Language Models as Catalysts in EEG-Based BCI Speller Systems: A Low-Cost Solution for Paralyzed Patients

Overview

Neuromuscular conditions, including amyotrophic lateral sclerosis, stroke, spinal cord injury, and cerebral palsy, are affecting nearly fourteen million people globally, hindering their ability to communicate. This study presents a novel non-invasive brain-computer interface (BCI) system, leveraging a commercially available EEG device (Muse 2) and a large language model (LLM) to enable communication for individuals with neuromuscular conditions. Utilizing EEG signals from eye movements, the system captures keywords which are then expanded into sentences using a pre-trained LLM (LLaMA-2). The EEG data, collected with the device's electrodes (TP9, AF7, AF8, TP10) under IRB approval, was used to train a 1 -Dconvolutional neural network. This model classifies eye navigation in four directions from 0.25-second signals with 96.9% accuracy. Users input keywords through a simple interface, and the LLM, considering context like emotions and conversation history, generates varied sentences. This research evaluates the quality of the generated sentences by creating a single-sentence dataset (hospital500) and extracting keywords from the multi-turn dialogue dataset (DailyDialog). Compared to ground -truth sentences in these datasets, our system's top generated sentences (top 1) achieved sentence transformer-based text similarity scores of 0.86 and 0.74, respectively. The proposed system can achieve speeds of 200 characters per minute (CPM), a significant improvement over the 50 CPM of current non -invasive systems, but it st ill lags behind the 310 CPM of invasive systems. However, it reduces the cost to approximately 1/370th of an average BCI system, representing a substantial lowering of the cost-prohibitive barrier for people in need.