Explainable Machine Learning Approaches Predict Frailty and Adverse Outcomes in Older Adults: Development and Validation with Two Longitudinal Cohorts

TL;DR

This study developed and validated a machine learning model to predict frailty and related health risks in older adults using explainable AI techniques.

Contribution

The novel contribution is an interpretable machine learning model for predicting frailty with high accuracy and generalizability across two cohorts.

Findings

XGBoost achieved high performance (AUC = 0.934 in internal validation and AUC = 0.792 in external validation).

Frailty prediction was associated with increased risks of falls, hospitalization, and disability.

Key predictors included lifestyle and psychological factors like BMI, self-rated health, and depression.

Abstract

Objectives: Early and accurate identification of frailty is essential for preventing adverse outcomes in older adults. However, existing frailty prediction models often lack reliability, interpretability, and generalizability. Methods: Participants aged 60 years and older between 2011 and 2015 (n = 3419) from the CHARLS were used to develop models, and participants from the CLHLS-HF between 2014 and 2018 (n = 1017) were used for external validation. The frailty was assessed 4 years after baseline in both cohorts by Fried’s Frailty Phenotype (FFP). Six machine learning models were applied to develop prediction models. The SHapley Additive exPlanations (SHAP) method was utilized to explain the final model. Clinical outcomes were evaluated between participants predicted as frail and non-frail by the final model. Results: The XGBoost (AUC = 0.934, 95% CI: 0.921–0.948; F1 = 0.712, 95% CI: 0.686–0.736 in internal validation; AUC = 0.792, 95% CI: 0.750–0.830; F1 = 0.702, 95% CI: 0.652–0.753 in external validation) performed best among six models. Key predictors included lifestyle factors (e.g., instrumental daily living activities, BMI, and self-rated health) and psychological traits (e.g., depression). Participants predicted as frail had significantly elevated risks of falls (OR = 2.11), hospitalization (OR = 1.75), and disability (OR = 1.42). Conclusions: The proposed model provided a robust and interpretable digital tool for predicting frailty among older adults and associated adverse outcomes.

Figures