Improving inter-session performance via relevant session-transfer for multi-session motor imagery classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Electroencephalography (EEG)-based brain-computer interfaces (BCIs) enable external device control via motor imagery (MI).
• EEG signal non-stationarity, or inter-session variability, significantly hinders BCI performance across multiple sessions for individual users.
• Existing methods struggle to effectively adapt to these signal changes over time.

Purpose of the Study:

• To introduce and evaluate a novel relevant session-transfer (RST) method to mitigate inter-session variability in multi-session MI classification.
• To enhance the robustness and accuracy of EEG-based BCIs for practical applications.

Main Methods:

• Developed the relevant session-transfer (RST) method, utilizing cosine similarity to transfer pertinent data from prior sessions to the current one.
• Compared RST against a self-calibrating method (current session data only) and a whole-session transfer method (all prior session data).
• Validated methods on a public MI dataset (Shu Dataset) and a custom dataset of gait-related MI from healthy and spinal cord injury participants.

Main Results:

• The RST method demonstrated statistically significant improvements: 2.29% on the Shu Dataset (p < 0.001) and up to 6.37% on the custom dataset compared to the self-calibrating method.
• RST outperformed the previous state-of-the-art method on the Shu Dataset.
• These results confirm the efficacy of RST in enhancing multi-session MI classification performance.

Conclusions:

• The proposed relevant session-transfer (RST) method effectively addresses EEG non-stationarity in motor imagery BCIs.
• RST offers a promising solution for improving the reliability and performance of BCIs across multiple sessions.
• This advancement has significant implications for practical BCI applications requiring sustained user interaction.