Survey: large language models for neurological signal interpretation

• A Frontiers survey reviews large language models for interpreting neurological signals (e.g. EEG).¹
• The survey informs near-term tooling for neuroinformatics and BCI; state of the art for interpretation and potential automation is summarized.¹
• Implementation path for LLM-based signal interpretation is still emerging.^{1 1}

Weekly enrichment (2026-07-20)

• The survey is Chandrasekharan & Jacob (2025), “Bridging neuroscience and AI: a survey on large language models for neurological signal interpretation,” published in Frontiers in Neuroinformatics, volume 19, article 1561401.^{2 3}
• Its central argument is that LLMs can reach expert-level EEG diagnostic performance with minimal training data, light fine-tuning, and lower compute than convolutional or sequence-to-sequence networks, while also producing human-readable reasoning for transparency.²
• The review traces the GPT lineage (GPT-3 ≈175B parameters with sparse attention and few-shot in-context learning; GPT-4 >1T parameters with RLHF and chain-of-thought) as the backbone for these signal-interpretation methods.²
• EEG-GPT (Kim et al., 2024) fine-tunes an LLM on quantitative EEG features to classify recordings as normal vs. abnormal using the Temple University Hospital Abnormal Corpus—about 1,140 hours of EEG from 2,993 subjects—matching deep-learning baselines with roughly 50× less labeled data and emitting a reasoning chain.^{2 4}
• A word-level study distinguished high- vs. low-relevance words from eye-gaze-gated EEG, reporting over 60% validation accuracy across 12 subjects, described as the first LLM-based classification of brain states at the word level.²
• A reading-comprehension model fusing eye-tracking and EEG via an attention-based transformer reached 71.2% best single-subject accuracy and 68.7% mean 5-fold cross-validation accuracy across nine people, with a fine-tuned BERT word-embedding step hitting 92.7%.²
• Practical deployment leans on parameter-efficient methods: Low-Rank Adaptation (LoRA) for fine-tuning, model pruning (e.g., a Qwen2-0.5B pipeline pruned ~50%), and Retrieval-Augmented Generation for local, privacy-preserving execution.^{2 5}
• The survey is candid about limits: for direct EEG-based sleep-stage and attention-state estimation, fine-tuned LLMs underperformed traditional models such as XGBoost, and it flags ethical, privacy, and optimization bottlenecks (broad clinical superiority not established).²
• BCI implication: LLMs are positioned as interpretive companions—flagging signal quality, translating EEG into clinical language, and enabling EEG-to-text decoding and emotion/mental-health monitoring—rather than turnkey replacements for existing decoders.^{2 6}

Footnotes

1. https://news.google.com/rss/articles/CBMimgFBVV95cUxQY0JWU3dLa3ZEdWxEcmxwRjFydTJNbW5nd050NGJ4cW5KYVlRUzBsdDVLblBmeDExWUZjbDVGdDdqSmVobUN3ODBVT052NWxya3V5NzRiUXJrYy1QaU4wRThRdndCbWZGVDFncUhvS2p6NmgtQS02clVFZnZ0N1FTUm9WdXg0NkFPTUY3QTQ0eHZESWFYM2ZQaXN3?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.frontiersin.org/journals/neuroinformatics/articles/10.3389/fninf.2025.1561401/full ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

3. https://doi.org/10.3389/fninf.2025.1561401 ↩

4. https://arxiv.org/abs/2401.18006 ↩

5. https://www.brainaccess.ai/from-literature-to-live-data-llms-as-full-cycle-research-companions/ ↩

6. https://arxiv.org/html/2506.06353v1 ↩