Near-optimal P300 speller performance using large language models: A multi-model analysis with performance bounds.

Language-model assistance substantially improved simulated P300 typing speed on existing EEG recordings. The result is promising for communication technology, but it was an offline analysis in healthy adults rather than a real-time trial in people with ALS.

Parthasarathy N, Soetedjo J, Panchavati S, Parthasarathy N, Lee D, Arnold C, Pouratian N, Speier W.
PLOS ONE · Publication date: 2026-10-01
Added to site: 2026-10-02
DOI: 10.1371/journal.pone.0349281

Electronic publication date verified in PubMed, Crossref and the full text. This source date is separate from the site's addition date.

Modalities

EEG, ERP, Brain-computer interface, AI

Clinical focus

P300 speller, Assistive communication, Cross-subject decoding

Population

Healthy young adults represented in an existing P300 speller dataset; no participant with ALS or another severe motor impairment was studied

Sample

EEG data from 78 healthy adults aged 20 to 35

Intervention / method

Multiple language models, word-completion methods, flashing schemes and within- or across-subject SWLDA classifiers were compared with an idealized language-model bound

Comparator

Conventional random and diagonal flashing approaches, Dijkstra-based completion and an oracle-style word-prediction bound

Clinician summary

Within-subject GPT-2 plus Dijkstra decoding reached 75.3 bits per minute, about 45% above the random baseline and within 4% of the word-performance bound. Across-subject analyses rose from about 34 to about 60 bits per minute, with gains generalizing across several high-performing model families.

How to interpret the findings

The work relied on offline simulation and healthy adults aged 20 to 35. It did not test online use, communication accuracy or fatigue in people with ALS, severe motor impairment or real-world assistive settings. Performance bounds apply to the evaluated framework, not to the information-theoretic capacity of P300 communication.

Equipment and disclosures

Previously collected P300 EEG data were analyzed offline. The checked full text evaluates decoding and language-model configurations rather than validating a commercial communication device.

The authors declared no competing interests and reported no specific funding. The article states that its dataset and code availability are provided in the paper.

Related evidence

Compare findings from the same clinical area, including mixed results.