Seizure Detection Framework via Multisubject Dynamic Adaptation and Structural Clustering

• Published in: IEEE transactions on instrumentation and measurement Vol. 74; pp. 1 - 12
• Main Authors: Cui, Xiaonan, Hu, Dinghan, Lai, Xiaoping, Jiang, Tiejia, Gao, Feng, Cao, Jiuwen
• Format: Journal Article
• Language: English
• Published: New York IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation; Adaptation models; Algorithms; Alignment; Attention; Brain modeling; Centroids; Clustering; Convolution; Datasets; Electroencephalography; Epilepsy; Feature extraction; Kernel; Modules; Multisubject dynamic transfer; Optimization; seizure detection; structurally regularized clustering; Time-frequency analysis; Training; unsupervised domain adaptation
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2025.3551437

Intersubject variation seriously affects the generalization ability of seizure detection models. Most current models need to be calibrated and trained with annotated data before application, making them strongly dependent on subject-specific features and difficult to directly generalize on new subjects. To overcome this limitation, we propose a multisubject dynamic adaptation and structural clustering (SCMDA) framework to perform offline seizure detection tasks. First, the backbone network is designed as a combination of the temporal encoder and multiple dynamic attention transfer (DAT) modules, where DAT is a parallel structure of squeeze-and-excitation (SE) residual and dynamic residual transfer (DRT). The designed DAT module can enhance the discriminability of the latent space and blur the distribution boundaries between source subjects to reduce the negative impact of domain information on distribution alignment. Then, the model is optimized by jointly discriminative feature alignment of the latent space and structurally regularized clustering of the target domain. The cluster centroids are generated by learning the self-attention feature interaction of the target data in a feedforward manner. Finally, to evaluate the effectiveness of SCMDA, we conduct extensive tests on the public available TUH dataset and the Children's Hospital, Zhejiang University School of Medicine (CHZU) dataset. The proposed method achieves 93.42% and 91.23% cross-subject classification accuracy on the TUH and CHZU datasets, outperforming the current state-of-the-art offline algorithms.