CLEF: A Clinically Grounded Long-Context EEG Foundation Model
Researchers have developed a new EEG model called CLEF, designed for clinical settings. This model can analyze entire EEG sessions by combining signal patterns with clinical data. Unlike existing models that only decode short segments, CLEF uses 3D multitaper spectrogram tokens for better session-level analysis. It also aligns its outputs with neurologist reports and structured electronic health records (EHR) using contrastive methods. CLEF was tested on a groundbreaking set of 234 tasks related to disease traits, medication use, and EEG results, based on over 260,000 EEG sessions from more than 108,000 patients. It surpassed previous models on 229 tasks and improved the average AUROC from 0.65 to 0.74.
Key facts
- CLEF is a clinically grounded long-context EEG foundation model.
- It represents EEG sessions as 3D multitaper spectrogram tokens.
- Aligns embeddings with neurologist reports and structured EHR data.
- Evaluated on a 234-task benchmark with over 260,000 EEG sessions from 108,000+ patients.
- Outperformed prior EEG foundation models on 229 of 234 tasks.
- Improved mean AUROC from 0.65 to 0.74.
- Reconstruction-only pretraining surpassed prior models.
- Report and EHR alignment yielded further gains.
Entities
—