Comparing Machine Learning Approaches for Fall Risk Assessment

Abstract

Traditional fall risk assessment tests are based on timing certain physical tasks, such as the timed up and go test, counting the number of repetitions in a certain time-frame, as the 30-second sit-to-stand or observation such as the 4-stage balance test. A systematic comparison of multifactorial assessment tools and their instrumentation for fall risk classification based on machine learning approaches were studied for a population of 296 community-dwelling older persons aged above 50 years old. Using features from inertial sensors and a pressure platform by opposition to using solely the tests scores and personal metrics increased the F-Score of Naive Bayes classifier from 72.85% to 92.61%. Functional abilities revealed higher association with fall level than personal conditions such as gender, age and health conditions.