CARMEN packs ten cardiorespiratory waveforms into one encoder, beating task-specific models on most of 19 tasks with frozen weights

One encoder now reads the cardiovascular and respiratory waveforms recorded at operating-room and ICU bedsides, including arterial, central venous, pulmonary artery and intracranial pressures plus capnography, airway pressure, respiratory flow and impedance respiration, rather than electrocardiogram and photoplethysmogram alone. Earlier physiological-waveform foundation models were built almost entirely on the electrocardiogram and photoplethysmogram, and ICU foundation models stayed confined to ECG, PPG and peripheral oxygen saturation, leaving invasive hemodynamic and respiratory waveforms out of pretraining. The pretraining corpus holds 526,061 records and 2,274,701 modality-hours in total, of which invasive hemodynamic waveforms are 21.7% and respiratory waveforms 28.2%; these are waveforms left on bedside monitors during routine care, not records selected for research.

With the encoder frozen and only a linear layer trained, it beat end-to-end supervised models and existing pretrained foundation models on most of the 19 downstream tasks. Comparators were CARMEN-scratch, the same architecture trained end-to-end from random initialization, a 1D-ResNet, the general-purpose time-series model MOMENT, and the biosignal models NormWear and SPOTR, the last three evaluated with the same linear probe. On 12 of the 15 tasks with supervised comparators, CARMEN was significantly higher than both supervised models; across the 19 tasks it was significantly higher than MOMENT on 16, NormWear on 14 and SPOTR on 17, and none of the three was significantly higher than CARMEN on any task.

The model keeps the absolute signal level that normalization removes, so it is markedly more accurate on tasks decided by absolute thresholds. Standard normalization can make two arterial pressure waveforms of identical shape with mean arterial pressures of 55 mmHg and 85 mmHg indistinguishable, yet clinicians intervene on exactly such absolute numbers. Replacing LSCNorm with RMSNorm dropped the AUROC for intraoperative hypotension from 0.894 to 0.780 and for EtCO2 derangement from 0.730 to 0.709, while the difference was only 0.1% and 0.2% on arrhythmia and atrial fibrillation, which do not depend on absolute level.

The demonstrated scope is label-free pretraining on three Seoul National University Hospital cohorts and 19 tasks on a Korean operating-room cohort and US ICU cohorts drawn mostly from MIMIC-III, with all results from a frozen encoder plus a linear probe.

The authors note that only two downstream tasks take central venous, pulmonary artery or intracranial pressure as input and both have few patients; in external validation the Korean cohorts are limited to the operating room and most US ICU tasks come from a single database collected in 2001-2012; some labels derive from nursing records or diagnosis codes and may carry noise and timing errors; the admission outcome tasks are weakest, with even the best model reaching an AUROC of only 0.63-0.66 on cardiac arrest and mechanical ventilation.