Research map / Large language models / AI and society

Neural Signals Generate Clinical Notes in the Wild

NeurIPS 2026 · Large language models › AI and society · approach: Data commons and human baselines · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CELM, an EEG-to-language model that generates clinical EEG reports from up to three-hour recordings by combining a frozen CBraMod encoder, epoch-aggregated tokenization, sequence-aware alignment, and a frozen Qwen3-4B LLM, trained on a new 9,922-report EEG-report benchmark.

Problem

Writing clinical EEG reports from long-term recordings is labor-intensive, and prior EEG report generation methods use short contexts, classification-then-decoding pipelines, and task-specific templates.

Approach

Epoch-level tokens from a pretrained EEG encoder are aggregated and aligned with a sequence-aware transformer (SCA) or compressed via Perceiver-style queries (SCC) before projection into an LLM, enabling end-to-end report generation from hour-long EEG with only the alignment module trained.

Key results

• On S0002 with clinical context, CELM reaches ROUGE-1 0.6408 vs 0.3084 for the best baseline (Gemma-3-4b with linear projector + CBraMod), roughly two-fold.
• In zero-context generation on S0002, CELM achieves ROUGE-1 0.5248 vs 0.2557 for the best baseline Gemma-3-4b with EEG features.
• Six clinical experts ranked CELM first with a weighted-average rank of 1.29, and CELM significantly outperformed baselines on all five dimensions (Wilcoxon p < 0.001).

Surprising findings

• Adding handcrafted EEG band-power features to text inputs degraded performance on S0001 relative to text-only input, attributed to multi-section report complexity.
• Linear and Perceiver projectors overfit (rising validation loss) while SCA and SCC converge stably.

Limitations

• Abnormality decision-making remains weak; model often misclassifies mildly abnormal recordings as normal.
• SCC compression substantially underperforms the non-compressed SCA-based CELM (0.4487 vs 0.6408 ROUGE-1 on S0002).
• Dataset restricted to two hospital sites and single-session matches; closed LLM judges unavailable due to data-usage constraints.

Experimental setup

Datasets: Harvard Electroencephalography Database v4.1 (S0001, S0002) BDSP EHR neurology reports

Benchmarks: EEG-Report benchmark (S0001, S0002)

Metrics: BLEU-1 BLEU-4 ROUGE-1 ROUGE-2 ROUGE-L ROUGE-LSUM BERTScore METEOR human expert 5-point ratings LLM-as-judge ratings

Baselines: Gemma-3-4b-it MedGemma-4b-it Llama-3.1-8B-Instruct Qwen3-4B-Instruct-2507 Linear projector + CBraMod

Builds on: CBraMod LLaVA Perceiver

Closest papers by meaning

• NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals (ICLR 2025 · similarity 0.88)
• EEG-Language Pretraining for Highly Label-Efficient Clinical Phenotyping (ICML 2025 · similarity 0.88)
• CLEF: EEG Foundation Models for Learning Clinical Semantics (NeurIPS 2026 · similarity 0.88)
• NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.87)
• What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.87)
• BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.86)