Effect of Large Language Models on P300 Speller Performance with Cross-Subject Training

Amyotrophic lateral sclerosis (ALS), a progressive neuromuscular degenerative disease, rapidly impairs communication within years of onset. This loss of communication necessitates assistive technologies to restore interaction and independence. One such technology, the P300 speller brain-computer interface (BCI), translates EEG signals into text by tracking a subjects neural responses to highlighted characters on a screen. A central challenge in P300-based research is enhancing performance to enable faster and more efficient user interaction. In this context, this study addresses key limitations, particularly in the training of multi-subject classifiers, and integrating advanced language models to optimize stimuli presentation and word prediction, thereby improving communication efficiency. Specifically, we introduce three key innovations: - Advanced multi-subject classifier training - Integrating and evaluating impact of numerous large language models (LLMs) on speller performance - Determining P300 LLM performance bounds using an ideal LLM with perfect prediction We conduct extensive simulations using randomly sampled EEG data. Our results demonstrate substantial speed improvements in typing passages that include rare and out-of-vocabulary (OOV) words. The magnitude of improvement depends on the type of language model used. More specifically, character-level models provide typing speed improvements of approximately 10%, while open-source LLMs such as Llama, Mistral and GPT2 achieve around 40% improvement through efficient word prediction. Additionally, we construct an ideal LLM to establish theoretical performance limits and show that many modern LLMs achieve performance levels within 10% of it. Further, we show that these LLM-driven speed improvements generalize across classifiers, including those designed to reduce subject-specific training.1