CNN-Autoformer: Automated EEG-Based Seizure Detection and Localization Using Hybrid Deep Learning.

Overview

abstract

Epilepsy is a neurological disorder characterized by transient and recurrent abnormal brain activity, often diagnosed through manual inspection extensive analysis of electroencephalogram (EEG) recordings. However, existing deep-learning seizure detection methods still struggle to model the complex spatiotemporal dynamics of noisy and nonstationary EEG signals, and their ability to localize seizure onset zones remains limited. To address this, we propose a novel hybrid deep learning framework, CNN-Autoformer, that combines the spatial feature extraction capabilities of Convolutional Neural Networks (CNN) with the temporal modeling power of an Autoformer architecture. The CNN captures inter-channel correlations across multi-channel EEG, while the Autoformer applies an auto-correlation mechanism to extract periodic dependencies and decomposes signals into trend and seasonal components for more accurate and interpretable classification. Evaluation on the public CHB-MIT dataset demonstrated segment-based metrics of 98.34% accuracy, 99.46% sensitivity, 97.12% specificity, 97.46% precision, and 0.983 F1-score. Event-based assessment achieved 100% sensitivity with a low false detection rate of 0.21 events/hour. Validation on the proprietary SH-SDU dataset yielded comparable performance (97.51% accuracy, 98.27% sensitivity, 96.81% specificity, 97.49% precision, 0.976 F1-score). Additionally, the model generates seizure-onset heatmaps aligned with expert annotations, enabling seizure localization. These results highlight the model's robustness, interpretability, and potential for real-world clinical integration in seizure detection and localization.