Skip to content
Conference Open access

A Lightweight Cross-Attention Fusion Network for EEG-based Seizure Detection with Unbiased Cross-Patient Evaluation

Sep 2026 · 2026 10th International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC) · pp. 259-265 · 0 citations · 22 references

Abstract

During routine monitoring, automated electroencephalogram analysis could reduce the time required by clinicians to recognize epileptic seizures. In this paper, we demonstrate Multi ViT Fusion, a compact network to combine 3 complementary feature streams from every scalp window. Short-time spectrogram, 12 waveform statistics and 5 clinical band powers are mapped into a common latent space. Then after each attention block, a single attention block with 4 heads shares information among the 3 stream tokens before classification. The model has only 217666 trainable parameters and takes less than 3 minutes to train on the laptop processor. Evaluation is performed on 2 publicly available corpora, with the same segment-wise split as well as a leakage-free split respecting recording. On the saturated Bonn corpus, the results of both protocols differ by only 0.18%, and the leakage at the segment level has little impact on the score. When the subject wise testing is done on the clinical CHB-MIT corpus, the values of each of the classifiers are reduced by 7% to 10% from their optimistic values. The proposed model achieves the highest mean accuracy of 85.94% under strict cross-patient testing while having the lowest variance among patients. Simple reference classifiers remain competitive, and the honest cross-patient gap instead of a saturated number is reported. This work contributes to the achievement of United Nations Sustainable Development Goal 3 (Good Health and Well-being) in under-served areas.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.