Research Information System
Artificial Intelligence Studies, vol.8, no.2, pp.148-157, 2025 (Peer-Reviewed Journal)
The similarity of clinical symptoms in neuromuscular diseases, particularly myopathy and neuropathy, significantly complicates diagnosis based on superficial electromyography (sEMG) signals. While deep learning approaches in the literature offer high success rates, their low explainability limits clinical confidence. This study aims to develop a diagnostic system with both high accuracy and interpretability using handcrafted features extracted from sEMG signals and hybrid artificial intelligence models. The study utilized a dataset of 241 participants with recordings from the Biceps Brachii and Deltoid muscles, encompassing healthy, myopathy, and neuropathy classes. Forty-three features encompassing time, frequency, and nonlinear dynamics were extracted from the signals, and the 20 most distinctive features were identified using the Mutual Information method. SVM, Random Forest, Deep-MLP, and a Stacking architecture combining these models were used in the classification phase. Experimental results showed that the Stacking model exhibited the best performance with 80.26% accuracy in the three-class classification. In binary distinctions, a 93.17% success rate was achieved in the "Healthy-Neuropathy" classification. Furthermore, in the most difficult-to-distinct pair, "Myopathy-Neuropathy," the Deep-MLP model successfully modelled the heterogeneous nature of myopathic signals with up to 91% accuracy. These findings demonstrate that multidimensional feature sets and ensemble learning methods offer a non-invasive, reliable, and clinically interpretable solution for the early diagnosis of neuromuscular diseases.