Optimizing P300 speller performance using language models for character and word prediction

Neurodegenerative diseases such as Amyotrophic Lateral Sclerosis (ALS) restrict an individual's ability to fully engage with their surroundings by interrupting crucial cell signaling processes between the brain and the peripheral nervous system. In such cases, brain-computer interfaces (BCI) such as the P300 speller use electroencephalography (EEG) to capture subject brain signals which are decoded to provide an alternate channel for communication. This paper develops new and sophisticated BCI techniques combining language models, probabilistic flashboard design, and optimal scanning techniques which together with smoothing algorithms and word suggestions can be effectively utilized as a form of predictive spelling. Detailed offline simulations based on subject EEG data allowed for testing a variety of methods on a large subject population producing extensive output text including both in vocabulary and out of vocabulary words. Significant performance gains were seen when applying language models to classification and optimizing the interface for stimulus presentation. The results demonstrate that feed-forward techniques with appropriate flashboard design and word completion algorithms can achieve similar or better performance as more complex feedback schemes. Consequently, one can employ lower complexity flashboard as well as control and yet achieve the performance gains offered by more complex techniques.