Claude Code for Health
A reinforcement learning (RL) environment for training and evaluating clinical reasoning agents.
- My role
- Developer · team of three
- Built with
- Python · OpenEnv · FastAPI · Gradio · Docker
The idea
Clinical reasoning involves deciding what to investigate next, interpreting the result, and revising a hypothesis. We built a reinforcement learning (RL) environment that trains and evaluates that sequence of decisions through a terminal interface.
The build
An agent issues commands to explore a patient chart, consult reference information, build a differential diagnosis, or submit a calculation. OpenEnv connects those actions to structured observations and programmatic graders, with a FastAPI server and Gradio dashboard for inspecting episodes.
The environment covers 15,000+ cases across three tasks: clinical note review using MEDEC, medical calculations using MedCalc-Bench, and diagnostic workups using MedCaseReasoning. Reward functions score intermediate steps as well as final answers, accounting for relevant exploration, correctness, and efficient tool use.
The outcome
Built with Siddhant Bharadwaj and Pratinav Seth, the project provides a repeatable environment for training and evaluating clinical reasoning agents, with a baseline inference script and containerized deployment.