title	Medical Triage OpenEnv
emoji	🩺
colorFrom	blue
colorTo	green
sdk	docker
app_port	7860
pinned	false

🩺 Clinical Triage Decision Environment (OpenEnv)

🚀 Overview

This project implements a deterministic reinforcement learning environment that simulates real-world clinical triage decision-making.

The agent is tasked with:

• Assessing patient symptoms
• Determining severity level
• Recommending the appropriate next action

⚠️ This is NOT a medical diagnosis system. It is a structured simulation designed for safe and testable decision-making.

---

🎯 Objective

Given patient data, the agent must:

1. Classify severity:

   • low
   • medium
   • high
   • emergency

2. Recommend next action:

   • self_care
   • visit_doctor
   • urgent_care
   • go_to_er

---

🧠 Environment Design

🔁 Multi-Step Interaction

This environment supports multi-step reasoning:

• Step 1: Initial decision based on symptoms
• Step 2: Additional clinical information revealed
• Step 3: Final refined decision

This mimics real-world triage workflows.

---

📥 Observation Space

{
  "symptoms": ["chest pain"],
  "duration": "30 minutes",
  "age": 55,
  "history": ["hypertension"],
  "step": 1
}

---

🎮 Action Space

{
  "severity": "emergency",
  "action": "go_to_er"
}

---

🧩 Tasks

🟢 Easy — Basic Triage

• Clear and obvious symptoms
• Direct mapping to severity and action

🟡 Medium — Conflicting Signals

• Mixed symptoms and patient history
• Requires contextual reasoning

🔴 Hard — Multi-Step Critical Reasoning

• Evolving symptoms across steps
• Requires updating decisions
• Penalizes unsafe underestimation

---

🏆 Reward Design

Component	Reward
Correct severity	+0.4
Correct action	+0.4
Early correct guess	+0.1
Improvement after update	+0.2
Underestimation	-0.6
Overestimation	-0.2

👉 Emphasis is placed on safety and risk-aware decisions.

---

🧪 Grading

Final score is normalized to 0.0 – 1.0:

• 1.0 → Perfect decision
• 0.7 → Partially correct
• 0.3 → Risky decision
• 0.0 → Unsafe outcome

---

🤖 Baseline Agent

A simple rule-based agent is provided:

• Uses symptom matching
• Updates decisions after new information
• Demonstrates partial reasoning capability

---

⚙️ Setup & Run

1. Install dependencies

pip install -r requirements.txt

2. Run API

uvicorn app.main:app --reload --port 7860

3. Open API docs

http://localhost:7860/docs

---

🐳 Docker

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

---

📊 Baseline Score

Typical baseline performance:

~0.55 – 0.70

---

🌍 Motivation

Clinical triage is a high-stakes decision-making process where:

• Underestimating severity can be dangerous
• Overestimating leads to resource misuse

This environment captures these trade-offs in a safe, deterministic framework.

---

✅ Key Features

• Deterministic and fully testable
• Multi-step reasoning environment
• Risk-sensitive reward design
• Real-world inspired task simulation
• OpenEnv compliant

---

🏁 Conclusion

This project demonstrates how reinforcement learning environments can model real-world decision-making systems while remaining:

• Safe
• Interpretable
• Evaluatable

---