ICareFM

A Foundation Model for Intensive Care Unlocking Generalization across Tasks and Domains at Scale

This is the root repository for ICareFM. We provide an overview of the associated resources.

News

• April 2. 2026: Full preprint of ICareFM model and dataset released on medRxiv: A Foundation Model for Intensive Care: Unlocking Generalization across Tasks and Domains at Scale
• December 1. 2025: Spotlight Talk and Best Findings Paper Award at ML4H in San Diego
• August 20. 2025: Preview version of 'A Foundation Model for Intensive Care: Unlocking Generalization across Tasks and Domains at Scale' posted on medRxiv
• December 14. 2024: Best Paper Award at AIM-FM @ NeurIPS for Towards Foundation Models for Critical Care Time Series

Data

We are working hard to create an easy to access and one-click-download version of our large-scale multi-center harmonized dataset available on the Physionet platform. Watch this space for news on this development and effort.

For data harmonization we use the ricu tool written in the R programming language:

• We are working on a python version of the ricu tool: {tbd}
• Our extended fork of the ricu data harmonization tool: {tbd}
• Original data harmonization tool ricu source repository: https://github.com/eth-mds/ricu

Model

Experiments and Results

Publications

• Preprint manuscript of our foundation model and dataset: A Foundation Model for Intensive Care: Unlocking Generalization across Tasks and Domains at Scale
• Early Workshop Paper highlighting data harmonization efforts and large benchmarking: Towards Foundation Models for Critical Care Time Series

Citations

The medRxiv preview of A Foundation Model for Intensive Care: Unlocking Generalization across Tasks and Domains at Scale:

@article {burger2026icarefm,
	author = {Burger, Manuel and Chopard, Daphne and Lichtner, Gregor and Londschien, Malte and Sergeev, Fedor and Fuchs, Moritz and Yeche, Hugo and Kuznetsova, Rita and Faltys, Martin and Gerdes, Eike and Leshetkina, Polina and Christ, Micha and Schanz, Moritz and Goebel, Nora and Buehlmann, Peter and Gruenewald, Elias and Balzer, Felix and Raetsch, Gunnar},
	title = {A Foundation Model for Intensive Care: Unlocking Generalization across Tasks and Domains at Scale},
	year = {2026},
	doi = {10.1101/2025.07.25.25331635},
	publisher = {Cold Spring Harbor Laboratory Press},
	URL = {https://www.medrxiv.org/content/early/2026/04/02/2025.07.25.25331635},
	eprint = {https://www.medrxiv.org/content/early/2026/04/02/2025.07.25.25331635.full.pdf},
	journal = {medRxiv}
}

Workshop Paper highlighting early progress on creating a large-scale harmonized critical care dataset Towards Foundation Models for Critical Care Time Series:

@misc{burger2024foundationmodelscriticalcare,
      title={Towards Foundation Models for Critical Care Time Series}, 
      author={Manuel Burger and Fedor Sergeev and Malte Londschien and Daphné Chopard and Hugo Yèche and Eike Gerdes and Polina Leshetkina and Alexander Morgenroth and Zeynep Babür and Jasmina Bogojeska and Martin Faltys and Rita Kuznetsova and Gunnar Rätsch},
      year={2024},
      eprint={2411.16346},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.16346}, 
}

Please also cite:

If you use our work on data harmonization please consider citing the original authors work on the ricu package:

@article{bennett2023ricu,
  title={ricu: R’s interface to intensive care data},
  author={Bennett, Nicolas and Ple{\v{c}}ko, Drago and Ukor, Ida-Fong and Meinshausen, Nicolai and B{\"u}hlmann, Peter},
  journal={GigaScience},
  volume={12},
  pages={giad041},
  year={2023},
  publisher={Oxford University Press}
}