Utilizing Contrastive Predictive Coding for the Detection of Epilepsy in EEG Data

Zihang Wang, Guangxian Zhu, Haohui Jia, Kenta T. Suzuki, Dafang Zhao, Naoaki Ono, Md. Altaf-Ul-Amin, Shigehiko Kanaya

Abstract: The difficulty in predicting epileptic seizures in the early stages stems from the absence of clear precursors to these episodes. Therefore, the capacity to identify the impending onset of epileptic seizures accurately at this early stage is crucial. To improve early-stage seizure identification, our study employs a Contrastive Predictive Coding (CPC) framework to identify key EEG regions. CPC uses an autoregressive model predicting future occurrences in latent space, improving data representation learning. We introduce a probabilistic contrastive loss, capturing information vital for future sample prediction, and positive sampling to enhance efficiency. Testing on the CHB-MIT EEG database validates CPC's efficacy in extracting useful representations and seizure pre-warning. This effective framework paves the way for advanced structures for precise epilepsy-related.