Deep Learning for Sensor-Based Rehabilitation Exercise Recognition and Evaluation.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Science

Background:

• Accurate recognition of rehabilitation exercises is crucial for effective patient monitoring and therapy.
• Existing methods often struggle with the complexity and variability of human movement data.

Purpose of the Study:

• To propose a novel multipath convolutional neural network (MP-CNN) for enhanced rehabilitation exercise recognition using sensor data.
• To develop a specialized evaluation matrix for assessing the quality of rehabilitation exercises.

Main Methods:

• The proposed MP-CNN integrates a dynamic convolutional neural network (D-CNN) using Gaussian mixture models (GMMs) and a state transition probability CNN (S-CNN) employing a modified Lempel-Ziv-Welch (LZW) algorithm.
• Sensor data is segmented into multiple paths within the D-CNN, and transition probabilities of hidden states are extracted as features by the S-CNN.
• A deep learning classifier and a custom evaluation matrix are used to learn general feature representations and score exercise performance.

Main Results:

• The MP-CNN achieved superior classification accuracy compared to other deep learning models on collected and public activity recognition datasets.
• The proposed evaluation scores proved effective for practical rehabilitation applications, enabling quantitative assessment of exercise quality.

Conclusions:

• The MP-CNN offers a robust and accurate solution for sensor-based rehabilitation exercise recognition.
• The developed evaluation framework provides a valuable tool for objective assessment in physical therapy and rehabilitation.