A hybrid deep transfer learning-based approach for Parkinson's disease classification in surface electromyography signals

Abstract

Parkinson's disease (PD) is known as a rampant neurodegenerative disorder, which has afflicted approximately 10 million people throughout the world. Surface Electromyography (sEMG) signal trials received from the upper extremities, such as the arm and wrist, would be an efficient way to assess neuromuscular function in the detection of PD. This paper mainly aimed to utilize pre-trained deep transfer learning (DTL) structures and conventional machine learning (ML) models as an automated approach to diagnose PD from sEMG signals. Primarily, we stacked the extracted features from three deep pre-trained architectures, including AlexNet, VGG-f, and CaffeNet, to generate the discriminative feature vectors. Although the number of stacked features from all the three deep structures was large, the proper features is effective in overcoming the challenge of over-fitting as well as increasing the robustness to added noise with different levels. Subsequently, we proposed a novel soft combination of subset feature selection methods, including receiver operating characteristic (ROC), entropy, and the signal-to-noise (SNR) procedures, in order to reduce the size of the extracted features. Finally, we utilized the support vector machine (SVM) with radial basis function (RBF) kernel for identifying PD disorder. The experimental results in different analysis frameworks illustrated that the hybrid deep transfer learning-based approach to PD classification could lead to hitting rates higher than 99%. Moreover, it can be of a competitive performance with the state-of-the-art SVM-based pattern even though the suggested model needs minimal processing in feature construction of sEMG signals to PD detection.