Predicting Falls and Mobility Decline Among Older Adults Using Longitudinal Functional Data and Ensemble Machine Learning

Abstract

Falls and mobility decline represent major threats to independence, safety, and health among older adults. This quantitative study developed and evaluated an ensemble machine-learning framework for predicting falls and mobility deterioration from longitudinal functional, clinical, psychological, and wearable-sensor data. A multisite prospective cohort design followed 1,178 adults aged 65 years and older across 12 community, outpatient, rehabilitation, and assisted-living sites for 24 months. Functional assessments were completed at baseline and at 6-, 12-, 18-, and 24-month intervals, while falls were monitored monthly. The dataset included 5,505 functional assessments, 5,268 valid wearable-monitoring periods, and 4,241 six-month prediction windows. During follow-up, 44.3% of participants experienced at least one fall, 19.7% experienced recurrent falls, 18.2% experienced an injurious fall, and 34.9% developed clinically meaningful mobility decline. Mean gait speed decreased by 0.082 meters per second, Short Physical Performance Battery scores declined by 0.91 points, and daily steps decreased by 632. Previous falls increased the adjusted odds of a subsequent fall by 118%, while frailty increased the odds of mobility decline by 105%. Stable, gradual-decline, accelerated-decline, fluctuating, and recovery trajectories accounted for 33.4%, 29.9%, 16.0%, 13.3%, and 7.4% of participants, respectively. Fall occurrence reached 68.8% in the accelerated-decline group compared with 25.4% in the stable group. Logistic regression and multiple individual and ensemble algorithms were compared using participant-level cross-validation, temporal validation, and geographically independent external testing. The stacked ensemble achieved external receiver operating characteristic areas of .842 for falls and .873 for mobility decline, exceeding logistic regression by .128 and .135. Corresponding precision–recall areas were .679 and .736, while Brier scores were .128 and .113. Calibration slopes of .96 and .98 indicated close agreement between predicted and observed risks. Longitudinal gait-speed slopes, physical-performance deterioration, gait variability, activity reduction, previous falls, frailty, and hospitalization contributed most strongly to prediction. The findings established that longitudinal functional information and ensemble machine learning predicted falls and mobility decline more accurately than baseline assessment and conventional regression alone.