CARMEN: A Multimodal Foundation Model for Cardiorespiratory Waveforms in Critical Care

Existing biosignal foundation models (FMs) are heavily skewed toward non-invasive signals, largely excluding the invasive hemodynamic and respiratory waveforms central to critical care. Their normalization pipelines discard the absolute units that anchor clinical decision-making, and they lack explicit mechanisms for learning physiological interactions across signals. We present CARMEN, a multimodal FM spanning nine biosignals across hemodynamics and respiration. Pretrained on 668,000 hours of intensive care unit (ICU) data, CARMEN operates on arbitrary signal combinations and lengths through a single encoder, preserves absolute-unit information via its LSCNorm mechanism, and learns cross-modal physiological interactions directly through dedicated objectives. Across multiple independent cohorts held out from pretraining, CARMEN not only outperforms supervised models on most acute event detection and prediction tasks and surpasses existing biosignal FMs across modality combinations, but also successfully transfers to clinical outcome prediction with a frozen encoder. These results suggest CARMEN's potential as a clinical biosignal FM. It flexibly accommodates the complex and heterogeneous scenarios of ICUs and operating rooms.