Machine Learning-Based ALS Detection Using HOG Features from sEMG Time-Frequency Images

Amyotrophic lateral sclerosis (ALS) is a degenerative disorder affecting motor neurons, resulting in muscular deterioration and eventual paralysis. Surface electromyography (sEMG) signals offer valuable neuromuscular data that can assist in identifying ALS. Our research introduces an innovative method utilizing time-frequency representations (TFRs) for classifying ALS. We used continuous wavelet transform (CWT) and wavelet synchrosqueezed transform (WSST) to create TFR-based images from sEMG signals. We then applied histogram of oriented gradients (HOG) to extract distinctive features from these images, which were subsequently classified using a support vector machine (SVM). Our approach yielded an accuracy of 89.2% with WSST-derived HOG features and 85% with CWT-derived HOG features. Furthermore, WSST exhibited superior performance compared to CWT in terms of precision, recall, F1-score, and AUC-ROC score. In comparison to previous studies, our method demonstrates enhanced classification performance, underscoring the effectiveness of wavelet-based TFRs and HOG feature extraction for detecting ALS. These results contribute to the advancement of automated ALS diagnosis, potentially enhancing clinical evaluations and facilitating early intervention.