# Prediction of frailty in community older adults based on machine learning: a systematic review and meta-analysis

## TL;DR

This paper reviews and analyzes machine learning models for predicting frailty in older adults, finding moderate to good performance but highlighting the need for better quality and applicability.

## Contribution

A systematic review and meta-analysis of ML-based frailty prediction models in community-dwelling older adults, evaluating their performance and quality.

## Key findings

- The pooled AUC for baseline frailty prediction was 0.878, indicating strong discrimination.
- Longitudinal prediction models had a lower pooled AUC of 0.730, suggesting room for improvement.
- Overall, the pooled AUC across all studies was 0.786, showing moderate predictive accuracy.

## Abstract

An increasing number of predictive models for frailty in community-dwelling older adults are now being developed using machine learning methods. Differences between model performances limit their practical application. Therefore, we conducted a systematic review and meta-analysis to summarize and evaluate the performance and clinical applicability of these risk prediction models.

PubMed, Web of Science, Embase, Cochrane Library, Scopus, CINAHL, SinoMed, VIP, CNKI, and Wanfang were searched. The search time was from the database establishment to June 10, 2025. The PROBAST+AI assessment tool was used to assess the study quality, and a meta-analysis of the area under the curve (AUC) was performed using Stata18.0 software.

A total of 10 studies were included, and 45 Machine Learning (ML) models were developed, of which 36 models were developed for internal validation and 9 for external validation. In the internal validation set, the pooled AUC for baseline frailty prediction studies was 0.878 (95% CI 0.799, 0.958), while the pooled AUC for longitudinal frailty prediction studies was 0.730 (0.670, 0.790). When all studies were pooled without distinguishing prediction time points, the overall pooled AUC was 0.786 (95% CI 0.697, 0.875).

Although most of the included models had good discrimination and calibration, the overall quality and applicability of the current study are still problematic. In future studies, researchers should follow the TRIPOD+AI statement and the PROBAST+AI list to construct high-quality, more applicable predictive models.

https://www.crd.york.ac.uk/PROSPERO/view/CRD420251071061, identifier CRD420251071061.