Optimizing motor imagery BCI models with hard trials removal and model refinement

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Deep learning models achieve high accuracy in motor imagery Brain-Computer Interface (BCI) systems.
• Model performance degrades due to sensitivity to challenging or "hard" trials.
• Existing methods lack transparency in identifying these hard trials.

Purpose of the Study:

• To propose and evaluate novel methods for identifying and mitigating the impact of hard trials in motor imagery BCI.
• To enhance the robustness and accuracy of deep learning models in BCI applications.
• To compare a quantitative explainable AI (XAI) approach against a prediction score-based method for hard trial identification.

Main Methods:

• Developed two methods for hard trial identification: one using model prediction scores and another employing quantitative XAI.
• Removed identified hard trials from the training and validation datasets.
• Re-trained deep learning models on datasets excluding hard trials and evaluated performance on the Open BMI dataset.

Main Results:

• The quantitative XAI-based hard trial removal method significantly improved subject-specific motor imagery classification accuracy from 63.77% to 68.70% (p-value = 7.66e-11).
• Analysis of relevance scores on scalp maps provided deeper insights into hard trial characteristics.
• The quantitative XAI approach demonstrated superior performance in hard trial identification compared to the prediction score-based method.

Conclusions:

• Quantitative XAI is an effective and interpretable method for identifying hard trials in motor imagery BCI.
• Removing hard trials identified by XAI significantly enhances deep learning model performance.
• This approach offers a promising direction for improving the reliability of BCI systems.