RealICU

Paper · Code · Project page

This repository contains a synthetic sample, not the benchmark data. The files here describe 5 synthetic patients in exactly the same file layout, Parquet schemas, and label JSON format as the full dataset. They contain no patient data: every identifier, date, measurement, note, and label was generated from hand-written scenarios, not from MIMIC records. Use the sample to inspect the format and to test data pipelines; do not use it to evaluate models.

The full dataset is distributed through PhysioNet. RealICU is derived from MIMIC-IV, so it is released under PhysioNet credentialed access: a credentialed PhysioNet account, completed human-subjects research training (CITI), and a signed data use agreement. It has been submitted to PhysioNet as MIMIC-IV-Ext-RealICU and is under review. The PhysioNet link will be added here once it is published.

RealICU is a benchmark for evaluating language models and agents on ICU decision support. For each evaluation window, it provides clinical events for constructing model inputs and aligned reference labels for four tasks: patient status, acute problems, recommended actions, and red flag actions.

The full dataset (version 1.0.0) contains 94 MIMIC-derived patients, one target ICU stay per patient, and 2,989,158 clinical events. The trajectories include available records before, during, and after the target ICU stay.

Annotation set	Reference source	Windows	Action ratings
RealICU-Gold	Paired physician consensus on Oracle candidates	930	2,189
RealICU-Scale	Oracle, an LLM hindsight annotator	11,862	25,387

Gold and Scale share the same patients and have partially overlapping windows; Gold is not a window subset of Scale. Both are evaluation sets and should not be used as separate training and test cohorts. Windows generally span 30 minutes, with shorter intervals at the end of a stay when needed.

Synthetic sample in this repository

Content	Sample	Full dataset
Patients (one target ICU stay each)	5	94
Clinical events	1,743	2,989,158
Gold windows / action ratings	35 / 46	930 / 2,189
Scale windows / action ratings	198 / 219	11,862 / 25,387

The five synthetic stays cover septic shock, hypoxemic respiratory failure, diabetic ketoacidosis, upper gastrointestinal bleeding, and decompensated heart failure. They also reproduce the format's edge cases: a truncated final window, a Gold window with an unavailable patient-status label, a Scale window without acute-problem references, an insufficient_data window, an ICU readmission, and in-hospital deaths after ICU discharge. Synthetic identifiers start with 9 (for example 90000001_99000001) and do not overlap MIMIC-IV identifier ranges.

Quick start

The standalone loader requires Python 3.10 or later, pandas, and pyarrow:

python -m pip install pandas pyarrow huggingface_hub
hf download leoshen/RealICU --repo-type dataset --local-dir realicu
cd realicu

Load one Gold window:

from loader import RealICULoader

loader = RealICULoader(".", "gold", "90000001_99000001")
sample = loader.load_window(
    loader.window_indices[0],
    icu_history_hours=None,
    pre_icu_history_hours=48,
    excluded_event_codes=("NOTE_DISCHARGESUMMARY", "META_DEATH"),
)

model_input = sample["input"]
labels = sample["labels"]
current_events = model_input["current_events"]
references = labels["reference_labels"]
availability = labels["label_availability"]

print(current_events.head())
if availability["patient_status"]:
    print(references["patient_assessment"]["overall"]["label"])

Use "scale" instead of "gold" to read Oracle annotations. The loader reads a stay once; call load_window for any ID in loader.window_indices. These IDs may be nonconsecutive. List available stays with pandas.read_parquet("trajectories/icu_stay.parquet")["stay_id"].

Each sample has two parts:

• input: sample and stay identifiers, actual window timestamps, and the three event tables below.
• labels: the four tasks under reference_labels, plus label_availability indicating which references can be used for evaluation.

Input event table	Included observations
pre_icu_events	Before target ICU admission, within the chosen pre-ICU lookback
history_events	During the target ICU stay, before the window, within the chosen ICU lookback
current_events	Within the actual half-open interval [window_start_time, window_end_time)

Both lookbacks and excluded_event_codes are required choices. Use a number of hours, None for all available history, or 0 for no history. The pre-ICU lookback is measured from ICU admission; the ICU lookback is measured from the window start. An empty tuple retains all event categories. The example's 48-hour lookback and exclusions are illustrative, not benchmark defaults.

