Mahal Systems MedChain
Fits end-to-end clinical workflow agents where an early mistake (wrong specialty, missed question) should hurt later stages. Borrow the five-stage sequential case format and the per-stage metrics (accuracy and IoU for referral and history, LLM scoring for diagnosis).
What it asks the agent to do
- refer a case to the right specialty
- take a history by asking a simulated patient questions
- choose examinations and interpret the resulting images
- make a diagnosis
- propose a treatment plan that follows from earlier stages
What to borrow from it
case templateuser simulator
Caveats
Cases are Chinese-language clinical cases scraped with permission from the iiYi website and de-identified by that site, so provenance is weaker than a hospital dataset; grading mixes set overlap with prompt-based LLM scoring; 7,338 cases need image input; the released train and test split is a 7 to 3 cut of the same source.
Details
| Workloads | healthcare, multi agent, knowledge qa |
|---|---|
| Interaction | multi turn simulated user |
| Environment | static |
| End state checked | No |
| Tools | LLM patient agent, medical images with reports |
| Grading | programmatic check, llm judge |
| Size | 12,163 |
| Data | real scraped |
| Personal data | possible |
| License | MIT |
| Harness | own repo |
| Status | active |