms to 75 potential variables, followed by comparative evaluation of 12 machine learning approaches, we developed a parsimonious assessment tool for predicting frailty diagnosis, chronic kidney disease progression, cardiovascular events, and all-cause mortality.

Results: Our analysis identified a minimal set of just eight readily available clinical parameters— age, sex, body mass index (BMI), pulse pressure, creatinine, hemoglobin, and preparing meals difficulty and lifting/carrying difficulty—that demonstrated robust predictive power. The extreme gradient boosting (XGBoost) algorithm exhibited superior performance across training (AUC 0.963, 95% CI: 0.951–0.975), internal validation (AUC 0.940, 95% CI: 0.924–0.956), and external validation (AUC 0.850, 95% CI: 0.832–0.868) datasets. This model significantly outperformed traditional frailty indices in predicting CKD progression (AUC 0.916 vs. 0.701, p < 0.001), cardiovascular events (AUC 0.789 vs. 0.708, p < 0.001), and mortality (time-dependent AUC 0.767 − 0.702 vs. 0.690 − 0.627, p < 0.001). SHAP analysis provided transparent insights into model predictions, facilitating clinical interpretation.

Conclusion: Our simplified frailty assessment tool demonstrates robust performance across multiple health outcomes while minimizing measurement burden. The model’s superior predictive capabilities for CKD progression, cardiovascular events, and mortality underscore its potential utility for risk stratification.

Supplementary Information: The online version contains supplementary material available at 10.1186/s12967-025-06728-4.

Keywords: Cardiovascular risk; Chronic kidney disease; Frailty assessment; Machine learning; Mortality prediction; Risk stratification; XGBoost.

Fig. 1

Comprehensive Methodology Framework for Machine Learning-Based Frailty Assessment Tool Development

Fig. 2

Intersection Analysis of Features Selected by Five Machine Learning Algorithms

Fig. 3

Comparison of ROC Curves Between Machine Learning-Based Frailty Diagnostic Model and Traditional Frailty Index Across Different Cohorts. (A) NHANES cohort and (B) CHARLS external validation cohort

Fig. 4

Clinical Utility of the Machine Learning Frailty Model for Predicting Multiple Health Outcomes Across Diverse Cohorts. (A) CKD progression prediction in SYSU3 CKD cohort; (B) Comparison with traditional frailty index for rapid kidney function decline prediction in CHARLS; (C) Cardiovascular event prediction in CHNS; (D) Comparison with traditional frailty index for cardiovascular outcomes in CHARLS; (E) Time-dependent mortality prediction in NHANES compared with traditional frailty metrics; (F) Kaplan-Meier survival curves stratified by predicted frailty risk categories

Fig. 5

SHAP Analysis and Clinical Web Application of the Machine Learning Frailty Model. (A) SHAP Values of Feature Variables and Their Impact on Prediction Results (B) Web Application Interface for clinical implementation