# A deep learning framework for gait-based frailty classification using inertial measurement units

## TL;DR

This paper introduces a deep learning method using wearable sensors to classify frailty in older adults based on gait data.

## Contribution

A novel participant-centric data partitioning framework and the use of InceptionTime for gait-based frailty classification.

## Key findings

- InceptionTime achieved 82% accuracy on the GSTRIDE dataset for frailty classification.
- The method showed 79% accuracy on the FRAILPOL dataset with strong AUC and F1-score results.
- Spatio-temporal features from IMU signals were effectively captured for frailty detection.

## Abstract

Frailty in older adults leads to heightened vulnerability to adverse health outcomes, significantly burdening individuals and society by increasing healthcare costs and dependency. To address this issue, an advanced frailty assessment method combining wearable sensors measurements with Deep Learning (DL) techniques is proposed to classify individuals into frail or non-frail stages. Wearable sensors provide real-time monitoring, facilitating early detection and timely interventions. Two diverse datasets, i.e., GSTRIDE and FRAILPOL, were utilized for enhanced frailty analysis, employing one to five Inertial Measurement Unit (IMU) sensors with varying configurations and mounting positions. A participant-centric data partitioning framework based on signal windows segmentation is proposed and applied to DL algorithms. Among the DL algorithms, InceptionTime outperformed, achieving 82% accuracy on GSTRIDE and 79% on the FRAILPOL dataset. Furthermore, the area under the ROC curve (AUC) and evaluation metrics such as precision, recall, and F1-score confirm InceptionTime’s effectiveness in classifying frail and non-frail stages by capturing spatio-temporal features from raw IMU signals.