{"title":"Efficient Seizure Detection by Complementary Integration of Convolutional Neural Network and Vision Transformer.","authors":"Jiaqi Wang, Haotian Li, Chuanyu Li, Weisen Lu, Haozhou Cui, Xiangwen Zhong, Shuhao Ren, Zhida Shang, Weidong Zhou","doi":"10.1142/S0129065725500236","DOIUrl":null,"url":null,"abstract":"<p><p>Epilepsy, as a prevalent neurological disorder, is characterized by its high incidence, sudden onset, and recurrent nature. The development of an accurate and real-time automatic seizure detection system is crucial for assisting clinicians in making precise diagnoses and providing timely treatment for epilepsy. However, conventional automatic seizure detection methods often face limitations in simultaneously capturing both local features and long-range correlations inherent in EEG signals, which constrains the accuracy of these existing detection systems. To address this challenge, we propose a novel end-to-end seizure detection framework, named CNN-ViT, which complementarily integrates a Convolutional Neural Network (CNN) for capturing local inductive bias of EEG and Vision Transformer (ViT) for further mining their long-range dependency. Initially, raw electroencephalogram (EEG) signals are filtered and segmented and then sent into the CNN-ViT model to learn their local and global feature representations and identify the seizure patterns. Meanwhile, we adopt a global max-pooling strategy to reduce the scale of the CNN-ViT model and make it focus on the most discriminative features. Given the occurrence of diverse artifacts in long-term EEG recordings, we further employ post-processing techniques to improve the seizure detection performance. The proposed CNN-ViT model, when evaluated using the publicly accessible CHB-MIT EEG dataset, reveals its outstanding performance with a sensitivity of 99.34% at a segment-based level and 99.70% at an event-based level. On the SH-SDU dataset we collected, our method yielded a segment-based sensitivity of 99.86%, specificity of 94.33%, and accuracy of 94.40%, along with an event-based sensitivity of 100%. The total processing time for 1[Formula: see text]h EEG data was only 3.07[Formula: see text]s. These exceptional results demonstrate the potential of our method as a reference for clinical real-time seizure detection applications.</p>","PeriodicalId":94052,"journal":{"name":"International journal of neural systems","volume":" ","pages":"2550023"},"PeriodicalIF":0.0000,"publicationDate":"2025-03-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"International journal of neural systems","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1142/S0129065725500236","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 0
Abstract
Epilepsy, as a prevalent neurological disorder, is characterized by its high incidence, sudden onset, and recurrent nature. The development of an accurate and real-time automatic seizure detection system is crucial for assisting clinicians in making precise diagnoses and providing timely treatment for epilepsy. However, conventional automatic seizure detection methods often face limitations in simultaneously capturing both local features and long-range correlations inherent in EEG signals, which constrains the accuracy of these existing detection systems. To address this challenge, we propose a novel end-to-end seizure detection framework, named CNN-ViT, which complementarily integrates a Convolutional Neural Network (CNN) for capturing local inductive bias of EEG and Vision Transformer (ViT) for further mining their long-range dependency. Initially, raw electroencephalogram (EEG) signals are filtered and segmented and then sent into the CNN-ViT model to learn their local and global feature representations and identify the seizure patterns. Meanwhile, we adopt a global max-pooling strategy to reduce the scale of the CNN-ViT model and make it focus on the most discriminative features. Given the occurrence of diverse artifacts in long-term EEG recordings, we further employ post-processing techniques to improve the seizure detection performance. The proposed CNN-ViT model, when evaluated using the publicly accessible CHB-MIT EEG dataset, reveals its outstanding performance with a sensitivity of 99.34% at a segment-based level and 99.70% at an event-based level. On the SH-SDU dataset we collected, our method yielded a segment-based sensitivity of 99.86%, specificity of 94.33%, and accuracy of 94.40%, along with an event-based sensitivity of 100%. The total processing time for 1[Formula: see text]h EEG data was only 3.07[Formula: see text]s. These exceptional results demonstrate the potential of our method as a reference for clinical real-time seizure detection applications.