Deep Learning for Frailty Classification Using IMU Sensor Data: Insights From FRAILPOL Database

Abstract

Frailty is a significant health issue among the elders that leads to adverse outcomes, such as disability, increased healthcare utilization, and diminished overall well-being. This results in significant economic and personal costs, adding a considerable burden on the healthcare system. Early frailty detection contributes to establish a sustainable society by addressing health challenges and enhancing healthcare systems to be more responsive to aging populations. The FRAILPOL dataset was assessed, which contains data of 682 elderly participants. The dataset includes recordings from five inertial measurement unit (IMU) sensors aimed at classifying individuals into frail, prefrail, or robust (nonfrail) categories. The IMU data, comprising accelerometer and gyroscope readings from gait, are directly input into the InceptionTime, a deep learning (DL) algorithm for the classification task. The InceptionTime algorithm achieved an average accuracy of 81% on the test data. For the critical early frailty stage (prefrail), it reported 81% precision, 81% recall, and an F1-score of 80%. These findings can provide a valuable diagnostic tool for identifying frailty in its early stages, which can significantly contribute to the mitigation of frailty progression. In addition, the relatively simple observation of an elderly person’s gait can be an important social factor toward recognizing frailty syndrome.