Multiscale activity recognition algorithms to improve cross-subjects performance resilience in rehabilitation monitoring systems

Abstract

Background and Objective:

This study introduces multiscale feature learning to develop more robust and resilient activity recognition algorithms, aimed at accurately tracking and quantifying rehabilitation exercises while minimizing performance disparities across subjects with varying motion-related characteristics.

Methods:

Advanced architectures designed to process multi-channel time series data using two parallel branches that extract features at different scales were developed and tested.

Results:

The results indicate that multiscale algorithms consistently outperform traditional approaches, demonstrating enhanced performance, particularly among patient subjects. Specifically, the multiscale tCNN and multiscale CNN-LSTM achieved accuracies of 91% and 90%, respectively, while the multiscale ConvLSTM maintained strong performance at 89%. Notably, the multiscale Transformer emerged as the most effective model, achieving the best average accuracy of 93%.

Conclusions:

This research underscores the need to explore advanced methods for enhancing activity recognition systems in healthcare, where accurate exercise monitoring and evaluation are becoming essential for effective and personalized treatment in telemedicine services.