ENHANCING CONSISTENCY IN EEG LATENT REPRESENTATIONS: UNSUPERVISED SHARED-PRIVATE DEEP LEARNING ADAPTATION TO MINIMISE INTER-SESSION VARIABILITY

Motor imagery (MI) electroencephalography (EEG) is a commonly-used functional measurement challenged by substantial performance drops across recording sessions, largely due to dynamic physiological fluctuations and electrode shifts. This inter-session variability limits the generalisability of deep learning-based decoders from EEG. Here we systematically applied unsupervised domain adaptation strategies to address these issues in a model-agnostic setting. We conducted a baseline experiment including source-only training, supervised fine-tuning (upper bound), adversarial adaptation (Domain-Adversarial Neural Networks), and discrepancy-based adaptation (Maximum Mean Discrepancy) in MI classification on the LEE multi-session MI-EEG dataset from 54 participants. Furthermore, we implemented a shared-private representation learning framework, decomposing latent features into domain-invariant (shared) and session-specific (private) components. Training involved a source-only warm-up, followed by unsupervised adaptation using distribution alignment and orthogonality constraints, without target labels. Notably, our framework incorporates second-order feature alignment under fully unsupervised cross-session conditions. Model performance was assessed on held-out target sessions using Area Under the Curve (AUC). Unsupervised adaptation improved target-session performance over source-only training (AUC=0.582), with the shared-private model achieving the highest AUC (0.667), significantly outperforming adversarial (0.558, p<0.001) and discrepancy-based (0.604, p<0.01) methods. While supervised fine-tuning yielded slightly higher accuracy, the unsupervised approach closely approached this upper bound and demonstrated a ~10% gain over unseen data. Joint learning also provided more stable cross-session generalisation. In sum, disentangling shared and session-specific EEG representations with second-order alignment offers an effective, stable, and unsupervised solution to reduce inter-session variability in EEG decoding, supporting practical deployment with minimal calibration.