ehr2rl
A Python library for converting MIMIC-IV electronic health record data into offline reinforcement learning datasets, filling the preprocessing gap between raw EHR data and offline RL training.
Timeline (last 90 days)
ehr2rl is a library written in Python for converting MIMIC-IV electronic medical record data into offline reinforcement learning datasets.
ehr2rl builds hourly patient trajectories from MIMIC-IV v3.1 via BigQuery, with each trajectory containing states, actions, rewards and termination flags.
Every ehr2rl query first performs a dry run and sets a hard maximum_bytes_billed cap to prevent unexpectedly high bills.
ehr2rl extracts vasopressor and fluid doses from ICU input events as actions and discretizes them according to user-defined bin boundaries.
ehr2rl supports replaceable reward functions, including mortality, per-step SOFA change, 30-day readmission, or a weighted combination of these rewards.
ehr2rl provides the capability to estimate the observed clinician behavior policy from data, for offline policy evaluation.
ehr2rl outputs directly as a d3rlpy dataset, and can optionally generate a JSON sidecar file recording the data provenance.
The ehr2rl project provides a synthetic MIMIC-style data generator, so that users without MIMIC access can also try it out.
The authors explicitly state that ehr2rl is a research tool and is not intended for treatment decisions.
In action discretization, the dose binning boundaries for vasopressors and fluids directly affect the results of policy learning, and the current binning approach is more heuristic.
1 more timeline events
All Facts (11)
ehr2rl is a library written in Python for converting MIMIC-IV electronic medical record data into offline reinforcement learning datasets.
50%Unverifiedehr2rl builds hourly patient trajectories from MIMIC-IV v3.1 via BigQuery, with each trajectory containing states, actions, rewards and termination flags.
50%UnverifiedEvery ehr2rl query first performs a dry run and sets a hard maximum_bytes_billed cap to prevent unexpectedly high bills.
50%Unverifiedehr2rl extracts vasopressor and fluid doses from ICU input events as actions and discretizes them according to user-defined bin boundaries.
50%Unverifiedehr2rl supports replaceable reward functions, including mortality, per-step SOFA change, 30-day readmission, or a weighted combination of these rewards.
50%Unverifiedehr2rl provides the capability to estimate the observed clinician behavior policy from data, for offline policy evaluation.
50%Unverifiedehr2rl outputs directly as a d3rlpy dataset, and can optionally generate a JSON sidecar file recording the data provenance.
50%UnverifiedThe ehr2rl project provides a synthetic MIMIC-style data generator, so that users without MIMIC access can also try it out.
50%UnverifiedThe authors explicitly state that ehr2rl is a research tool and is not intended for treatment decisions.
50%UnverifiedIn action discretization, the dose binning boundaries for vasopressors and fluids directly affect the results of policy learning, and the current binning approach is more heuristic.
50%Unverifiedehr2rl chooses BigQuery as its data access layer, lowering the barrier to local storage and processing, but also binding users to the Google Cloud ecosystem.
50%