ehr2rl

ehr2rl turns MIMIC-IV-style electronic health record data into datasets ready for offline reinforcement learning research. It loads longitudinal EHR tables, represents patient trajectories, shapes rewards, estimates observed clinician behavior, and exports data for tools such as d3rlpy.

pip install ehr2rl

Synthetic quickstart

This example needs only the base install. No MIMIC-IV access or cloud credentials are required. The full quickstart explains each step.

from ehr2rl import BehaviorPolicy, MortalityReward, make_synthetic_dataset

ds = make_synthetic_dataset(n_patients=25, trajectory_length=24, seed=7)
policy = BehaviorPolicy().fit(ds)
ds = MortalityReward().shape(ds)

print(f"{len(ds)} trajectories")                    # 25 trajectories
print(ds[0].states.shape)                           # (24, 4)
print(policy.predict_proba(ds[0].states).shape)     # (24, 2)

Where to go next

Researchers

Build datasets for offline RL experiments.

• Installation and optional extras

• MIMIC-IV v3.1 through BigQuery, with cost guards

• Actions, rewards, behavior policy, and export

• Dataset model, provenance, and limitations

Contributors

Maintain and extend the library.

• Development setup and quality checks

• Architecture and data flow

• Testing conventions

• Adding features, actions, rewards, and exporters

Both paths share one API reference.

GitHub · PyPI · Issues

Important

ehr2rl is research infrastructure. It is not clinical decision support and does not validate treatment recommendations. It does not ship or provide access to MIMIC-IV: obtain clinical data through credentialed channels such as PhysioNet and follow the applicable data use agreements.