Inter-participant transfer learning with attention based domain adversarial training for P300 detection

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Brain-computer interface (BCI) systems enable communication between the brain and computers.
• Event-related potential (ERP)-based BCIs often require lengthy user training for decoding models.
• Individual variability in EEG signals poses a challenge for deep learning model generalization in BCI applications.

Purpose of the Study:

• To propose a novel One-source domain transfer learning method, Attention Domain Adversarial Neural Network (OADANN).
• To mitigate data distribution discrepancies for cross-participant classification in ERP-based BCIs.
• To enhance the accuracy and robustness of participant-independent BCI models.

Main Methods:

• Developed and implemented the Attention Domain Adversarial Neural Network (OADANN) model.
• Trained and validated OADANN on the OpenBMI dataset and a self-collected dataset.
• Utilized a leave-one-participant-out cross-validation scheme for robust evaluation.

Main Results:

• OADANN achieved the highest and most robust classification performance.
• Demonstrated significant improvements over baseline methods (CNN, EEGNet, ShallowNet, DeepCovNet) and domain generalization techniques (ERM, Mixup, Groupdro).
• Effectively addressed individual differences in EEG signals for improved cross-participant classification.

Conclusions:

• The proposed OADANN method is highly effective for participant-independent ERP classification in BCIs.
• OADANN successfully mitigates data distribution discrepancies across participants.
• This approach offers a promising solution to the challenges of BCI training and generalization.