Classification of Hand-Grasp Movements of Stroke Patients using EEG Data

Electroencephalography (EEG) based Brain Controlled Prosthetics can potentially improve the lives of people with movement disorders, however, the successful classification of the brain thoughts into correct intended movement is still a challenge. In recent years, machine learning based methods, especially deep neural networks, have improved the pattern recognition and classification performance of computer vision systems. However, there is a need to evaluate the classical EEG signal processing algorithms against advanced machine learning based variants, specifically in the domain of Brain Computer Interfaces (BCI). This study aims to be a benchmark where we evaluate the performance of 5 popular motor imagery BCI pipelines and compare classical signal processing techniques (wavelet transform and power spectral density) with motor imagery specific algorithms (Common Spatial Patterns and Filter Bank Common Spatial Patterns (FBCSP)) and a state-of-the-art deep neural network based EEGNet algorithm. The experiments were performed on an open-source EEG dataset of hemiparetic stroke patients and both within subject and cross subject performance of the aforementioned algorithms was evaluated based on kappa scores. We empirically found that, for within subject classification, FBCSP method still is the gold-standard for motor imagery task with a mean kappa score of about 0.70 (84.8% accuracy) while for cross subject classification, due to the availability of a large amount of data, deep learning based EEGNet method outperformed all the other methods with a large margin and gave kappa value of 0.54 (77.0% accuracy). We believe these results would help BCI researchers to select a suitable BCI pipeline for their task that could help in the development of robot assisted therapies or as an interface for assistive devices.