Abstract
BACKGROUND: Physical frailty is a pressing public health issue that significantly increases the risk of disability, hospitalization, and mortality. Early and accurate detection of frailty is essential for timely intervention, reducing its widespread impact on healthcare systems, social support networks, and economic stability. OBJECTIVE: This study aimed to classify frailty status into binary (frail vs. non-frail) and multi-class (frail vs. pre-frail vs. non-frail) categories. The goal was to detect and classify frailty status at a specific point in time. Model development and internal validation were conducted using data from wave 8 of the English Longitudinal Study of Ageing (ELSA), with external validation using wave 6 data to assess model generalizability. METHODS: Nine classification algorithms, including Logistic Regression, Random Forest, K-nearest Neighbor, Gradient Boosting, AdaBoost, XGBoost, LightGBM, CatBoost, and Multi-Layer Perceptron, were evaluated and their performance compared. RESULTS: CatBoost demonstrated the best overall performance in binary classification, achieving high recall (0.951), balanced accuracy (0.928), and the lowest Brier score (0.049) on the internal validation set, and maintaining strong performance externally with a recall of 0.950, balanced accuracy of 0.913, and F1-score of 0.951. Multi-class classification was more challenging, with Gradient Boosting emerging as the top model, achieving the highest recall (0.666) and precision (0.663) on the external validation set, with a strong F1-score (0.664) and reasonable calibration (Brier Score = 0.223). CONCLUSION: Machine learning algorithms show promise for the detection of current frailty status, particularly in binary classification. However, distinguishing between frailty subcategories remains challenging, highlighting the need for improved models and feature selection strategies to enhance multi-class classification accuracy.