T-TIME: Test-Time Information Maximization Ensemble for Plug-and-Play BCIs.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Electroencephalogram (EEG)-based brain-computer interfaces (BCIs) offer direct brain-computer communication.
• BCIs typically require time-consuming, user-unfriendly subject-specific calibration due to EEG signal variability.
• Transfer learning (TL) aims to reduce or eliminate calibration, but existing methods often assume offline data availability.

Purpose of the Study:

• To address the challenge of online transfer learning for EEG-based BCIs.
• To develop a method for immediate classification of EEG data from new users in a streaming context.
• To enable calibration-free, plug-and-play electroencephalogram (EEG)-based brain-computer interfaces (BCIs).

Main Methods:

• Proposes Test-Time Information Maximization Ensemble (T-TIME) for online transfer learning in BCIs.
• Initializes multiple classifiers using source data and updates them with incoming unlabeled EEG trials.
• Employs ensemble learning for prediction and conditional entropy minimization with adaptive marginal distribution regularization for classifier updates.

Main Results:

• T-TIME demonstrated superior performance compared to approximately 20 classical and state-of-the-art TL approaches.
• Experiments were conducted on three public motor imagery-based BCI datasets.
• The proposed method effectively handles the online, streaming nature of EEG data for immediate classification.

Conclusions:

• This work presents the first test-time adaptation method for calibration-free EEG-based BCIs.
• The T-TIME approach facilitates the development of plug-and-play BCIs.
• The study highlights the potential of online transfer learning to overcome calibration barriers in BCI applications.