Apppublic

arpann09/med-reconciliation-hard

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

Medication Reconciliation — OpenEnv Environment

An OpenEnv environment where an AI agent acts as a clinical pharmacist, identifying dangerous discrepancies between a patient's home medications and their hospital discharge prescription.

Why This Matters

Medication reconciliation errors cause ~1.5 million patient injuries per year in the US. They most commonly occur during care transitions — when a patient leaves the hospital and nobody systematically compares the two medication lists. This is a real problem that costs lives and billions in preventable harm annually.

This environment trains and evaluates AI agents on exactly that task — giving the RL/agent community a medically grounded, high-stakes benchmark that doesn't exist anywhere else.

The Task

The agent receives two medication lists:

—Home medications — what the patient was taking before hospitalization
—Discharge medications — what the hospital prescribed on discharge

The agent must identify all discrepancies by issuing flag actions, then call submit when done. Rewards are given at every step — the agent gets signal on each flag, not just at the end.

Action Space

Field	Type	Description
`action_type`	string	One of: `flag_duplicate`, `flag_interaction`, `flag_dose_mismatch`, `flag_missing`, `submit`
`drug_a`	string	Primary drug name involved
`drug_b`	string	Secondary drug name (for duplicate/interaction flags)
`reasoning`	string	Agent's explanation — used for partial credit scoring

Observation Space

Field	Type	Description
`task_id`	string	Current task identifier
`task_difficulty`	string	`easy`, `medium`, or `hard`
`home_medications`	list[dict]	Home med list: `{name, dose, frequency}`
`discharge_medications`	list[dict]	Discharge list: `{name, dose, frequency}`
`flags_submitted`	list[dict]	All flags submitted so far this episode
`step_feedback`	string	Feedback on the last action taken
`total_issues`	int	Total planted issues in this scenario
`issues_found`	int	Correct issues identified so far
`false_positives`	int	Incorrect flags submitted so far
`done`	bool	Whether the episode is complete
`reward`	float	Step reward

Tasks

Easy — Exact Duplicate (1 issue)

Patient on warfarin has it listed twice in discharge. Double dosing a blood thinner = fatal bleeding risk.

Medium — Brand/Generic Duplicate (1 issue)

Patient takes Ultram (brand) at home, hospital prescribes tramadol (generic) — same drug, double dose. Patient is also on sertraline (SSRI), making the duplicate life-threatening via serotonin syndrome.

Hard — Three Hidden Issues (3 issues)

1.Coumadin + aspirin interaction (Coumadin = warfarin, major bleeding risk)
2.Digoxin dose doubled 0.125mg → 0.25mg (narrow therapeutic index, fatal arrhythmia risk)
3.Metoprolol missing from discharge (abrupt beta-blocker withdrawal = heart attack risk)

Reward Function

Event	Reward
Correct flag	+0.30
Partial flag (right drugs, wrong type)	+0.10
False positive	-0.15
Submit bonus	+0.10

Setup

bash

pip install openenv-core pydantic openai
docker build -t med-recon-env .
docker run -p 7860:7860 -e MED_RECON_TASK=easy med-recon-env

Run Inference

bash

export API_BASE_URL="https://router.huggingface.co/v1"
export MODEL_NAME="Qwen/Qwen2.5-72B-Instruct"
export HF_TOKEN="your_token"
export MED_RECON_TASK="all"
python inference.py

Baseline Scores

Task	Model	Score	Steps
Easy	Qwen/Qwen2.5-72B-Instruct	0.300	1
Medium	Qwen/Qwen2.5-72B-Instruct	0.300	1
Hard	Qwen/Qwen2.5-72B-Instruct	0.300	1

Scores normalized to [0.0, 1.0]. Success threshold is 0.5. The hard task has 3 planted issues — a perfect score requires identifying all 3.