A Cardiorespiratory Foundation Model Pretrained on Ten Physiological Signals for Perioperative and Critical Care
November 13, 2026
CARMEN is a single-encoder foundation model pretrained on 2.27 million modality-hours of ten distinct cardiorespiratory waveforms. Using the LSCNorm conditioning mechanism to preserve absolute signal levels, the encoder achieves an 0.894 AUROC for intraoperative hypotension detection and outperforms existing supervised and foundation models across 19 clinical tasks.