Machine learning-based fall risk assessment using gait data

Introduction: Falls affect approximately 30% of older adults annually and represent a major public health burden. Existing research has developed standardised tests for fall risk (FR) assessment or correlated movement tests with fall events using machine learning (ML). However, there is a lack of studies including large and heterogeneous populations to test the generalisability of different ML methods. The aim of this study was to compare gait analysis-based classification of FR using ML with comprehensive, established tests in a large multicentre cohort.

Methods: A total of 443 participants were recruited (mean age 71 ± 9 years), of whom 145 experienced falls. Participants completed a standardised test battery comprising 6-metre walks, the 4-Stage Balance Test, Step Test, Functional Forward Reach, Timed Up-and-Go, 30 s Sit-to-Stand, and handgrip strength testing. Additional data were collected via a structured medical interview covering comorbidities, fall-risk-increasing drugs, lower extremity pain and stiffness, sleep habits, physical activity (GPAQ), fall history, and STEADI fall-risk screening. Falls were prospectively followed up after 6 months. Movement tests were recorded using a markerless motion capture system with eight cameras. Participants were classified as ‘at-risk’ or ‘not-at-risk’ of falls based on falls in the 12 months prior to and 6 months following assessment. Gait features were derived from event-based analysis of walking trials and included spatiotemporal parameters, gait properties such as asymmetries, clearance, range of motion, and derivatives. Statistical preprocessing accounted for multiple comparisons and parameter dependencies using the Benjamini–Hochberg procedure and correlation filtering. Feature selection was performed on gait parameters using LASSO logistic regression with cross-validation.

Results: Gait-based classification achieved higher accuracy than the best-performing clinical test, improving accuracy from 0.79 to 0.85 using only two walking trials.

Conclusion: Automated gait analysis enables accurate FR classification and may offer practical advantages over conventional screening methods in everyday clinical settings.