Brain Activity Classification using Retentive Networks with Explainable AI

Harmful brain activity can significantly impact an individual’s life, leading to debilitating seizures, epileptic episodes, and long-term cognitive impairments that affect daily functioning and overall quality of life. These abnormalities can be effectively detected using electroencephalography (EEG) image data, which captures the underlying neural activity of the brain in a non-invasive manner. In this study, Retentive Networks (RetNets) are employed to classify EEG spectrogram images for identifying different types of harmful brain activity. While deep learning architectures such as EfficientNet and Vision Transformers have demonstrated strong performance in image-based classification tasks, RetNets offer a compelling advantage in terms of reduced computational complexity and efficient sequence modelling. The dataset used in this work is a publicly available EEG dataset comprising approximately 17,000 spectrogram images derived from EEG recordings, ensuring sufficient diversity and representation for robust model training and evaluation. Experimental evaluation demonstrates that the proposed RetNet model achieved the highest overall performance, with a precision of 94%, recall of 100%, and an F1-score of 97%, indicating balanced and highly reliable classification capability. In comparison, EfficientNet achieved a precision of 91%, recall of 90%, and an F1-score of 96%, while the Vision Transformer achieved precision, recall, and F1-scores of 94%, 86%, and 90%, respectively. Furthermore, the proposed approach integrates model interpretability through Explainable Artificial Intelligence (XAI) techniques. The novelty of the proposed work lies in combining RetNet-based EEG spectrogram classification with XAI-driven interpretability for harmful brain activity detection.