Skip to content
Modelwire
Subscribe

Foundation model for ten cardiorespiratory waveforms outperforms supervised baselines in OR and ICU

Source published ·Modelwire updated

Original coverage: arXiv cs.LG ↗·How Modelwire adds context

The development

CARMEN represents a shift in foundation model design for clinical settings, moving beyond the ECG-dominated paradigm to encode ten cardiorespiratory waveforms simultaneously. The model's key innovation, LSCNorm, preserves absolute signal magnitudes that standard normalization discards, a detail that swung intraoperative hypotension detection from 0.780 to 0.894 AUROC. Across 19 clinical tasks spanning Korean and US datasets, the frozen encoder outperformed both supervised baselines and competing foundation models on most benchmarks. This work signals growing maturity in multimodal physiological AI for perioperative and ICU contexts, where real-time decision support at scale remains underexplored compared to imaging or text domains.

Modelwire’s AI-generated summary of coverage from arXiv cs.LG.

Modelwire analysis

Explainer

Our AI-generated reading of the wider context and the next developments to watch.

CARMEN's real contribution isn't multimodal encoding itself, but that LSCNorm recovers clinically actionable signal structure that standard normalization destroys. The 0.114 AUROC jump on hypotension detection hinges entirely on this one design choice, not on scale or breadth of waveforms.

This work sits at the intersection of two recent threads in Modelwire coverage. Like the single-cell foundation model (scOntoFM, Nov 7) that embedded hierarchical biological structure into representation learning, CARMEN bakes domain-specific inductive bias into its normalization layer rather than relying on generic preprocessing. More directly, it extends the clinical AI trajectory from ICare-DT (Nov 6), which showed that continuous state refinement improves ICU forecasting. CARMEN targets the same perioperative and critical care setting but takes a foundation model approach, freezing the encoder and testing transfer across 19 tasks rather than retraining per-cohort. The causal graph fusion work (Nov 6) also operates in fragmented clinical data, but CARMEN sidesteps that problem by learning shared representations across multiple signal types upfront.

If CARMEN's frozen encoder maintains its performance advantage when evaluated on prospective ICU data collected after model training (not just held-out retrospective splits), that confirms the learned representations generalize to distribution shift. If performance degrades significantly, the gains may reflect dataset artifacts rather than robust physiological understanding. Watch for external validation announcements from the authors or independent teams within the next 6 months.

This interpretation is generated from the summary above and available source metadata. Our methodology · Report an error

MentionsCARMEN · LSCNorm

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “A Cardiorespiratory Foundation Model Pretrained on Ten Physiological Signals for Perioperative and Critical Care”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

Foundation models excel at forecasting but fail at causal inference

arXiv cs.LG·

Foundation model embeds cell lineage hierarchy into single-cell biology

Google's Walrus unifies speech and audio with data-centric training recipe

arXiv cs.CL·
Foundation model for ten cardiorespiratory waveforms outperforms supervised baselines in OR and ICU · Modelwire