Core Concepts
Unsupervised domain adaptation using physics-informed methods enhances EEG data analysis.
Abstract
This article introduces a novel approach to combining heterogeneous EEG datasets for machine learning tasks. The core message revolves around leveraging physics-informed unsupervised techniques to address challenges in EEG data analysis. The content is structured into sections focusing on the introduction, Riemannian geometry concepts, matching EEG data dimensions, experimental evaluation, and conclusion. Key highlights include:
Proposal of an unsupervised approach leveraging EEG signal physics.
Utilization of field interpolation and Riemannian geometry for domain adaptation.
Comparative analysis against statistical-based approaches and common channel selection.
Evaluation on six public BCI datasets with leave-one-dataset-out validation.
Results showing enhanced classification performance with field interpolation.
Discussion on covariance matrices, transfer learning, and dimensionality mismatch solutions.
Stats
"Combining electroencephalogram (EEG) datasets for supervised machine learning (ML) is challenging due to session, subject, and device variability."
"Numerical experiments show that in the presence of few shared channels in train and test, the field interpolation consistently outperforms other methods."