Fine-Tuning Strategies for Continual Online EEG Motor Imagery Decoding: Insights from a Large-Scale Longitudinal Study

TL;DR

The paper addresses the challenge of continual online EEG motor imagery decoding across many users and multiple sessions by evaluating how offline fine-tuning and online test-time adaptation interact under a causal deployment. Using a large longitudinal dataset (61 subjects, 7–11 sessions per subject) and a BaseNet with Real-Time Adaptive Pooling, it compares exemplar-free and joint data settings with independent or sequential fine-tuning, augmented by online Euclidean alignment and AdaBN. The most effective approach is joint sequential fine-tuning, which leverages all prior subject-specific data and preserves a progressive adaptation trajectory, and is complemented by OTTA to achieve calibration-free operation across sessions. The study demonstrates improved performance and stability, offering practical guidelines for designing robust, real-world longitudinal MI decoding systems with potential neurorehabilitation and assistive-tech impact.

Abstract

This study investigates continual fine-tuning strategies for deep learning in online longitudinal electroencephalography (EEG) motor imagery (MI) decoding within a causal setting involving a large user group and multiple sessions per participant. We are the first to explore such strategies across a large user group, as longitudinal adaptation is typically studied in the single-subject setting with a single adaptation strategy, which limits the ability to generalize findings. First, we examine the impact of different fine-tuning approaches on decoder performance and stability. Building on this, we integrate online test-time adaptation (OTTA) to adapt the model during deployment, complementing the effects of prior fine-tuning. Our findings demonstrate that fine-tuning that successively builds on prior subject-specific information improves both performance and stability, while OTTA effectively adapts the model to evolving data distributions across consecutive sessions, enabling calibration-free operation. These results offer valuable insights and recommendations for future research in longitudinal online MI decoding and highlight the importance of combining domain adaptation strategies for improving BCI performance in real-world applications. Clinical Relevance: Our investigation enables more stable and efficient long-term motor imagery decoding, which is critical for neurorehabilitation and assistive technologies.

Figures (5)

• Figure 1: Fine-tuning process and task vectors across different settings in a two-dimensional weight space. Solid arrows represent task vectors, dashed lines illustrate the fine-tuning trajectory.
• Figure 2: Accuracy across the sessions for each fine-tuning strategy compared to the source model and the baseline. Each dot represents the test accuracy of a single session, averaged across all subjects. The x-axis denotes session progression, while the y-axis represents the test accuracy (%).
• Figure 3: Average test accuracy over all sessions (2 - 11) per setting. Stars above the brackets indicate a significance level ($p<0.001$ (***)) when compared to the baseline and source model, respectively.
• Figure 4: Average test accuracy (over all subjects and both paradigms) for each fine-tuning setting. The X on the x-axis refers to no fine-tuning, i.e., using the source model.
• Figure 5: Cosine distance between task vectors for all strategies.