Motor Imagery EEG Decoding for Brain-Computer Interfaces: Structured Representation, Transfer, and Drift

Abstract

Motor imagery EEG decoding is often summarized by the accuracy of a final classifier, but the classifier is only the last stage of the pipeline. Before classification, the signal has already been shaped by preprocessing, feature extraction, source-session organization, and adaptation. This dissertation studies how feature representation, source-session transfer, and drift shape reliable MI-EEG decoding for brain-computer interfaces.

It first studies within-session decoding on public MI-EEG datasets using nested validation that keeps preprocessing, feature fitting, and model selection inside the training folds. This analysis separates gains from feature representation from gains due to nonlinear classification, and relates both comparisons to latency, operation-count proxies, and model size. Structured feature representations, including spatial-filter and covariance-based features, improve over simple log-variance summaries in several settings, while the strongest feature family varies across datasets. At fixed feature extraction, the nonlinear classifier comparison does not support replacing the selected linear-family classifier once practical margins and deployment costs are considered.

The analysis then moves to repeated-session settings. When target-session labels are unavailable, historical sessions are not automatically interchangeable sources. The transfer study compares full pooling, distance-weighted pooling, and confidence-gated source selection under a matched feature, classifier, and adaptation grid. The main benefit of selective source use is not a uniform accuracy improvement, but the ability to keep selected-configuration performance close to strong pooling baselines while reducing source use and computation in the larger longitudinal dataset. The drift study then measures pre-adaptation feature-space drift from a fixed reference session and compares it with fixed-model loss, domain-adaptation gain, and retraining gain from target-session labels. Drift is informative in some longitudinal common spatial pattern (CSP) settings, where larger measured drift is associated with larger fixed-model loss, but its interpretation depends on the dataset and feature setting. Domain adaptation improves average accuracy in the main longitudinal settings, but it does not remove the value of target-session labels.

The final research chapter extends the representation theme through tensor multiple canonical correlation analysis (TMCCA) for structured multiview tensor data. It formulates rank-1 canonical weight tensors, develops computational procedures for the resulting nonconvex problem, and gives statistical interpretations and recovery results. In this dissertation, TMCCA provides the methodological part of the representation theme by extending the analysis from MI-EEG pipelines to structured multiview data.

Taken together, the chapters show that reliable MI-EEG decoding is better understood as a sequence of representation, transfer, drift measurement, adaptation, and recalibration choices than as a classifier comparison alone.

Chapters and companion software

• Chapter 1. Allocating Model Complexity in Within-Session MI-EEG Decoding: A Cost-Aware Comparison Under Strict Nested Validation. Related: research track · BCIFeatR
• Chapter 2. A Confidence-Gated Source Selection Strategy for Cross-Session Transfer in Brain-Computer Interfaces. Accepted for publication in Frontiers in Human Neuroscience - Brain-Computer Interfaces, Aug 2026. Related: research track · CrossDA
• Chapter 3. Drift Diagnostics, Adaptation, and Recalibration in Multi-Session Motor-Imagery EEG. Manuscript under review. Related: research track · ShiftDx
• Chapter 4. Tensor Multiple Canonical Correlation Analysis for Structured Multiview Data. Related: research track · tensorMCCA