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

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Longitudinal adaptation in electroencephalography (EEG) motor imagery (MI) decoding is crucial for brain-computer interfaces (BCIs).
• Current adaptation strategies are often limited to single-subject settings, hindering generalizability.
• Developing robust, long-term decoding requires effective continual fine-tuning and online adaptation methods.

Purpose of the Study:

• To investigate continual fine-tuning strategies for deep learning in online longitudinal EEG-MI decoding across a large user group.
• To evaluate the impact of different fine-tuning approaches on decoder performance and stability.
• To integrate online test-time adaptation (OTTA) to complement fine-tuning and enable calibration-free operation.

Main Methods:

• Explored various continual fine-tuning techniques for deep learning models in a causal EEG-MI decoding setting.
• Implemented online test-time adaptation (OTTA) to dynamically adjust models during deployment.
• Assessed performance and stability across multiple sessions and a large participant cohort.

Main Results:

• Successive fine-tuning leveraging prior subject-specific information significantly enhanced decoder performance and stability.
• Online test-time adaptation (OTTA) effectively adapted models to evolving data distributions across sessions.
• The combined approach enabled stable, calibration-free longitudinal motor imagery decoding.

Conclusions:

• Combining continual fine-tuning with online test-time adaptation is a promising strategy for robust longitudinal EEG-MI decoding.
• These findings provide recommendations for improving brain-computer interface performance in real-world applications.
• The developed methods are critical for advancing neurorehabilitation and assistive technologies.