Apppublic

prashant-9457/my-openenv-task

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

App README

🏥 ICU Resource Allocation — OpenEnv

A real-world OpenEnv environment modelling a 20-bed ICU in a 500-bed Indian tertiary-care hospital. An AI agent acts as the ICU charge coordinator, making resource allocation decisions every 30 minutes over a 24-hour duty cycle.

🌍 Real-World Grounding

This environment is not a simulation of a game — it models actual clinical practice:

Element	Real-World Source
SOFA scoring (0–24)	Vincent et al., JAMA 1996 — gold-standard ICU triage
SOFA → mortality	Ferreira et al., JAMA 2001
Nurse:patient ratio	NABH ICU Standard (India) — max 1:2
Bed turnover times	Agnihotri et al., Indian J Crit Care Med 2019
Arrival patterns	Arias-Verdú et al., Critical Care Medicine 2017
Cost calibration	CGHS ICU package rates 2023 (Central Govt Health Scheme)

🎮 Action Space (7 discrete actions)

#	Action	Effect	Cost
0	HOLD	Observe, no change	₹0
1	ADMIT_CRITICAL	Admit highest-SOFA patient	—
2	ADMIT_FIFO	Admit longest-waiting patient	—
3	TRANSFER_OUT	Move stable patient to step-down	—
4	CALL_EXTRA_NURSE	+1 nurse, improves ratio	₹1,200
5	SPECIALIST_CONSULT	−15% mortality risk on sickest patient	₹3,500
6	EXPEDITE_BED	Faster bed turnover	₹600

👁️ Observation Space (27 fields)

Rich, clinically meaningful state:

—Bed status: occupied / available / in turnover
—Queue: total waiting, broken down by SOFA severity (critical/severe/moderate), longest wait
—Patient acuity: avg SOFA score, avg mortality risk of current ICU patients
—Equipment: ventilators & dialysis machines available/in-use
—Staff: nurses on duty, nurse:patient ratio, doctors
—Time: hour, shift (Day/Evening/Night), step
—Budget: remaining INR, utilisation %
—Outcomes: admissions, transfers, deaths in queue, adverse events, wait violations

📋 Tasks

🟢 Easy — Prevent Preventable Deaths

Zero queue deaths + nurse:patient ratio ≤ 2.0 in ≥90% of steps.

🟡 Medium — NABH-Compliant Critical Care

All Easy criteria + all critical patients (SOFA ≥ 11) admitted within 2 hours of arrival. Maps to NABH Grade-B ICU standard.

🔴 Hard — JCI-Grade ICU Excellence

All Medium criteria + zero adverse events + budget ≤ 85% + average SOFA non-increasing. Maps to JCI / NABH Grade-A accreditation benchmarks.

🔌 API

bash

# Reset
curl -X POST http://localhost:7860/reset -H "Content-Type: application/json" -d '{"seed": 42}'

# Step (action 1 = ADMIT_CRITICAL)
curl -X POST http://localhost:7860/step -H "Content-Type: application/json" -d '{"action": 1}'

# State
curl http://localhost:7860/state

🚀 Local Setup

bash

pip install -r requirements.txt

# Run server
python app.py

# Run graders (validates all scores in [0,1])
python graders/task_graders.py

# Run LLM inference (needs env vars)
export API_BASE_URL=...
export MODEL_NAME=...
export HF_TOKEN=...
python inference.py

📁 Structure

hospital_env/
├── env.py                 # ICU environment (SOFA, NABH, Poisson arrivals)
├── app.py                 # FastAPI: /reset /step /state
├── inference.py           # LLM agent via OpenAI client
├── openenv.yaml           # OpenEnv spec
├── requirements.txt
├── Dockerfile
├── README.md
└── graders/
    └── task_graders.py    # Easy / Medium / Hard graders → scores [0,1]