Two-Stream Attention 3-D Deep Network-Based Childhood Epilepsy Syndrome Classification

• Published in: IEEE transactions on instrumentation and measurement Vol. 72; pp. 1 - 12
• Main Authors: Cao, Jiuwen, Feng, Yuanmeng, Zheng, Runze, Cui, Xiaonan, Zhao, Weijie, Jiang, Tiejia, Gao, Feng
• Format: Journal Article
• Language: English
• Published: New York IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: 3-D convolution neural network; attention module; childhood epilepsy syndrome; Classification; Convulsions & seizures; Disorders; Electroencephalography; Epilepsy; interictal EEG; Machine learning; Modules; Nervous system; Optimization; Recording; Representation learning; Seizures; Three-dimensional displays; Time-frequency analysis; Transforms
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2022.3220287

Epilepsy syndromes are typical childhood nervous system diseases, usually containing several different seizure types commonly seen. There exist around 20 known childhood epilepsy syndromes and accurate classifying these epilepsy syndromes plays a vital guidance to diagnosis and treatment. Current research mainly focused on seizure detection, while few have been presented on epilepsy syndrome classification. In this article, a novel two-stream 3-D attention module-based deep network (TSA3-D) is developed to exploit multichannel time-frequency and frequency-space features of interictal EEGs for epilepsy syndrome classification. Two multichannel montage transforms are applied for EEG feature optimization to reduce artifact affection. In TSA3-D, both the channelwise attention (CA) and dual spatial attention (DSA)-based 3-D attention modules are developed to enhance the EEG feature learning capability. The study is conducted on the long-term interictal EEGs of 115 subjects from the Children's Hospital, Zhejiang University School of Medicine (CHZU). These data belong to seven typical childhood epilepsy syndromes and one normal control group. Experimental results show that TSA3-D can achieve an overall accuracy of 99.52%, the highest among when compared with the state-of-the-art (SOTA) methods.