MedACE: Medical Agentic Concept Extraction

MedACE is a minimal proof‑of‑concept that demonstrates how to leverage AG2 multi‑agent workflows to extract immune‑related acute kidney injury (irAKI) evidence from de‑identified clinical notes.

Role	Description
AKIExtractor	LLM agent that parses note bundles and produces a strict JSON record (ICI exposure → AKI events → management → outcome + confidence).
ClinicianReviewer	LLM (or human) agent that approves the JSON or requests fixes.
GroupChatManager	Orchestrates a round‑robin chat until the reviewer issues APPROVE.

---

Quick start

# 1. install AG2 with OpenAI support
pip install "ag2[openai]"

# 2. set your key
export OPENAI_API_KEY="sk‑..." 

# 3. run the demo
python run.py 2      # 2 = classic irAKI example, 1 = non‑irAKI

You’ll see the full dialogue and the final JSON printed to the console.

---

File overview

File	Purpose
run.py	Defines prompts, agents, and group chat; contains two demo note bundles.
requirements.txt (optional)	Pin exact versions for reproducible runs.

---

Customising

• Human‑in‑the‑loop review: swap clinician = AssistantAgent(...) → UserProxyAgent(...).
• Real notes: replace load_demo_notes() with your own loader.
• Different schema: edit the JSON template in a_extractor_sys and update the reviewer prompt.

---

Contact

Questions or suggestions? Open an issue or email Haining Wang at hw56@iu.edu.

---

License

MIT