Improving Multi-Class Motor Imagery Classification Using EEG Signals and a Hybrid Deep Learning Model in Brain–Computer Interface Systems for Patients with Motor Disabilities

Abstract

EEG (Electroencephalography)-based BCI (brain-computer interface) systems are widely used to help patients with motor disabilities. Motor imagery (MI) is a major BCI paradigm since it allows the user to imagine the movement of limbs without its muscular execution. Nonetheless, the multi-class MI-EEG classification remains challenging because EEG signals are nonlinear and non-stationary in nature.  This research proposes a model for four-class MI-EEG classification using a compact hybrid CNN-BiLSTM-attention in compact form. The framework integrates spatial CNN filtering, bidirectional temporal sequence modeling, and attention-based temporal weighting within a single end-to-end model. The BCI Competition IV Dataset 2a was used to evaluate the proposed model and compared it with CSP-LDA, SVM, CNN and CNN-LSTM baselines. Based on the experimental results, the proposed model achieved the best overall performance with a mean accuracy of 83.89% ±4.20% and Cohen’s kappa of 0.7077 ±0.0635. The results of the study show that bidirectional temporal modeling and attention weighting can enhance the robustness and discriminative ability of EEG-based MI classification for assistive BCI applications.