A Cardiorespiratory Foundation Model Pretrained on Ten Physiological Signals for Perioperative and Critical Care
Abstract
Arterial, venous, pulmonary artery and intracranial pressures and ventilator waveforms guide decisions in operating rooms and intensive care units, yet foundation models for physiological waveforms have been built almost entirely on the electrocardiogram and photoplethysmogram. We present CARMEN, a single encoder pretrained without labels on ten cardiorespiratory waveforms (2.27 million modality-hours). CARMEN accepts any combination of co-recorded signals from 10 s to 5 min, and a conditioning mechanism, LSCNorm, restores the absolute signal level that normalization removes; without it, the AUROC for intraoperative hypotension fell from 0.894 to 0.780. Across 19 tasks from Korean and US cohorts, the frozen encoder outperformed supervised models on 12 of 15 and three foundation models, one general-purpose and two for biosignals, on 14 to 17 of 19, with none exceeding it on any task. Given the electrocardiogram or photoplethysmogram alone, it matched or exceeded signal-specific foundation models in most comparisons, and it retained its performance when channels were missing.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
CARMEN is a multimodal foundation model for cardiorespiratory signals that introduces LSCNorm to preserve absolute physiological levels and employs a two-phase curriculum to learn cross-modal coupling. The paper demonstrates significant improvements over existing foundation models and supervised baselines across 19 clinical tasks, particularly for invasive hemodynamic signals, establishing a new standard for biosignal foundation models in critical care.
Methodology
The paper introduces CARMEN, a multimodal foundation model for cardiorespiratory signals. The core architectural contribution is LSCNorm (Loc/Scale-conditioned Layer Normalization), which re-injects the mean and standard deviation removed by instance normalization into every transformer layer via FiLM-style conditioning. This is a clever and necessary design choice for physiological signals where absolute amplitude (e.g., blood pressure levels) carries critical clinical meaning, unlike standard normalized representations. The pretraining strategy employs a two-phase curriculum: Phase 1 focuses on intra-modality representation learning using masked patch modeling and block next prediction, while Phase 2 introduces cross-modal coupling via Cross-Modal Patch Modeling (CMPM). This curriculum prevents the model from taking shortcuts by relying on synchronized signals before learning individual morphologies. The use of a shared patch-based encoder for heterogeneous signals (ECG, ABP, PPG, CVP, PAP, ICP, etc.) is a significant step forward from single-signal models.
Results & evaluation
The evaluation is extensive, covering 19 downstream tasks across clinical event detection, prediction, outcome estimation, and phenotyping. The model is compared against supervised baselines (1D-ResNet, scratch training) and three other foundation models (MOMENT, NormWear, SPOTR). CARMEN demonstrates superior performance, particularly on tasks involving invasive hemodynamic signals where existing FMs have no pretraining data. The robustness to missing channels is a strong practical feature, validated by ablation studies. The comparison with 11 specialized ECG/PPG foundation models on shared modalities shows CARMEN matching or exceeding them, validating the benefit of multimodal pretraining. The inclusion of a US cohort (MIMIC-III) for external validation strengthens the generalizability claims, although the pretraining data is from a single Korean institution.
Reproducibility
The authors provide a GitHub repository with code and pretrained weights (Small, Base, Large). The data availability statement is transparent, distinguishing between public datasets (MIMIC, PhysioNet) and restricted hospital data. While the pretraining data is not public, the release of weights and code allows for partial reproduction and fine-tuning on other datasets. The detailed description of the pretraining objectives, loss functions, and data augmentation strategies enhances reproducibility for researchers with access to similar data.
Limitations
The primary limitation is the reliance on a single institution for pretraining (SNUH), which may limit generalizability to diverse hospital settings and equipment, although the US cohort results suggest some transferability. The downstream tasks for invasive signals (CVP, PAP, ICP) have small sample sizes due to data scarcity. The model is evaluated primarily with linear probes; fine-tuning results are not reported, which might yield higher performance but obscures the full potential of the representation. The use of a deceased-patient cohort (K-MIMIC-MORTAL) for pretraining is ethically and methodologically interesting but may bias the model toward pathological states.
Broader impact
This paper addresses a critical gap in biomedical AI: the lack of foundation models for invasive hemodynamic and respiratory signals, which are central to ICU and OR decision-making. By demonstrating that a single model can handle diverse signal combinations and preserve absolute physiological levels, CARMEN provides a template for developing clinically useful FMs in other domains where signal amplitude is meaningful. The work bridges the gap between data availability (non-invasive ECG/PPG) and clinical need (invasive pressures), potentially accelerating the deployment of AI in critical care monitoring. CARMEN is a multimodal foundation model for cardiorespiratory signals that introduces LSCNorm to preserve absolute physiological levels and employs a two-phase curriculum to learn cross-modal coupling. The paper demonstrates significant improvements over existing foundation models and supervised baselines across 19 clinical tasks, particularly for invasive hemodynamic signals, establishing a new standard for biosignal foundation models in critical care.
Author of this paper? Send a correction.