Integrating dynamic stopping, transfer learning and language models in an adaptive zero-training ERP speller

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCIs) typically require extensive calibration sessions for decoder training.
• Zero-training methods are emerging as a way to reduce setup time and improve user experience.

Purpose of the Study:

• To propose a probabilistic framework for zero-training BCIs that leverages event-related potentials (ERPs).
• To demonstrate the framework's effectiveness using a visual P300 speller paradigm.
• To investigate the contributions of transfer learning, unsupervised adaptation, language models, and dynamic stopping.

Main Methods:

• A probabilistic zero-training framework was developed for BCI applications using ERPs.
• A simulation study compared the proposed framework against a state-of-the-art supervised model.
• The influence of individual components (transfer learning, unsupervised adaptation, language model, dynamic stopping) was analyzed.

Main Results:

• The zero-training framework achieved performance competitive with supervised methods without requiring calibration.
• Inter-subject transfer learning was a key factor in the framework's high decoding quality.
• Continuous unsupervised adaptation further enhanced performance.

Conclusions:

• High-performing zero-training BCIs are feasible for ERP-based paradigms like P300 spelling.
• Eliminating calibration saves valuable user time, enabling immediate use.
• The framework has potential applications in both clinical and non-clinical settings.