Auto-MD-bench: task packages

Auto-MD-bench evaluates autoresearch agents that build prediction models for medical tasks. An agent receives a task description and labeled training data, develops a model on its own within a fixed time budget, and submits predictions for a hidden test set. Submissions are scored on the task's primary metric, on risk calibration and on performance under population shift.

This repository holds the ready-built packages of the 10 benchmark tasks that use public data, plus the two objective variants of uci_diabetes used in the calibration experiment. The other 5 tasks use MIMIC-IV, eICU-CRD and SEER, which cannot be redistributed; build them from the scripts in the code repository after obtaining access.

Paper: https://arxiv.org/abs/2610.25069

Code, agents, grader and paper results: https://github.com/Edwin0080/Auto-MD-bench

Tasks in this repository

Task	Data	Outcome	Validation	Primary metric	Train / test
support	SUPPORT	Death	Internal	Ctd	7,098 / 1,775
metabric	METABRIC	Death	Internal	Ctd	1,523 / 381
rotterdam	Rotterdam	Recurrence or death	Internal	Ctd	2,385 / 597
rotterdam_gbsg	Rotterdam → GBSG	Recurrence or death	External (cohort)	Ctd	2,982 / 686
nhanes	NHANES	Death	Internal	Ctd	11,639 / 2,910
uci_diabetes	UCI Diabetes, 130 US hospitals	30-day readmission	Internal	AUROC	81,347 / 20,419 encounters
physionet2012	PhysioNet/CinC 2012, ICU	In-hospital death	Internal	min(Se, PPV)	4,000 / 8,000 stays
sepsis	PhysioNet/CinC 2019, ICU	Sepsis onset	Internal	Utility	32,268 / 8,068
sepsis_new_hospital	PhysioNet/CinC 2019, ICU	Sepsis onset	External (hospital)	Utility	20,336 / 4,000
sepsis_500_labels	PhysioNet/CinC 2019, ICU	Sepsis onset	External (hospital)	Utility	20,336 + 500 / 4,000

uci_diabetes_brier and uci_diabetes_ece have the same data as uci_diabetes; only the stated objective in description.md differs (Brier score; AUROC and ECE).

Not here (build locally): seer, seer_later_period (SEER Research Data Agreement), mimiciv, eicu, eicu_new_hospitals (PhysioNet credentialed access).

Layout

<task>/
  public/    what the agent sees: description.md, training data, test features,
             sample_submission.csv, validate.py (format checker)
  private/   hidden test labels, used only by the grader

Give an agent only public/. The data are the same as in the paper's experiments.

Usage

pip install -U huggingface_hub
git clone https://github.com/Edwin0080/Auto-MD-bench && cd Auto-MD-bench
pip install -e .
hf download Edwin0080/Auto-MD-bench --repo-type dataset --local-dir packages
python -m automdbench.grade support packages/support/public/sample_submission.csv

See the code repository for running agents and for the scoring protocol.

Data sources and licenses

Each task keeps the license of its source data. Cite the sources below when you use the tasks.

Tasks	Source	License or terms
support	SUPPORT (Knaus et al., 1995), as processed by DeepSurv (Katzman et al., 2018) from the Vanderbilt Biostatistics data repository	"Permission is granted to anyone wishing to use the data sets"; data obtained from https://hbiostat.org/data courtesy of the Vanderbilt University Department of Biostatistics
metabric	METABRIC (Curtis et al., 2012; Pereira et al., 2016; Rueda et al., 2019), as processed by DeepSurv; the same values are in cBioPortal brca_metabric	ODC Open Database License (ODbL) 1.0, as distributed by cBioPortal
rotterdam, rotterdam_gbsg	Rotterdam tumour bank and GBSG trial (Royston and Altman, 2013), as distributed in the R package survival	LGPL (≥ 2), the license of the survival package
nhanes	NHANES 2011–2016 and NCHS public-use linked mortality files	U.S. public-use data; NCHS data use restrictions apply (no attempt to identify individuals)
uci_diabetes*	Diabetes 130-US Hospitals for Years 1999–2008, UCI Machine Learning Repository (Strack et al., 2014)	CC BY 4.0
physionet2012	PhysioNet/CinC Challenge 2012 (Silva et al., 2012; Goldberger et al., 2000)	Open Data Commons Attribution License 1.0
sepsis*	PhysioNet/CinC Challenge 2019 (Reyna et al., 2020; Goldberger et al., 2000)	CC BY 4.0

The sepsis* packages include utility_score.py, the official PhysioNet/CinC Challenge 2019 scoring functions (BSD 2-Clause). The validate.py format checkers are part of the benchmark code (MIT).

Citation

@misc{zhou2026automdbenchevaluatingautoresearchagents,
      title={Auto-MD-bench: Evaluating Autoresearch Agents for Real-World Medical Tasks},
      author={Shenqi Zhou and Qiran Zou and Wenhao Zhao and Zhenyu Liao and Dianbo Liu},
      year={2026},
      eprint={2610.25069},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2610.25069},
}