A multi-feature fusion model with temporal convolution and vision transformer for epileptic seizure prediction

Abstract

Epilepsy is a disease that affects the brain’s nervous system and is characterized by sudden onset, recurrence, and intractability. Epilepsy seizure prediction through electroencephalogram (EEG) signals and early intervention can greatly improve the quality of life of patients. However, recent seizure prediction methods based on deep learning commonly extract only the temporal feature of EEG signals, which disregard the global feature of EEG signals from all of channels. Besides, appropriate fusion strategy of different features is usually ignored in existing methods. To overcome above issues, we propose a multi-feature fusion model with Temporal Convolution and Vision Transformer (TConv-ViT) for epileptic seizure prediction. Specifically, we first use Wavelet Convolution (WaveConv) and Short-Time Fourier transform (STFT) to extract different EEG features. Then we calculate each channel’s attention and put the weighted features into temporal CNN and vision transformer separately to further extract the local and global features. We also develop a feature coupling unit to guide the two branch’s features flow to each other, and obtain better feature representations. On CHB-MIT dataset, our method achieves a sensitivity of 94.2%, a specificity of 99.7% and our false prediction rate is less than 0.007. We also validate the method on Xuanwu Hospital intracranial EEG dataset and get a sensitivity of 93% on average for three different experimental setups. Experimental results show that compared with the existing methods, the proposed method has a high predictive performance and a low false positive rate, which provides a feasible scheme for the clinical application of EEG-based seizure prediction.