LG-TriCapsNet: A lightweight graph capsule framework with nearest neighbor graphs for multi-disease EEG classification.

EEG signal classification for neurological disorders is a very critical task in the healthcare field, demanding accuracy and efficiency. Due to the diversity of these disorders and the complexity of the EEG signals, the task of diagnosing these disorders is very challenging. Traditional methods usually cannot capture the intricate relationships between temporal and spatial features, which are inherent in EEG data. To overcome this challenge, the proposed Lightweight Graph Triplet Capsule Networks combined with Nearest Neighbor Graphs effectively leverage both temporal and spatial information in EEG signals. The graph-based representation enhances the spread of information across the network, which thus performs efficient feature extraction and good classification performance. By integrating NNG, the model dynamically captures spatial-temporal dependencies between EEG channels, significantly improving feature discrimination across neurological conditions. The novelty of the approach lies in the combination of graph structures and capsule networks, which particularly makes it suited for handling the complexities of EEG data. LG-TriCapsNet is performing quite well as the F1 score is 98.32 %, accuracy is 98.34 %, sensitivity is 98.30 %, and specificity is 98.40 %, and it also outperforms the current state-of-the-art models, which points to the fact that this framework is really effective for the classification of such a variety of neurological disorders. LG-TriCapsNet, by providing a computationally efficient, real-time solution, holds great potential for advancing clinical decision-making and patient care by offering a robust tool for automated neurological disease detection and diagnosis. At https://github.com/jainshraddha12/LG-TriCapsNet, the source code will be available to the public.