Spatiotemporal distributionally robust optimization for improved cross-patient EEG seizure analysis

Zongpeng Zhang, Xiang Li, Mingqing Xiao, Haoxuan Li

Keywords: Multivariate Signals, Electroencephalography, Epilepsy Seizure, Distributionally robust optimization

Abstract: Automatic seizure detection and classification from electroencephalography (EEG) hold significant potential to enhance epilepsy diagnosis and treatment. However, deep learning approaches often suffer from limited generalization ability to unseen patients due to inter-patient variability in EEG. While existing studies primarily focus on model architecture design or pre-training strategies to alleviate the problem, the optimization framework for robust cross-patient generalization, especially under the inherently spatiotemporal structure of EEG, remains underexplored. In this work, we propose SpatioTemporal Distributionally Robust Optimization (STDRO), a novel method to improve cross-patient seizure analysis in parallel to existing architectural/pre-training solutions. STDRO constructs and learns structured uncertainty sets that explicitly capture the spatial and temporal characteristics of EEG signals, thereby inducing data-adaptive worst-case distributions for robust optimization and improving cross-patient generalization. Extensive experiments demonstrate the effectiveness of STDRO as a plug-and-play approach to consistently enhance state-of-the-art seizure detection and classification models across diverse evaluation scenarios. Our work advances robust EEG-based seizure analysis toward practical applications with cross-patient scenarios.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 9166