Research map / Large language models / Language, safety and interpretability

CLEF: EEG Foundation Models for Learning Clinical Semantics

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Linguistic encoding in LLM representations

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 CLEF, a long-context EEG foundation model that tokenizes full 20-minute multitaper spectrograms with a VQGAN and aligns patient embeddings with neurologist reports and structured EHR via contrastive losses, evaluated on a 234-task clinical benchmark from HEEDB.

Problem

Existing EEG foundation models operate on short BCI-style windows and ignore clinical context, while clinical EEG interpretation requires reasoning over full sessions integrated with clinical information.

Approach

Represent full EEG sessions as compact 3D spectrogram tokens (multi-channel VQGAN tokenization with a channel-decomposed reconstruction loss), pretrain by masked code prediction, then align session embeddings with LLM-summarized reports and structured EHR code sets via symmetric contrastive losses.

Key results

• CLEF outperforms prior EEG foundation models on 229 of 234 tasks, improving mean AUROC from 0.65 to 0.74 (0.737 ± 0.006).
• Reconstruction-only CLEF reaches 0.671 AUROC, above retrained REVE at 0.655 on the same data.
• On external cohorts CLEF reaches AUROC 0.939 on TUAB, 0.960 on TUEP and 0.689 on HSP, the best of all methods reported.

Surprising findings

• Reconstruction-only pretraining already surpasses prior EEG foundation models, with alignment adding further gains.
• Some tasks such as propofol and anoxic brain injury are already high with reconstruction alone and gain little from report/EHR alignment.

Limitations

• Not suitable for real-time monitoring; summarizes a session as one patient-level vector.
• Context windows of 20 minutes may be too short for full sleep architecture or multi-day trends.
• Embeddings are not interpretable; no attribution mechanisms provided.

Experimental setup

Datasets: Harvard Electroencephalography Database (HEEDB) TUAB TUEP HSP (PhysioNet Challenge 2026)

Benchmarks: 234-task HEEDB clinical benchmark TUAB TUEP HSP

Metrics: AUROC balanced accuracy

Baselines: BIOT LaBraM CBraMod REVE NeuroLM

Builds on: VQGAN CLIP masked image modeling LaBraM REVE

Closest papers by meaning

• What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.89)
• Neural Signals Generate Clinical Notes in the Wild (NeurIPS 2026 · similarity 0.88)
• EEG-Language Pretraining for Highly Label-Efficient Clinical Phenotyping (ICML 2025 · similarity 0.87)
• CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.87)
• Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.87)
• NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.87)