EMG-Spectrogram-Empowered CNN Stroke-Classifier Model Development

TL;DR

This paper introduces a new CNN model using EMG spectrograms to accurately classify stroke patients and healthy individuals, offering a promising tool for home-based rehabilitation.

Contribution

A novel Tri-CCNN model using EMG spectrograms achieves 93.33% classification accuracy, outperforming existing CNN architectures for stroke assessment.

Findings

The Tri-CCNN model achieved 93.33% classification accuracy, the highest among tested models.

Spectrogram amplitude distributions showed distinct patterns in stroke patients, supporting objective assessment.

The method shows potential for automating stroke classification and rehabilitation monitoring in home settings.

Abstract

Stroke is a leading cause of death and long-term disability worldwide, with ischemic stroke accounting for approximately 62.4% of all cases. This condition often results in persistent motor dysfunction, significantly reducing patients’ productivity. The effectiveness of rehabilitation therapy is crucial for post-stroke motor recovery. However, limited access to rehabilitation services particularly in low- and middle-income countries remains a major barrier due to a shortage of experienced professionals. This challenge also affects home-based rehabilitation, an alternative to conventional therapy, which primarily relies on standard evaluation methods that are heavily dependent on expert interpretation. Electromyography (EMG) offers an objective and alternative approach to assessing muscle activity during stroke therapy in home environments. Recent advancements in deep learning (DL) have opened new avenues for automating the classification of EMG data, enabling differentiation between post-stroke patients and healthy individuals. This study introduces a novel methodology for transforming EMG signals into time–frequency representation (TFR) spectrograms, which serve as input for a convolutional neural network (CNN) model. The proposed Tri-CCNN model achieved the highest classification accuracy of 93.33%, outperforming both the Shallow CNN and the classic LeNet-5 architecture. Furthermore, an in-depth analysis of spectrogram amplitude distributions revealed distinct patterns in stroke patients, demonstrating the method’s potential for objective stroke assessment. These findings suggest that the proposed approach could serve as an effective tool for enhancing stroke classification and rehabilitation procedures, with significant implications for automating rehabilitation monitoring in home-based rehabilitation (HBR) settings.

Figures