EMBC Special Issue: ChatBCI-Assist: An Intent-Based P300 Speller with A Locally-Deployed LLM and Adaptive Stopping Strategy Enabling Record Online Spelling Performance.

Abstract

P300-based speller brain computer interfaces (BCIs) provide promising communication solutions for individuals with severe motor impairments such as those with amyotrophic lateral sclerosis (ALS). However, existing P300 spellers are constrained by slow typing speed and limited efficiency. Here, we present ChatBCI-Assist, an intent-based P300 speller that integrates a locally-deployed large language model (LLM), fine-tuned for the task at hand, with an adaptive stopping strategy for key selection and a graphical user interface (GUI) designed for efficient message composition, to achieve record-level online spelling performance. The LLM, trained on an ALS-specific communication corpus using low-rank adaptation (LoRA), produces context-aware, semantically coherent, and prefix-constrained word and phrase predictions in real time. The proposed GUI supports efficient, user-adaptive message composition, while the adaptive stopping strategy dynamically adjusts stimulus presentation based on each subject's classification performance. Combined with a subject specific stepwise linear discriminant analysis (SWLDA) classifier, ChatBCI-Assist enhances spelling efficiency. Results from online experiments demonstrate that ChatBCI-Assist achieves record performance, with an average information transfer rate (ITR) of 105.2 bits/min, an overall character-level mutual information rate (MIR) of 52.9 bits/min and characters per minute (CPM) of 19.7 in copy-spelling tasks, and 30.7 CPM in semantic spelling tasks. Evaluated using semantic ITR (SITR), a metric proposed to characterize semantic communication efficiency, ChatBCI-Assist achieved SITR of 147.1 bits/min. User experience evaluations further confirm reduced workload and higher usability from LLM-based semantic spelling configurations, compared to traditional copy-spelling paradigms (dictionary or LLM). This work demonstrates that integrating locally-adapted LLMs with intent driven design and subject-specific decoding optimization can substantially improve the speed, efficiency, and user experience of BCI-based communication systems.