HandEEGnet: A Kolmogorov-Arnold Network for Hand Movements Classification from Motor Imagery EEG

Motor imagery-based BCI plays a vital role in controlling hand movements by processing and recognizing features from EEG signals associated with the imagination of those movements. In this paper, we present a Kolmogorov-Arnold Networks (KAN) based framework in order to classify the left and right-hand movements along with the absence of movement's intention from the electroencephalogram (EEG) signal. The emergence of KAN is expected to be a breakthrough concept in the artificial intelligence paradigm considering its learnable activation functions mechanism and interpretability opportunities. Preprocessing of the collected raw EEG data has been accomplished with band-pass filtering of EEG components of five well-matched channels and corresponding power spectral density feature extraction. We have used 15282 samples with annotated classes from a publicly available dataset to train, validate, and test the classifier. A comparative analysis of the performance of KAN and conventional Multi-Layer Perceptron (MLP) networks has been conducted in order to investigate the effectiveness of the KAN for this classification task. The proposed framework attains 96.14% classification accuracy for 3057 blind test samples, which is found to be better than the conventional MLP neural network approach (76.64% accuracy) and compatible with the existing techniques. The finding of this research might be useful in hand movement-related applications for robotics, prosthetics, and neuro-rehabilitation supports.