Validation of an accelerometer-based fall prediction model

Falls are a common and serious problem faced by older populations. There is a growing interest in estimating the risk of falling for older people using body-worn sensors and simple movement tasks, allowing appropriate fall prevention programs to be administered in a timely manner to the high risk population. This study investigated the capability and validity of using a waist-mounted triaxial accelerometer (TA) and a directed routine (DR) that includes three movement tasks to discriminate between fallers and non-fallers and between multiple fallers and non-multiple fallers. Data were collected from 98 subjects who were stratified into two separate groups, one for model training and the other for model validation. Logistic regression models were constructed using the TA features from the entire DR and from each single DR task, and were validated using unseen data. The best models were obtained using features from the alternate step test to classify between fallers and non-fallers with κ = 0.34-0.41, sensitivity = 68%-71% and specificity = 63%-73%. However, the overall validation performances were poor. The study emphasizes the importance of independent validation in fall prediction studies.