MFTSGNet: A Dual-Branch Spatio-Temporal Network for Robust Seizure Detection

Zhiheng Zhang, XiangLi, Zhaonian Guo, Xin Wang, Ke Zhang, Shixiong Chen

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Epileptic seizure detection, EEG analysis, Wavelet Transform, TCN, Graph Learning

Abstract: Epileptic seizure detection from EEG signals remains a challenging task due to the complex spatial-temporal dependencies in neural activities. In this paper, we propose a novel model named Multi-Frequency TCN Spatial Graph Network(MFTSGNet), which incorporates a dual-branch architecture combining Wavelet Transform with Temporal Convolution (WT+TCN) for time-frequency analysis and a Spatial Attention with Top-k pooling module for spatial feature extraction. The fused features are further processed using a Graph Convolutional LSTM (GC-LSTM) to capture inter-window dynamics. Evaluated on CHB-MIT and Siena datasets using 10-fold cross-validation, our MFTSGNet model achieved 99.24% accuracy on the CHB-MIT dataset and 99.33% accuracy on the Siena Dataset, outperforming existing methods. For further validation, we test the performance using 10-fold cross-validation and Leave-one-patient-out-cross-validation(LOPOCV) on both CHB-MIT and Siena Datasets, showing high accuracy and stability under various validation schemes. These results demonstrate the model’s strong potential for reliable and accurate automated seizure detection in clinical settings.

Track: 7. General Track

Registration Id: Y6NTJB48SZH

Submission Number: 65