Adaptive Multi-Scale Phase-Aware Fusion Network (AMS-PAFN) for EEG Seizure Recognition

Yanting Liang¹Jingyuan Liu²Xinzhou Zhang^1*

• ¹Shenzhen People’s Hospital, shenzhen, China
• ²Wenzhou Medical University, wenzhou, China

Epilepsy is a neurological disorder characterized by sudden, abnormal discharges of neuronal activity in the brain. Electroencephalogram (EEG) analysis serves as the primary technique for identifying epileptic seizures, and accurate seizure detection plays a vital role in clinical diagnosis, therapeutic intervention, and treatment planning for epilepsy. However, traditional detection methods often rely on manual feature extraction, while current deep learning-based approaches continue to face challenges in terms of frequency adaptability, multi-scale feature integration, and phase alignment.This study proposes the Adaptive Multi-Scale Phase-Aware Fusion Network (AMS-PAFN) to address these limitations. The AMS-PAFN integrates three novel modules: (1) a Dynamic Frequency Selection (DFS) module utilizing Gumbel-Softmax for adaptive spectral filtering to enhance seizurerelated frequency bands; (2) a Multi-Scale Feature Extraction (MCFE) module employing hierarchical downsampling and temperature-controlled multi-head attention to capture both macrorhythmic and micro-transient EEG patterns; and (3) a Multi-Scale Phase-Aware Fusion (MCPA) module that aligns temporal features across scales via phase-sensitive weighting. Evaluated on the CHB-MIT dataset, AMS-PAFN achieved state-of-the-art performance: 98.97% accuracy, 99.53% sensitivity, and 95.21% specificity (Subset 1), surpassing STFTormer by 1.58% absolute accuracy gain (97.39%→98.97%) with 2.66% higher specificity (92.55%→95.21%). Ablation studies confirmed each module's necessity, with DFS improving specificity by 6.87% and MCPA enhancing cross-scale synchronization by 5.54%. The framework's adaptability to spectral variability and spatiotemporal dynamics demonstrates robust potential for clinical seizure recognition.