The same example is available from the command line:

python loader.py --root . --label-set gold --stay-id 90000001_99000001 \
  --window-index 0 --icu-history-hours all --pre-icu-history-hours 48 \
  --excluded-event-codes NOTE_DISCHARGESUMMARY META_DEATH

The prediction and evaluation code in RealICU-Bench reads this layout directly; set data.root to the downloaded directory.

Tasks and reference labels

The paths below are relative to reference_labels:

Task	JSON path	Reference content
Patient Status	patient_assessment.overall	stable, improving, deteriorating, or insufficient_data, with a rationale
Acute Problems	patient_assessment.active_risks	Active clinical problems or emerging risks, with supporting evidence
Recommended Actions	action_review.evaluations	Recorded actions rated best_practice, acceptable, potentially_harmful, or insufficient_data
Red Flag Actions	action_review.red_flags	Actions to avoid, with a rationale and supporting evidence

Actions rated best_practice or acceptable form the positive reference set for recommendation evaluation. Every retained action refers to an event within its assigned window. Numeric action and evidence IDs refer to icu_event_index within the same target stay.

Check label_availability for each task before scoring. Its keys are patient_status, active_risks, recommended_action, and red_flags. A false value means the reference is unavailable. An unavailable patient-status label is stored as unavailable; this is not a clinical class. An available insufficient_data label remains a valid reference. An available empty list means no retained reference items, not a comprehensive finding that no clinical problems or contraindications exist.

In the full dataset, Gold contains physician-consensus labels and selected items from Oracle candidates; its rationale and key_evidence were retained from Oracle. Scale labels and explanations are Oracle outputs. Annotations use retrospective context, so evidence may refer to events before or after the observation window, or to pre-ICU history and discharge summaries.

Files and alignment

README.md
loader.py
croissant.json
trajectories/
    events.parquet
    icu_stay.parquet
gold/labels/<stay_id>.json
scale/labels/<stay_id>.json

events.parquet contains one clinical event per row. icu_stay.parquet contains one target ICU stay per row, including admission/discharge times, trajectory bounds, event counts, and outcomes. Join trajectories and annotations using stay_id, formatted as <subject_id>_<icu_stay_id>. An event's own icu_stay_id may refer to a different encounter; use the added stay_id for joins.

Each annotation set has one JSON file per target stay. The file contains stay identifiers, trajectory_metadata, and a window_outputs list. Each window has sample_id, window_index, window_key, window_metadata, reference_labels, and label_availability. Use the actual timestamps in window_metadata to align events; window_key gives nominal hours relative to ICU admission.

Event identifier	Meaning
event_idx, source_event_index	Original source event identifier
event_index	Zero-based position in the published trajectory
icu_event_index	Zero-based chronological event ID within the target ICU stay; null outside it

croissant.json provides Croissant 1.1 metadata with field descriptions and Responsible AI documentation. Its record sets expose ICU stays, events, windows, action evaluations, active risks, and red flags; each label record set covers Gold and Scale, distinguished by label_set. Its file URLs point to this repository, so it can be loaded directly with mlcroissant.

Evaluation input construction

Keep reference answers and retrospective outcomes out of model inputs. For Action Recommendation, the paper removes recorded actions from the current window before prediction to prevent answer leakage (Appendix A.1); the loader does not perform this masking automatically.

Citation

@article{shen2026realicu,
  title={RealICU: Do LLM Agents Understand Long-Context ICU Data? A Benchmark Beyond Behavior Imitation},
  author={Shen, Chengzhi and Shen, Weixiang and Susetzky, Tobias and Chen, Chen (Cherise) and Li, Jun and Liu, Yuyuan and Zhang, Xuepeng and Gong, Zhenyu and Rueckert, Daniel and Pan, Jiazhen},
  journal={arXiv preprint arXiv:2605.13542},
  year={2026}
}