CLEF: EEG Foundation Model for Learning Clinical Se... | AI Research

CLEF: EEG Foundation Model for Learning Clinical Semantics
Clinical EEG interpretation is a complex task that requires neurologists to analyze entire recording sessions—often lasting 20 minutes or longer—while integrating the patient’s medical history, medications, and specific clinical context. Current EEG foundation models are primarily designed for short-window tasks, such as brain-computer interfaces, which focus on brief segments of neural activity. CLEF (Clinical EEG Foundation model) is designed to bridge this gap by enabling long-context modeling at the scale of a full EEG session and grounding the model in clinical semantics, allowing it to reason about patient-level health data.

Moving Beyond Short Segments

To handle the massive amount of data in a full EEG session, CLEF avoids processing raw waveforms, which are computationally expensive and contain stochastic noise. Instead, it converts EEG data into 3D multitaper spectrograms. This approach discards irrelevant phase information while preserving the stable spectral features that neurologists rely on for diagnosis. By using a VQGAN-based tokenizer, the model compresses a 20-minute recording into a compact sequence of 2,048 tokens, making it possible to use Transformer architectures to analyze the entire session at once.

Grounding in Clinical Context

Signal analysis alone is often insufficient to capture a patient's full clinical profile. To address this, CLEF uses a two-stage training process. After learning to reconstruct EEG signals, the model is aligned with two critical clinical data sources: neurologist reports and structured electronic health records (EHR). By aligning EEG embeddings with these sources, the model learns to associate specific neural patterns with clinical concepts like medication effects, disease phenotypes, and diagnoses. The reports are processed using an LLM-based summarization pipeline to ensure the model focuses on relevant electrographic content, while EHR data is encoded using its native structure to maintain the integrity of patient information.

Performance and Generalization

The researchers evaluated CLEF on a new, large-scale benchmark consisting of 234 clinical tasks, utilizing over 260,000 EEG sessions from more than 108,000 patients. CLEF outperformed existing EEG foundation models on 229 of these 234 tasks, significantly improving the mean AUROC from 0.65 to 0.74. The results indicate that the model’s representations are robust, transferring effectively to new diseases, medications, and external datasets that were not included in the training process.

A New Paradigm for EEG

The success of CLEF suggests that the future of clinical EEG foundation models lies in session-scale representation learning. By combining long-context signal modeling with clinical grounding, the system moves closer to the way human experts interpret EEG data. This approach provides a scalable, benchmarkable framework that could help democratize expert-level neurophysiological assessment, potentially assisting clinicians in interpreting the millions of EEG recordings performed annually.