This is the authors' abstract. We don't add key points for this paper.
Automated epileptic seizure classification from electroencephalogram (EEG) recordings can assist the analysis of large volumes of neurological data and reduce the burden associated with manual interpretation. However, distinguishing EEG patterns across different seizure-related and non-seizure classes remains challenging because relevant information occurs at multiple temporal scales and may involve both local and long-range dependencies. To address these challenges, we propose STCA-Net, a compact spatiotemporal neural architecture that combines multi-scale convolutional feature extraction with cross-attention-based feature fusion. The network consists of four principal components: a Depthwise Separable Convolutional Attention (DSCA) module with multiple kernel sizes for extracting local temporal patterns, a Temporal Pyramid Pooling (TPP) module for multi-scale contextual aggregation, a Compact Temporal Encoder (CTE) for modelling long-range temporal dependencies with reduced computational cost, and a Bidirectional Cross-Attention Fusion (CAF) module for exchanging information between spatial and temporal representations. Experiments are conducted on the University of Bonn EEG dataset using stratified cross-validation for binary, three-class, and five-class classification tasks. STCA-Net achieves an accuracy of 97.58% ± 0.77% for binary classification, 93.09% ± 1.31% for three-class classification, and 75.64% ± 1.27% for five-class classification, with a corresponding five-class F1-score of 75.50% ± 1.26%. In the five-class setting, the proposed method improves accuracy and F1-score over the MASF baseline by 3.14 and 2.88 percentage points, respectively. Ablation experiments further demonstrate the contribution of bidirectional cross-attention compared with direct feature concatenation. With approximately 46,885 trainable parameters, STCA-Net provides a compact architecture for EEG classification while maintaining competitive predictive performance.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Bandale et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: