{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T18:09:07Z","timestamp":1790878147615,"version":"4.1.0"},"reference-count":48,"publisher":"Public Library of Science (PLoS)","issue":"10","license":[{"start":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T00:00:00Z","timestamp":1790812800000},"content-version":"vor","delay-in-days":0,"URL":"http:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["www.plosone.org"],"crossmark-restriction":false},"short-container-title":["PLoS One"],"abstract":"<jats:p>\n                    Amyotrophic lateral sclerosis (ALS), a progressive neurodegenerative disease, severely impairs communication, requiring assistive technologies that restore interaction. The P300 speller brain-computer interface (BCI) enables communication by translating EEG responses into text; however, its practical adoption is limited by slow typing speed and the need for subject-specific calibration. Recent work has demonstrated that large language models (LLMs), such as GPT-2, can significantly improve the performance of the P300 speller by predicting words. However, it remains unclear whether these gains are model-specific or represent a broader trend across language models. Furthermore, the framework-specific performance bounds of LLM-assisted P300 spellers have not been systematically characterized within a unified decoding framework. In this study, we address these gaps through a systematic multi-model theoretical analysis framework. We evaluate a wide range of language models and introduce an idealized LLM to establish upper bounds on achievable performance. In addition, we incorporate cross-subject classifier training to reduce calibration requirements and assess generalization across subjects. Using extensive simulations on EEG data from 78 subjects, we demonstrate that the evaluated models consistently achieve substantial improvements in typing speed, with gains of up to\n                    <jats:inline-formula id=\"pone.0349281.e001\">\n                      <jats:alternatives>\n                        <jats:graphic xmlns:xlink=\"http:\/\/www.w3.org\/1999\/xlink\" id=\"pone.0349281.e001g\" mimetype=\"image\" position=\"anchor\" xlink:href=\"info:doi\/10.1371\/journal.pone.0349281.e001\" xlink:type=\"simple\"\/>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" display=\"inline\" id=\"M1\">\n                          <mml:mrow>\n                            <mml:mi>~<\/mml:mi>\n                            <mml:mspace width=\"negativethinmathspace\"\/>\n                            <mml:mn>45<\/mml:mn>\n                            <mml:mi>%<\/mml:mi>\n                          <\/mml:mrow>\n                        <\/mml:math>\n                      <\/jats:alternatives>\n                    <\/jats:inline-formula>\n                    (within-subject training) and\n                    <jats:inline-formula id=\"pone.0349281.e002\">\n                      <jats:alternatives>\n                        <jats:graphic xmlns:xlink=\"http:\/\/www.w3.org\/1999\/xlink\" id=\"pone.0349281.e002g\" mimetype=\"image\" position=\"anchor\" xlink:href=\"info:doi\/10.1371\/journal.pone.0349281.e002\" xlink:type=\"simple\"\/>\n                        <mml:math xmlns:mml=\"http:\/\/www.w3.org\/1998\/Math\/MathML\" display=\"inline\" id=\"M2\">\n                          <mml:mrow>\n                            <mml:mi>~<\/mml:mi>\n                            <mml:mspace width=\"negativethinmathspace\"\/>\n                            <mml:mn>75<\/mml:mn>\n                            <mml:mi>%<\/mml:mi>\n                          <\/mml:mrow>\n                        <\/mml:math>\n                      <\/jats:alternatives>\n                    <\/jats:inline-formula>\n                    (across-subject training) over conventional approaches. More importantly, we show that several high-performing models, despite architectural differences, operate within 5% of the theoretical performance bound, indicating diminishing returns from further model scaling. These improvements generalize across both within-subject and across-subject classifiers. Our results suggest that LLM-assisted P300 spellers are approaching oracle-based upper bounds within the considered decoding framework, shifting the primary bottleneck from language modeling to neural signal decoding. This work provides both a practical framework for improving BCI communication and a theoretical perspective on its achievable limits.\n                  <\/jats:p>","DOI":"10.1371\/journal.pone.0349281","type":"journal-article","created":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T17:50:50Z","timestamp":1790877050000},"page":"e0349281","update-policy":"https:\/\/doi.org\/10.1371\/journal.pone.corrections_policy","source":"Crossref","is-referenced-by-count":0,"title":["Near-optimal P300 speller performance using large language models: A multi-model analysis with performance bounds"],"prefix":"10.1371","volume":"21","author":[{"ORCID":"https:\/\/orcid.org\/0000-0003-2478-1123","authenticated-orcid":true,"given":"Nithin","family":"Parthasarathy","sequence":"first","affiliation":[{"name":"University of Illinois at Urbana-Champaign","place":["Urbana, Illinois, United States of America"],"department":["Department of Computer Science"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"James","family":"Soetedjo","sequence":"additional","affiliation":[{"name":"University of Washington","place":["Seattle, Washington, United States of America"],"department":["Department of Bioengineering"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Saarang","family":"Panchavati","sequence":"additional","affiliation":[{"name":"University of California","place":["Los Angeles, California, United States of America"],"department":["Department of Radiological Sciences"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Nitya","family":"Parthasarathy","sequence":"additional","affiliation":[{"name":"PE Investments","place":["Boston, Massachusetts, United States of America"]},{"name":"Massachusetts Institute of Technology","place":["Cambridge, Massachusetts, United States of America"],"department":["Department of Computer Science"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Dongwoo","family":"Lee","sequence":"additional","affiliation":[{"name":"University of California","place":["Los Angeles, California, United States of America"],"department":["Department of Radiological Sciences"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Corey","family":"Arnold","sequence":"additional","affiliation":[{"name":"University of California","place":["Los Angeles, California, United States of America"],"department":["Department of Radiological Sciences"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Nader","family":"Pouratian","sequence":"additional","affiliation":[{"name":"University of Texas Southwestern Medical Center","place":["Dallas, Texas, United States of America"],"department":["Department of Neurological Surgery"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-0890-8684","authenticated-orcid":true,"given":"William","family":"Speier","sequence":"additional","affiliation":[{"name":"University of California","place":["Los Angeles, California, United States of America"],"department":["Department of Radiological Sciences"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"340","published-online":{"date-parts":[[2026,10,1]]},"reference":[{"issue":"10107","key":"pone.0349281.ref001","doi-asserted-by":"crossref","first-page":"2084","DOI":"10.1016\/S0140-6736(17)31287-4","article-title":"Amyotrophic lateral sclerosis","volume":"390","author":"MA van Es","year":"2017","journal-title":"Lancet"},{"issue":"5","key":"pone.0349281.ref002","doi-asserted-by":"crossref","first-page":"1436","DOI":"10.1109\/TBME.2014.2300164","article-title":"Visual and auditory brain-computer interfaces","volume":"61","author":"S Gao","year":"2014","journal-title":"IEEE Trans Biomed Eng"},{"issue":"4","key":"pone.0349281.ref003","doi-asserted-by":"crossref","first-page":"731","DOI":"10.1016\/j.clinph.2010.10.029","article-title":"The P300-based brain-computer interface (BCI): effects of stimulus rate","volume":"122","author":"DJ McFarland","year":"2011","journal-title":"Clin Neurophysiol"},{"issue":"6","key":"pone.0349281.ref004","doi-asserted-by":"crossref","first-page":"510","DOI":"10.1016\/0013-4694(88)90149-6","article-title":"Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials","volume":"70","author":"LA Farwell","year":"1988","journal-title":"Electroencephalogr Clin Neurophysiol"},{"issue":"7","key":"pone.0349281.ref005","doi-asserted-by":"crossref","first-page":"1109","DOI":"10.1016\/j.clinph.2010.01.030","article-title":"A novel P300-based brain-computer interface stimulus presentation paradigm: moving beyond rows and columns","volume":"121","author":"G Townsend","year":"2010","journal-title":"Clin Neurophysiol"},{"issue":"1","key":"pone.0349281.ref006","doi-asserted-by":"crossref","first-page":"13","DOI":"10.1080\/2326263X.2017.1410418","article-title":"Improving P300 Spelling Rate using Language Models and Predictive Spelling","volume":"5","author":"W Speier","year":"2018","journal-title":"Brain Comput Interfaces (Abingdon)"},{"issue":"3","key":"pone.0349281.ref007","doi-asserted-by":"crossref","first-page":"031002","DOI":"10.1088\/1741-2560\/13\/3\/031002","article-title":"Integrating language models into classifiers for BCI communication: a review","volume":"13","author":"W Speier","year":"2016","journal-title":"J Neural Eng"},{"key":"pone.0349281.ref008","doi-asserted-by":"crossref","unstructured":"Speier W, Arnold CW, Deshpande A, Knall J, Pouratian N. Incorporating advanced language models into the P300 speller using particle filtering. J Neural Eng. 2014.","DOI":"10.1088\/1741-2560\/12\/4\/046018"},{"key":"pone.0349281.ref009","unstructured":"Kindermans PJ, Verschore H, Verstraeten D, Schrauwen B. A P300 BCI for the masses: Prior information enables instant unsupervised spelling. In: Advances in Neural Information Processing Systems. vol. 25. 2012. p. 719\u201327."},{"issue":"3","key":"pone.0349281.ref010","doi-asserted-by":"crossref","first-page":"035005","DOI":"10.1088\/1741-2560\/11\/3\/035005","article-title":"Integrating dynamic stopping, transfer learning and language models in an adaptive zero-training ERP speller","volume":"11","author":"P-J Kindermans","year":"2014","journal-title":"J Neural Eng"},{"key":"pone.0349281.ref011","doi-asserted-by":"crossref","unstructured":"Parthasarathy N, Arnold C, Pouratian N, Speier W. Optimizing P300 speller performance using language models for character and word prediction. In: IEEE Bioinformatics and Bioengineering Conference. 2020.","DOI":"10.1109\/BIBE50027.2020.00137"},{"issue":"4","key":"pone.0349281.ref012","doi-asserted-by":"crossref","first-page":"210","DOI":"10.1080\/2326263X.2024.2413214","article-title":"High performance P300 spellers using GPT2 word prediction with cross-subject training","volume":"11","author":"N Parthasarathy","year":"2024","journal-title":"Brain-Comput Interfaces"},{"key":"pone.0349281.ref013","doi-asserted-by":"crossref","unstructured":"Lebedev MA, Makarova A, Kleeva D, Maysuradze AI. Connecting a P300 speller to a large language model. BioRxiv. 2025. doi: 10.1101\/2025.11.06.686984","DOI":"10.1101\/2025.11.06.686984"},{"issue":"1","key":"pone.0349281.ref014","doi-asserted-by":"crossref","first-page":"6379","DOI":"10.1038\/s41598-025-25660-7","article-title":"ChatBCI, a P300 speller BCI with context-driven word prediction leveraging large language models, from concept to evaluation","volume":"16","author":"J Hong","year":"2026","journal-title":"Sci Rep"},{"key":"pone.0349281.ref015","unstructured":"Liu Y, Ott M, Goyal N, Du J, Joshi M, Chen D, et al. RoBERTa: A Robustly Optimized BERT Pretraining Approach. ArXiv:1907.11692."},{"key":"pone.0349281.ref016","unstructured":"Devlin J, Chang M, Lee K, Toutanova K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding;. ArXiv:1810.04805."},{"key":"pone.0349281.ref017","volume-title":"Language models are unsupervised multitask learners","author":"A Radford","year":"2019"},{"key":"pone.0349281.ref018","unstructured":"LLaMA 3.1: Open-Source Large Language Model. 2024. Available from: https:\/\/ai.meta.com\/blog\/meta-llama-3-1\/"},{"key":"pone.0349281.ref019","unstructured":"LLaMA 3.3: Multilingual Large Language Model. 2024. Available from: https:\/\/huggingface.co\/meta-llama\/Llama-3.3-70B-Instruct"},{"key":"pone.0349281.ref020","unstructured":"Almazrouei E, Alobeidli H, Alshamsi A, Cappelli A, Cojocaru R, Debbah M, et al. The Falcon Series of Open Language Models. 2023. ArXiv:2311.16867 p. Available from: https:\/\/arxiv.org\/abs\/2311.16867"},{"key":"pone.0349281.ref021","unstructured":"Lewis M, Liu Y, Goyal N, Ghazvininejad M, Abdelrahman M, Levy O, et al. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. ArXiv:1910.13461."},{"key":"pone.0349281.ref022","unstructured":"Yang Z, Dai Z, Yang Y, Carbonell J, Salakhutdinov R, Le Q. XLNet: Generalized Autoregressive Pretraining for Language Understanding. ArXiv:1906.08237."},{"key":"pone.0349281.ref023","unstructured":"Black S, Gao L, Wang P, Connor L, Biderman S. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow. 2021. Available from: doi: 10.5281\/zenodo.5297715"},{"key":"pone.0349281.ref024","unstructured":"Scao TL, Fan A, Akiki C, et al. BLOOM: A 176B-Parameter Open-Access Multilingual Language Model. 2022. ArXiv:2211.05100 p. Available from: doi: 10.48550\/arXiv.2211.05100"},{"key":"pone.0349281.ref025","unstructured":"Zhang H, Niu X, et al. OPT: Open Pretrained Transformer Language Models. Available from: doi: 10.48550\/arXiv.2205.01068"},{"key":"pone.0349281.ref026","author":"B Wang","year":"2021"},{"key":"pone.0349281.ref027","unstructured":"Black S, Mccarthy J. GPT-NeoX: Large-scale autoregressive language model. Available from: doi: 10.48550\/arXiv.2204.09216"},{"key":"pone.0349281.ref028","unstructured":"MosaicML NLP Team. Introducing MPT-30B: Raising the bar for open-source foundation models. MosaicML Technical Blog. 2023. Available from: https:\/\/www.mosaicml.com\/blog\/mpt-30b"},{"key":"pone.0349281.ref029","doi-asserted-by":"crossref","DOI":"10.1090\/clrm\/016","volume-title":"Cryptological Mathematics","author":"RE Lewand","year":"2000"},{"key":"pone.0349281.ref030","volume-title":"Speech and language processing: an introduction to natural language processing, computational linguistics, and speech recognition","author":"D Jurafsky","year":"2009"},{"issue":"10","key":"pone.0349281.ref031","doi-asserted-by":"crossref","first-page":"2696","DOI":"10.1109\/TBME.2013.2262524","article-title":"A unified probabilistic approach to improve spelling in an event-related potential-based brain-computer interface","volume":"60","author":"P-J Kindermans","year":"2013","journal-title":"IEEE Trans Biomed Eng"},{"key":"pone.0349281.ref032","unstructured":"Francis W, Kucera H. Brown Corpus Manual. 1979."},{"issue":"2","key":"pone.0349281.ref033","doi-asserted-by":"crossref","first-page":"306","DOI":"10.1016\/j.clinph.2012.08.002","article-title":"The effects of stimulus timing features on P300 speller performance","volume":"124","author":"J Lu","year":"2013","journal-title":"Clin Neurophysiol"},{"issue":"3","key":"pone.0349281.ref034","doi-asserted-by":"crossref","first-page":"242","DOI":"10.1016\/j.biopsycho.2006.04.007","article-title":"A P300 event-related potential brain-computer interface (BCI): the effects of matrix size and inter stimulus interval on performance","volume":"73","author":"EW