Smart fall prediction paradigm for community-dwelling seniors through fitness screening protocols and machine learning

Abstract

This study aimed to predict the fall risk of community-dwelling seniors through fitness screening protocols and machine learning algorithms. Specifically, a customized questionnaire, simple functional performance tests (six items), and physical fitness tests (six items) were employed. A data processing was conducted to enhance the performance, including a deep neural network-based imputation approach (DataWig), an adaptive index filtration algorithm (multi-ReliefF), and decision tree (DT)-based prediction models. Historical data contained 215 community-dwelling seniors were used, and results showed that the proposed paradigm achieved promising and stable performance (e.g., accuracy: 0.933–0.950, F1-score: 0.951–0.955). The developed DT models presented satisfactory performance (top three in almost all evaluation metrics), and the utilized multi-ReliefF enhanced the accuracy (from 1.0% to 3.1%). These results indicated that the proposed paradigm could be a helpful tool for monitoring the health conditions of seniors during their daily exercises, which could help reduce the risk of falls.