These files are model outputs derived from MIMIC-III and MIMIC-IV. They include per-patient rows keyed by SUBJECT_ID, and the chain-of-thought and self-reflection runs contain generated text that restates parts of the source patient record.

MIMIC is distributed by PhysioNet under a credentialed data use agreement. By requesting access you confirm that you will treat these files under the same terms as the source databases: you will not redistribute them, will not attempt to re-identify any individual, and will not use them to train or fine-tune models for redistribution.

Access to the underlying MIMIC databases must be obtained separately from PhysioNet.

ClinicalBench Results

Model outputs behind ClinicalBench: Can LLMs Beat Traditional ML Models in Clinical Prediction? (KDD 2026).

3,015 result files covering 29 LLMs and 12 traditional ML models across three clinical prediction tasks and two databases. They are published so that every number in the paper can be re-scored without re-running inference.

💻 Code: github.com/canyuchen/ClinicalBench
🌐 Website: clinicalbench.github.io
📄 Paper: arXiv:2411.06469

The finding

Traditional ML models still win. Mortality prediction on MIMIC-III, mean F1 over five cohort reshuffles, computed from the files in this repository:

Model	Type	F1
XGBoost	baseline	65.75
SVM	baseline	63.97
LogisticRegression	baseline	63.09
Gemma2-9B	LLM	43.03
Mistral-v0.3-7B	LLM	38.11

The gap holds across model scale, prompting strategy and fine-tuning. See the paper for the full picture.

What is in here

summary.csv        one row per run, with metrics already computed
results/           the raw per-row model outputs
├── length_pred/{mimic3,mimic4}/
├── mortality_pred/{mimic3,mimic4}/
└── readmission_pred/{mimic3,mimic4}/

`summary.csv`

3,006 rows, one per run, so the headline numbers need no download. This is the file the dataset viewer shows.

Column	Meaning
`task`	`length_pred`, `mortality_pred`, `readmission_pred`
`dataset`	`mimic3` or `mimic4`
`model`	checkpoint name, or baseline name
`model_type`	`llm` or `baseline`
`random_index`	`0`–`4` cohort reshuffles, or `6` for the 500-sample cohort
`mode`	`ORI`, `ICL`, `COT`, `RP`, `SR`, `LORA`
`temperature`	set only for the decoding-temperature sweep
`train_ratio`	set only for the training-set scaling runs
`n`	rows scored
`f1`	macro-F1 for length-of-stay, positive-class F1 otherwise
`auroc`	blank when the run recorded no probability column
`n_invalid`, `invalid_rate`	rows where the model produced no parseable answer
`file`	path to the raw file under `results/`

Nine legacy files are absent from summary.csv: four *_withprob.csv and five missing a cohort index in their filename. They predate the current naming scheme and back no published number. The files themselves are included under results/.

`results/`

results/{task}/{dataset}/{task}_result_data_{model}_{index}{mode}{temp}{ratio}.csv

Column	In	Meaning
`SUBJECT_ID`	LLM runs	patient id from the source database
`ANSWER`	all	gold label
`PREDICTION`	all	scored prediction
`ORIGINAL`	LLM runs	raw model output before parsing
`PROB`	logits-scored runs, all baselines	softmax over the answer tokens only

Traditional-baseline files have ANSWER, PREDICTION, PROB only; they are scored positionally against the test split rather than by patient id.

Usage

Quickest look, no download:

import pandas as pd

df = pd.read_csv("hf://datasets/canyuchen/clinicalbench-results/summary.csv")
best = (df.query("task == 'mortality_pred' and dataset == 'mimic3' and mode == 'ORI'")
          .groupby(["model", "model_type"])["f1"].mean()
          .sort_values(ascending=False))
print(best.head())

Full download, then re-score with the benchmark's own tooling:

hf download canyuchen/clinicalbench-results --repo-type dataset --local-dir clinicalbench-results

git clone https://github.com/canyuchen/ClinicalBench && cd ClinicalBench
pip install -e .
python -m clinicalbench.eval.aggregate configs/paper/table_1.yaml \
    --task mortality_pred --dataset mimic3 \
    --result_root ../clinicalbench-results/results

Two things to know before quoting a number

Unparseable answers are scored wrong, not dropped. When a model returns prose or a number outside the label space, the row is recorded as a deliberately incorrect prediction rather than discarded. Otherwise a model could raise its score by refusing the cases it finds hard. summary.csv carries invalid_rate for exactly this reason: a model at 7% invalid is being measured on a different thing than one at 0%.

Two scoring paths extract answers differently. Runs scored from logits read the final character; runs scored from generated text scan backwards for the last valid digit, so a chain-of-thought answer is read from its conclusion. They disagree on output like "1.". Each path keeps the behaviour that produced its published numbers.

Both are documented in docs/methodology.md.

A caveat the paper states: features come from ICD codes, which are administrative rather than purely clinical data. Performance may differ with richer clinical representations.

Provenance and terms

Derived from MIMIC-III v1.4 and MIMIC-IV v3.0, distributed by PhysioNet under a credentialed data use agreement. This repository contains model outputs only, no raw MIMIC tables, but the outputs are patient-level and are gated accordingly. Access to the source databases must be obtained separately from PhysioNet.

Citation

@inproceedings{chen2026clinicalbench,
  title     = {ClinicalBench: Can LLMs Beat Traditional ML Models in Clinical Prediction?},
  author    = {Chen, Canyu and Yu, Jian and Chen, Shan and Liu, Che and Wan, Zhongwei
               and Zhou, Shuang and Luo, Yuan and Zhang, Rui and Bitterman, Danielle S.
               and Wang, Fei and Shu, Kai},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery
               and Data Mining (KDD '26)},
  year      = {2026}
}

Downloads last month: 13

Total file size:

304 MB

Paper for canyuchen/clinicalbench-results

ClinicalBench: Can LLMs Beat Traditional ML Models in Clinical Prediction?

Paper • 2411.06469 • Published Nov 10, 2024 • 17