3-Channel Motor Imagery Classification using Conventional Classifiers and Deep Learning Models

Abstract

Brain-computer interfaces (BCIs) are one of the important applications based on motor imagery classification using EEG signals. BCIs are designed to help patients afflicted with motor disabilities. The purpose of this study is to assess how well various conventional machine learning and deep learning models work for motor imagery task classification from EEG data analyzed by three channels C3, C4, and Cz. A comprehensive methodology employed including preprocessing of raw EEG signals (Time, Frequency, Time-frequency domains) multi-feature extraction followed by classification based on conventional models (decision Tree, SVM, Random Forest) as well as deep learning methodologies like CNN, RNN, and TSFFnet-based architectures. The results indicate that random forest is consistently performed well across different domains. As it achieves high accuracy and the lowest mean absolute error among other conventional classification models. The accuracy of TSFFnet among deep learning models was 99.75%, precision is maximum seems like it has been configured to have a good recall with the values for recall being close to that, and mean absolute error is minimal at 0.0038. These results reveal that deep learning models especially the TSFFnet model outperform in the tasks of motor imagery classification.