JMIR Publications

ABSTRACT

Background:

Machine learning (ML) and deep learning (DL) show promise for fall risk prediction, but prior reviews focused mainly on real-time fall detection, in-hospital falls, or conventional statistical models. The predictive performance and methodological quality of ML/DL-based models for predicting future falls in community-dwelling older adults remain unclear.

Objective:

To systematically review ML/DL studies for predicting future falls among community-dwelling older adults and meta-analyze discrimination where feasible.

Methods:

PubMed, Embase, Web of Science Core Collection, CINAHL, Cochrane Library, and IEEE Xplore were searched from inception to September 23, 2024, with an update on August 31, 2025. We included longitudinal studies developing or validating ML/DL models using baseline predictors to predict future falls in community-dwelling adults aged ≥60 years, and excluded real-time detection, simulated/no-fall, and inpatient studies. Risk of bias was assessed using PROBAST. We meta-analyzed logit-transformed AUCs using Hartung-Knapp-Sidik-Jonkman (HKSJ) random-effects models, reporting 95% confidence intervals (CIs). Heterogeneity was quantified using I2, τ2, τ, and Q statistic. To convey expected out-of-sample performance, we additionally reported 95% prediction intervals (PIs) using four estimators (HTS, HTS-HK, HTS-SJ, and confidence-distribution bootstrap). Robustness was assessed by leave-one-out analyses. Small-study effects were examined using funnel plots and Egger-type tests. Prespecified subgroup analyses were conducted.

Results:

After screening 6,865 records, 27 studies were included; 18 studies focused on general community-dwelling older adults. Prediction horizons range from 3 months to 4 years, and fall incidence ranged from 1.6% to 46.6%. Twenty-two studies applied ML, and five used DL. Input modalities included text (n=17), sensor (n=5), image (n=1), and multimodal data (n=4). Common predictors included age (n=14), sex (n=13), fall history (n=12), depression (n=10), and basic activities of daily living (n=8). Only one study reported external validation. Calibration reporting was sparse. All models were rated at high risk of bias. Ten internally validated models were meta-analyzed, yielding a pooled AUC of 0.79 (95%CI 0.69-0.87) with extreme heterogeneity (τ2 = 0.64; τ = 0.80; I2 = 99.8%; Q = 4128.99). Corresponding 95% PIs were wide (HTS: 0.28-0.98; HTS-HK: 0.40-0.96; HTS-SJ: 0.40-0.96; confidence-distribution bootstrap: 0.20-0.99). Leave-one-out analyses were stable (AUC range 0.75-0.80), with no strong evidence of small-study effects. Subgroup analyses indicated moderation by sample size (≤500 vs ＞500) and population type (general vs specific conditions). All findings should be interpreted cautiously.

Conclusions:

Current ML/DL models show potential for identifying community-dwelling older adults at elevated future fall risk; however, extremely wide PIs, limited external validation, and high risk of bias suggest real-world performance may be optimistic. Before implementation, models should undergo external validation, calibration, and prospective impact evaluation. By integrating PROBAST with HKSJ meta-analysis and multiple PI estimators, and mapping algorithm families to data modalities, this review provides practical guidance for selecting and evaluating fall-risk prediction approaches. Clinical Trial: PROSPERO registration: CRD42024580902.