EEG Image Transformer for Harmful Brain Activity Classification

Timely detection of harmful brain activities, such as seizures, lateralized periodic discharges (LPD), and generalized rhythmic delta activity (GRDA), is critical for patient outcomes, especially in neurocritical care. This paper proposes an EEG-DeiT framework to classify these activities using EEG data. By fusing raw EEG images and spectrogram images(akin to the manual inspection process used by clinicians), the proposed framework effectively captures both spatial and temporal characteristics. Extensive preprocessing, including log transformation, standardization, and colormap application, prepares the data for model training. The DeiT-EEG architecture enhances learning efficiency, particularly for small, noisy clinical datasets. Experimental results demonstrate that EEG-DeiT achieves 97.65% accuracy, outperforming traditional Vision Transformers (ViT). The model demonstrates expert-level performance on well-defined patterns and converges 3.8× faster than conventional transformers, reaching optimal performance in just 5 hours when using pre-trained weights.Clinical relevance-EEG-DeiT enables real-time, automated classification of harmful brain activities, improving diagnostic speed and accuracy. Aiding early intervention can help prevent secondary brain injuries(strokes) and support clinical decision-making, ultimately enhancing patient outcomes in critical care settings.