Daily fall risk assessment for older adults using a balance sensor with machine learning

Fall-risk assessment is important for preventing falls in older adults, but many conventional balance tests are time-consuming and difficult to use frequently in daily practice. This study aimed to evaluate a rapid 30-s quiet-standing assessment using a balance sensor combined with machine learning for fall-history classification and fall-risk assessment.

A total of 394 older adults (mean age 71.56 ± 6.69 years) participated, including 91 individuals in a laboratory cohort, 291 in an independent hospital cohort, and 12 in a 10-day home-based feasibility study. The balance sensor captured plantar pressure-distribution images during quiet standing. From these recordings, center-of-pressure (CoP) trajectories and model-derived center-of-gravity (CoG) descriptors were obtained. Spatial, stability, frequency, temporal, and statistical features were extracted, and multiple machine-learning models were evaluated. Model development in the laboratory cohort used nested stratified cross-validation, and the final pipeline was externally validated in the hospital cohort.

In the laboratory cohort, the highest area under the receiver operating characteristic curve (AUC) was 0.71, achieved by the random forest model under the eyes-open condition. Under the eyes-closed condition, the highest AUC was 0.68. In the independent hospital cohort, external validation of the final laboratory-developed model yielded an AUC of 0.68 and an accuracy of 72.0%, which was numerically higher than those of the conventional functional assessments evaluated in this study. The home-based component supported the feasibility of repeated short-duration balance measurement outside the laboratory.

These findings suggest that a brief sensor-based standing assessment combined with machine learning may provide a practical approach for fall-history classification and fall-risk assessment in older adults. However, the main analyses were based on retrospective fall history or documented clinical fall risk rather than prospectively observed future falls. Larger prospective and multi-center studies are needed to determine predictive validity and broader clinical utility.