Multivariate Time Series Representation Learning for Electrophysiology Classification with Procedure State Cross-domain Embedding

Electrophysiological time series data often exhibit complex dynamic patterns, strong inter-channel correlations and various non-stationary characteristics accompanied by significant noise interference. How to accurately and robustly extract deep representations of electrophysiological time series to improve the performance of downstream analytical tasks has become an interesting problem in the field of bioinformatics. In this paper, we present a novel framework named TimPSC for learning electrophysiological time series representation, which offers a procedure state cross-domain representation strategy that captures the dynamic procedure features during the denoising process. The contrastive representation learning model is proposed to leverage the diffusion process, extracting conditional features from Cross-domain Space Structure (CSS) and procedure features from the diffusion process. It also employs a Cross-domain Conditional Diffusion Model (CCDM) with cross-domain constraints in 2D space for efficient denoising. The experiments are conducted on electrophysiological datasets of UEA, our framework exhibits significant higher performance over the state-of-the-art self-supervised learning techniques. The TimePSC framework promises to provide a fresh perspective, indicating that embracing the procedure state embedding strategy has the potential to enhance the model’s capability in discerning meaningful features during Multivariate Time Series (MTS) representation learning.