Feature selection for elderly faller classification based on wearable sensors.

The study used a convenience sample of 100 community-dwelling older adults, so the people may not match a therapist's usual patients. Participants had to walk six minutes without an assistive device, so it does not apply to patients who cannot walk that far. Fall status was based on six-month recall, which can be inaccurate, and gait may change after a fall or because of fear of falling. Sensor data were missing for two non-fallers at the pelvis and one non-faller at the left shank. Feature selection was done on the entire dataset, which may have made the selected feature sets fit this sample too well; the authors said results need confirmation in a new population. The single train-test split gave performance at the upper end of the repeated model performances, and model performance varied widely across splits. The walking trial was only 7.62 m, which may not reflect everyday walking.

The supplied text contains a funding and conflicts heading with 'Funder named' but no funder name or author declarations.

Do not read the single train-test split result of 96% accuracy and 100% faller detection as evidence this is a usable fall screen. When the same model was tested across 10,000 random train-test splits, it classified 74% of people correctly and detected 44% of fallers, and the authors said the lower repeated performance is likely more indicative of future model performance.