Fusion with language models improves spelling accuracy for ERP-based brain computer interface spellers.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) utilize electroencephalography (EEG) event-related potentials (ERPs) for text input.
• Current BCI-typewriter systems often lack single-trial accuracy, necessitating multi-trial approaches that reduce speed.

Purpose of the Study:

• To investigate the impact of Bayesian fusion between n-gram language models and ERP detectors on EEG-based BCI performance.
• To enhance both accuracy and speed in BCI-based text input systems.

Main Methods:

• Bayesian fusion of an n-gram language model with a regularized discriminant analysis ERP detector.
• Evaluation of letter classification accuracy across varying language model orders and ERP trial counts.

Main Results:

• Language models significantly improve letter classification accuracy in EEG-based BCIs.
• A 4-gram language model allows for 3-trial classification for initial letters and single-trial for subsequent letters, matching performance levels.
• Fusion of EEG and language model data substantially increases the word rate of BCI typing systems.

Conclusions:

• The integration of language models is crucial for advancing the performance of EEG-based BCI spellers.
• Combining evidence from neural signals and linguistic context offers a significant pathway to faster and more accurate BCI communication.