An Attention-Residual Convolutional Network for Real-Time Seizure Classification on Edge Devices

Epilepsy affects over 50 million people globally, with accurate seizure type classification directly influencing treatment selection as different seizure types respond to specific antiepileptic medications. Manual electroencephalogram (EEG) interpretation remains time-intensive and requires specialized expertise, creating clinical workflow bottlenecks. This work presents EEG-ARCNet, an attention-residual convolutional network integrating residual connections with channel attention mechanisms to extract discriminative temporal and spectral features from multi-channel EEG recordings. The model combines nine statistical temporal features with five frequency-band power measures through Welch's spectral decomposition, processed through attention-enhanced convolutional pathways. Evaluated on the Temple University Hospital Seizure Corpus, EEG-ARCNet achieved 99.65% accuracy with 99.59% macro-averaged F1-score across five seizure types (absence, focal non-specific, simple partial, tonic-clonic, and tonic). To validate practical deployment, the model was implemented on Raspberry Pi 4, achieving a 2.06 ms average inference time per 10 s segment with 35.4% CPU utilization and 499.4 MB memory consumption. The combination of high classification accuracy and efficient edge deployment demonstrates technical feasibility for resource-constrained seizure-monitoring applications.