Machine learning-based risk of fall estimation using insole with force sensors while performing a sequence of activities in the TUG test

Abstract

Several methods combining biomedical and computer-based approaches have been used to address the risk of falls among the elderly using instrumented insoles. Machine-learning techniques in gait analysis has proven to be a promising solution when using instrumented insoles. However, no study has investigated the risk of falls associated with a sequence of activities. Indeed, it can be observed that an important amount of energy is required by individuals preparing to get out of bed or toilet. The main goal of this work is to detect and associate different risk levels by analyzing the sit-to-start-of-walk (STSOW) sequence. Data were acquired during a Timed Up and Go test using an instrumented insole. The proposed approach compares six types of classifiers to the STSOW sequence signals. Then, a recursive clustering approach based on statistical features and the Kruskal Wallis test was implemented to define different levels of risk. The results show the capacity of the proposed approach to associate different risk levels of falls to an STSOW sequence. The accuracies of the classifiers ranged from 69% to 95.2%, and the best accuracy was achieved using both decision tree and ensemble classifiers. For the sit-to-start of the walk sequence identification phase, the best accuracy was achieved using the support vector machine model.