AkashKaralingannavar/sequential-medical-triage-ai

🏥 AI Medical Triage Assistant — OpenEnv RL Environment

Real-World Motivation

In low-resource healthcare settings worldwide, patients face critical delays in receiving appropriate care. This OpenEnv simulates a first-level AI triage assistant that must make sequential decisions under partial observability — a genuine real-world problem.

Unlike simple classifiers, this is a true sequential decision-making RL environment where the agent must:

• —Decide what information to gather (at a cost)
• —Balance investigation thoroughness vs. time pressure
• —Handle patients who deteriorate if not triaged quickly

Observation Space (TriageObservation)

Action Space (TriageAction)

Reward Design

• —Urgency match: up to 0.45 points (partial credit for being 1 level off)
• —Action match: up to 0.45 points (partial credit for adjacent actions)
• —Investigation bonus: +0.05 for thorough information gathering
• —Time penalty: -0.06 per investigation step (-0.10 in time-critical)
• —Deterioration penalty: -0.10 if volatile patient worsens
• —All scores clamped to strict (0.01, 0.99) — never 0.0 or 1.0

Tasks & Difficulty

1. direct_triage — Easy

Full patient information (symptoms + vitals) is visible from the start. The agent simply needs to classify urgency and recommend an action.

2. investigative_triage — Medium

Only the chief complaint is visible. The agent must spend investigate actions to uncover symptoms and vitals. Each investigation incurs a time penalty (-0.06). Blind guessing is penalized.

3. time_critical_triage — Hard

High-volatility patients deteriorate over time. Investigation costs more (-0.10 per step). If the agent takes too long, the patient's condition worsens, potentially changing the correct answer. The agent must triage quickly while still gathering minimum critical info.

Setup & Usage

Docker (recommended)

docker build -t openenv-triage .
docker run -p 7860:7860 openenv-triage

Access the API at http://localhost:7860/

API Endpoints

• —GET / — Health check (returns 200)
• —GET/POST /reset?task_name=direct_triage&case_index=0 — Reset environment
• —POST /step — Submit an action (JSON body matching TriageAction)
• —GET /state — Get current environment state

Run Inference

export HF_TOKEN="your_api_key"
export MODEL_NAME="Qwen/Qwen2.5-72B-Instruct"
python inference.py

Run a specific task

export TASK_NAME="investigative_triage"
python inference.py

Baseline Scores (Heuristic Agent)

Scores improve significantly with capable LLMs (GPT-4, Qwen-72B).

Grader Verification

python grader.py

All scores will be printed and verified to be in the strict (0.0, 1.0) range.