Large language models (LLMs) are starting to be coupled with brain-computer interfaces (BCIs) for assistive communication, but the resulting systems differ widely in where the model sits in the pipeline and in what they actually measure. We performed a systematic review, prepared according to PRISMA, of eleven studies that combine an LLM with a BCI for communication or control. The included work covers P300, SSVEP, cVEP, passive affective and auditory paradigms, and five integration patterns: autocomplete, post-edit correction, intent expansion, dynamic interface generation and affective support. For each study we extracted the hardware and decoding pipeline, the LLM and prompting strategy, latency reporting and outcomes; we used scenario-appropriate metrics rather than a single common metric. Risk of bias was judged with an adapted ROBINS-I framework that stratified studies into online, offline-simulation and system-proposal categories. In the copy-spelling scenario, two studies that measured keystroke savings directly reported values above 50%, with one study exceeding 60% in a multi-turn condition; on an intent-based ALS message-bank task, one online study reached 42 characters per minute with a semantic accuracy of 88%. None of the eleven studies enrolled motor-impaired patients, seven of eleven relied on remote OpenAI endpoints, and reporting of end-to-end latency and failure modes was sparse. We propose a five-category taxonomy of BCI/LLM integration, separate findings that are supported from those that are still speculative, and give a checklist of metrics that should be reported by future studies. The taxonomy and the reporting checklist are the main contributions; clinical benefit for the target population remains to be shown.
Gorenshtein et al. (Wed,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: