Comparative evaluation of deep learning models for three-class frailty assessment using gait metrics

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• Traditional frailty assessment tools are subjective and time-consuming.
• Wearable sensor-based gait analysis offers an objective alternative for frailty screening.
• Limited studies comprehensively evaluate deep learning for multi-class frailty classification using gait metrics.

Purpose of the Study:

• To evaluate deep learning architectures for three-class frailty classification.
• To utilize structured gait features from a single wearable inertial measurement unit (IMU).
• To assess model performance using a standardized protocol and cross-validation.

Main Methods:

• Four deep learning architectures (Transformer, ShapeFormer, InceptionTime, LSTM-CNN) were tested.
• The GSTRIDE dataset of 163 older adults was used, with gait variables extracted from a single IMU.
• A 10-fold participant-level cross-validation ensured robust performance estimation.

Main Results:

• All four architectures showed comparable mean accuracy (65.9%–73.1%).
• No significant differences in performance were found between models.
• Classification accuracy was higher for non-frail and pre-frail individuals, with frail individuals posing a challenge.

Conclusions:

• Multi-class frailty classification using wearable gait sensors is feasible.
• Model performance varies across architectures and frailty classes.
• Model stability and class-specific sensitivity are crucial considerations for clinical application.