Adaptive Convolutional Neural Networks for Enhanced EEG Motor Imagery Task Classification

Recent advances highlight the importance of electroencephalogram (EEG) based motor imagery (MI) for brain-computer interface (BCI) applications, especially in improving the quality of life of people with disabilities. However, limited decoding, high computational need, subject invariability and inaccurate classification constrain the widespread growth of the brain imaging domain. This research introduces an adaptive convolutional neural network (ACNN) model for the classification of MI tasks specifically left-hand, right-hand tasks and foot without the need for extensive preprocessing and feature extraction techniques. The proposed model automatically adjusts the filters and parameters during training, which enables it to optimize the extraction of relevant features. This architecture leverages spatial and spectral convolution, along with the use of adaptive kernels, to effectively capture both the spatial and temporal characteristics of MI-EEG signals A comparison is made between the proposed $A C N N$ model and various state of art models which ensure higher performance of model with low computation demand. Using BCI Competition IV Dataset 1, the proposed ACNN achieved a classification accuracy of 97.3\%, outperforming other models such as CNN-LSTM, BLSAN, and MBGA-Net, thereby establishing a new state-of-the-art in the field.