ChatBCI, a P300 speller BCI with context-driven word prediction leveraging large language models, from concept to...

Area of Science:

• Neuroscience and Artificial Intelligence
• Human-Computer Interaction

Background:

• P300 speller brain-computer interfaces (BCIs) enable communication by detecting P300 brainwaves for key selection.
• Current P300 spellers often require extensive keystrokes, increasing user time and cognitive load.
• There is a need for more efficient and user-friendly BCI communication methods.

Purpose of the Study:

• To introduce ChatBCI, a novel P300 speller BCI system.
• To leverage large language models (LLMs) for predictive word suggestions and accelerate sentence composition.
• To reduce keystrokes and cognitive load for BCI users.

Main Methods:

• ChatBCI utilizes LLM (GPT-3.5 API) zero-shot learning for word suggestions.
• A modified graphical user interface (GUI) displays LLM-generated word suggestions as selectable keys.
• P300 classification is performed using stepwise linear discriminant analysis (SWLDA).

Main Results:

• ChatBCI reduced time and keystrokes by 22% and 33% respectively in copy-spelling tasks.
• Information transfer rate increased by 15% in copy-spelling tasks.
• ChatBCI achieved 45% keystroke savings in improvised sentence composition tasks.

Conclusions:

• ChatBCI significantly outperforms traditional letter-by-letter BCI spellers in efficiency.
• The system leverages remote LLM queries, eliminating the need for local model training or storage.
• ChatBCI offers a promising pathway for developing next-generation, efficient speller BCIs for real-time communication, particularly for individuals with disabilities.