CLEF: A Clinically Grounded EEG Foundation Model for Long-Context Semantic Learning
Researchers have introduced CLEF, a novel foundation model designed to enhance clinical electroencephalogram (EEG) interpretation by addressing the limitations of existing short-window decoding models. Unlike previous approaches, CLEF incorporates full-session clinical context by representing EEG data as 3D multitaper spectrogram tokens, enabling efficient Transformer modeling at scale. The model aligns these embeddings with neurologist reports and structured electronic health record (EHR) data using contrastive objectives. Evaluated on a comprehensive benchmark comprising 234 tasks across disease phenotypes, medication exposures, and EEG findings, CLEF demonstrated superior performance. Using a dataset of over 260,000 EEG sessions from more than 108,000 patients, the model outperformed prior foundations on 229 tasks, raising the mean AUROC from 0.65 to 0.74. Results indicate that combining reconstruction-only pretraining with clinical report and EHR alignment significantly boosts accuracy. Furthermore, experiments suggest these representations generalize well to unseen concepts and external cohorts, establishing session-scale, clinically grounded representation learning as a promising paradigm for advancing automated clinical EEG analysis and improving diagnostic precision in neurological care.
Wire timeline
CLEF: A Clinically Grounded EEG Foundation Model for Long-Context Semantic Learning
Researchers have introduced CLEF, a novel foundation model designed to enhance clinical electroencephalogram (EEG) interpretation by addressing the limitations of existing short-window decoding models. Unlike previous approaches, CLEF incorporates full-session clinical context by representing EEG data as 3D multitaper spectrogram tokens, enabling efficient Transformer modeling at scale. The model aligns these embeddings with neurologist reports and structured electronic health record (EHR) data using contrastive objectives. Evaluated on a comprehensive benchmark comprising 234 tasks across disease phenotypes, medication exposures, and EEG findings, CLEF demonstrated superior performance. Using a dataset of over 260,000 EEG sessions from more than 108,000 patients, the model outperformed prior foundations on 229 tasks, raising the mean AUROC from 0.65 to 0.74. Results indicate that combining reconstruction-only pretraining with clinical report and EHR alignment significantly boosts accuracy. Furthermore, experiments suggest these representations generalize well to unseen concepts and external cohorts, establishing session-scale, clinically grounded representation learning as a promising paradigm for advancing automated clinical EEG analysis and improving diagnostic precision in neurological care.
cs.AI updates on arXiv.org