Sellers","year":"2006","journal-title":"Biol Psychol"},{"key":"pone.0349281.ref035","volume-title":"Applied Regression Analysis","author":"NR Draper","year":"1981"},{"key":"pone.0349281.ref036","volume-title":"Introduction to Algorithms","author":"TH Cormen","year":"2009","edition":"3"},{"issue":"1","key":"pone.0349281.ref037","doi-asserted-by":"crossref","first-page":"269","DOI":"10.1007\/BF01386390","article-title":"A note on two problems in connexion with graphs","volume":"1","author":"EW Dijkstra","year":"1959","journal-title":"Numer Math"},{"key":"pone.0349281.ref038","author":"A Radford","year":"2018"},{"issue":"6","key":"pone.0349281.ref039","doi-asserted-by":"crossref","first-page":"1171","DOI":"10.1016\/j.clinph.2014.09.021","article-title":"A method for optimizing EEG electrode number and configuration for signal acquisition in P300 speller systems","volume":"126","author":"W Speier","year":"2015","journal-title":"Clin Neurophysiol"},{"key":"pone.0349281.ref040","unstructured":"Declaration of Independence. Available from: http:\/\/ushistory.org\/declaration\/document"},{"issue":"3","key":"pone.0349281.ref041","doi-asserted-by":"crossref","first-page":"237","DOI":"10.1016\/S0301-0511(03)00073-5","article-title":"Brain-computer interface (BCI) operation: optimizing information transfer rates","volume":"63","author":"DJ McFarland","year":"2003","journal-title":"Biol Psychol"},{"key":"pone.0349281.ref042","volume-title":"An introduction to information theory","author":"J Pierce","year":"1980"},{"issue":"14","key":"pone.0349281.ref043","first-page":"743","article-title":"Cognitive Load with Different Origins: An EEG Model-Based Explanation","volume":"20","author":"Y Gogna","year":"2022","journal-title":"NeuroQuantology"},{"issue":"11","key":"pone.0349281.ref044","doi-asserted-by":"crossref","first-page":"2992","DOI":"10.1109\/TBME.2018.2875024","article-title":"Improving Performance of Devanagari Script Input-Based P300 Speller Using Deep Learning","volume":"66","author":"GB Kshirsagar","year":"2019","journal-title":"IEEE Trans Biomed Eng"},{"issue":"3","key":"pone.0349281.ref045","doi-asserted-by":"crossref","first-page":"678","DOI":"10.1109\/TNSRE.2014.2300091","article-title":"Integrating language information with a hidden Markov model to improve communication rate in the P300 speller","volume":"22","author":"W Speier","year":"2014","journal-title":"IEEE Trans Neural Syst Rehabil Eng"},{"key":"pone.0349281.ref046","volume-title":"Nonparametric Statistics: A Step-by-Step Approach","author":"G Corder","year":"2014"},{"issue":"6","key":"pone.0349281.ref047","doi-asserted-by":"crossref","first-page":"80","DOI":"10.2307\/3001968","article-title":"Individual Comparisons by Ranking Methods","volume":"1","author":"F Wilcoxon","year":"1945","journal-title":"Biom Bull"},{"key":"pone.0349281.ref048","article-title":"Language Models are Few-Shot Learners","volume":"33","author":"T Brown","year":"2020","journal-title":"Adv Neural Inform Process Syst"}],"container-title":["PLOS One"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/dx.plos.org\/10.1371\/journal.pone.0349281","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T17:50:59Z","timestamp":1790877059000},"score":1,"resource":{"primary":{"URL":"https:\/\/dx.plos.org\/10.1371\/journal.pone.0349281"}},"subtitle":[],"editor":[{"given":"Onder","family":"Aydemir","sequence":"first","affiliation":[{"name":"Karadeniz Technical University: Karadeniz Teknik Universitesi","place":["T\u00dcRKIYE"]}],"role":[{"vocabulary":"crossref","role":"editor"}]}],"short-title":[],"issued":{"date-parts":[[2026,10,1]]},"references-count":48,"journal-issue":{"issue":"10","published-online":{"date-parts":[[2026,10,1]]}},"URL":"https:\/\/doi.org\/10.1371\/journal.pone.0349281","relation":{},"ISSN":["1932-6203"],"issn-type":[{"value":"1932-6203","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,10,1]]}}}