Foundation model

CARMEN

cardiorespiratory foundation model pretrained on ten perioperative/ICU waveforms (SNUH)

PreprintCatalog paper date: 2026Added to atlas: Nov 13, 2026Source

Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.

Single Transformer encoder pretrained on ten OR/ICU waveforms (2.27M modality-hours, SNUH). With PPG alone it matched or beat AnyPPG, PaPaGei and Pulse-PPG on most VitalDB tasks. Code and weights are public but under PolyForm Noncommercial.

Signals
PPG, ECG, BP, Resp, AWP, CVP, ICP, PAP
Year
2026
Architecture

Single Transformer encoder over any combination of 10 co-recorded signals, 10 s–5 min; all signals resampled to 100 Hz, patch 25 samples (0.25 s); LSCNorm conditioning restores absolute signal level (not used for PPG / RESP_Impedance)

Form factor

bedside OR/ICU monitor waveforms (finger/clinical PPG, not wrist)

License

PolyForm Noncommercial 1.0.0 (repository LICENSE); commercial use needs a separate licence from the authors

At a glance
Type
Foundation model
Parameters
Not reported
Architecture
Single Transformer encoder over any combination of 10 co-recorded signals, 10 s–5 min; all signals resampled to 100 Hz, patch 25 samples (0.25 s); LSCNorm conditioning restores absolute signal level (not used for PPG / RESP_Impedance)
Venue
arXiv 2611.15004v1 (cs.LG), 2026-11-11; README: accepted to GenAI4Health Workshop at NeurIPS 2026
Year
2026
DOI
Not reported
Licence
PolyForm Noncommercial 1.0.0 (repository LICENSE); commercial use needs a separate licence from the authors
Caveats
Not reported
Added
Nov 13, 2026
Updated
Nov 13, 2026

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.

Verbatim fromarXiv· arXiv metadata (abstracts are released under CC0)

CARMEN: ChatPPG research summary card

How it works

How CARMEN works: Input signals → Foundation model → Downstream tasksInput signals: PPG (Photoplethysmogram); ECG (Electrocardiogram); ABP (Arterial pressure); Respiration (Breathing effort). Datasets: VitalDB, VitalDB Arrhythmia, MIMIC-III Waveform Database v1.0 (PhysioNet), MIMIC-III-Ext-PPG, PhysioNet/CinC Challenge 2018: arousal detection from overnight PSG (MGH sleep lab, 1,985 subjects), Aurora-BP. Foundation model: CARMEN (Transformer).Input signalsPPG: PhotoplethysmogramPPGPhotoplethysmogramECG: ElectrocardiogramECGElectrocardiogramABP: Arterial pressureABPArterial pressureRespiration: Breathing effortRespirationBreathing effortDatasets: VitalDB, VitalDB Arrhythmia, MIMIC-III Waveform Database v1.0 (PhysioNet), MIMIC-III-Ext-PPG, PhysioNet/CinC Challenge 2018: arousal detection from overnight PSG (MGH sleep lab, 1,985 subjects), Aurora-BPDatasetsVitalDB, VitalDB Arrhythmia +4 moreFoundation model×NCARMENTransformer

Datasets linked in the hub; pretraining vs evaluation roles follow the paper.

Architecture
Single Transformer encoder over any combination of 10 co-recorded signals, 10 s–5 min; all signals resampled to 100 Hz, patch 25 samples (0.25 s); LSCNorm conditioning restores absolute signal level (not used for PPG / RESP_Impedance)
Inputs
PPG, ECG, BP, Resp, AWP, CVP, ICP, PAP
Approach
foundation model · multimodal · icu · perioperative · ppg · ecg

Devices· 1

Full photo credits, sources and licences are on each device page.

Datasets used· 5

Cite this model

Cite the publication recorded for this model. Check it against the publisher before submitting.

CARMEN. (2026). arXiv 2611.15004v1. https://arxiv.org/abs/2611.15004
@misc{carmen2026carmen,
  title         = {{CARMEN}},
  year          = {2026},
  howpublished  = {arXiv 2611.15004v1},
  eprint        = {2611.15004},
  archivePrefix = {arXiv}
}

Source metadata: 2026

Record updated Nov 13, 2026

History

  1. Added to the ChatPPG research hub