title	ClinicalTrialEnv
emoji	🧬
colorFrom	blue
colorTo	green
sdk	docker
app_port	7860

ClinicalTrialEnv 🧬

ClinicalTrialEnv is a live OpenEnv reinforcement-learning environment for adaptive Phase II clinical trial design. Agents act as autonomous trial statisticians: they choose cohort size, allocation across dose arms, early stopping, arm dropping, and patient inclusion strictness to reach significance quickly while strictly protecting patient safety.

The Challenge: Ethics vs. Optimization
Standard environments optimize for a simple score. ClinicalTrialEnv forces AI agents to formulate high-stakes trade-offs. The agent must balance mathematical optimization (powering a trial, achieving statistical significance, managing budget) against severe ethical constraints (limiting adverse patient events, avoiding over-enrollment in rare diseases, and swiftly dropping unsafe high-dose arms).

Built by Manas Dutta.

🚀 Quick Links

• Live Hugging Face Space
• GitHub Repository

Tasks

1. task_1: effective drug with a clear signal and a 200-patient budget
2. task_2: efficacy versus safety tradeoff with a risky high-dose arm
3. task_3: rare-disease setting with only 150 patients and weak effects

Observation Space

The environment exposes trial progress, response rates, adverse-event rates, patient counts, p-values, Bayesian posterior probabilities, estimated power, active-arm flags, and stop metadata. The full schema is available from GET /tasks and GET /schema.

Action Space

Agents submit:

• n_next_cohort
• allocation_control
• allocation_low
• allocation_mid
• allocation_high
• stop_for_success
• stop_for_futility
• drop_arm
• inclusion_criteria_strictness

Setup

cd clinical_trial_env
pip install openenv-core
pip install -e .
uvicorn server.app:app --port 7860 --reload

Open:

• http://localhost:7860/ for the custom dashboard
• http://localhost:7860/web for the OpenEnv built-in UI

Run the LLM agent against the live environment:

API_BASE_URL=<url> MODEL_NAME=<model> HF_TOKEN=<token> python inference.py

LLM Agent (inference.py)

inference.py is the hackathon submission agent used by judges against the deployed environment.

• It uses the OpenAI client for all model calls through an OpenAI-compatible endpoint.
• It emits exact [START], [STEP], and [END] structured logs for evaluation.
• By default it targets meta-llama/Llama-3.1-8B-Instruct through the Hugging Face Inference API.
• It can be redirected to any OpenAI-compatible API by setting API_BASE_URL and MODEL_NAME.
• If model calls fail, it falls back to a built-in heuristic so the run still completes.

API Reference

Endpoint	Method	Purpose
/	GET	Serve the custom dashboard directly with HTTP 200
/web	GET	OpenEnv built-in web UI
/health	GET	Healthcheck for deployment validators
/tasks	GET	List tasks plus action_schema
/reset	POST	Start an HTTP episode for validator compatibility
/step	POST	Step the latest active HTTP episode
/grader	POST	Return the final normalized score for the latest completed episode
/baseline	POST	Run the heuristic agent across all three tasks
/ws	WebSocket	Main OpenEnv interaction channel

Example Session

WebSocket reset:

{"type": "reset", "data": {"task_id": "task_1"}}

WebSocket step:

{"type": "step", "data": {"n_next_cohort": 25, "allocation_control": 0.25, "allocation_low": 0.25, "allocation_mid": 0.25, "allocation_high": 0.25, "stop_for_success": false, "stop_for_futility": false, "drop_arm": null, "inclusion_criteria_strictness": 0.5}}

Baseline Results

Task	Typical Score	Outcome	Notes
task_1	~0.8	success	Clear efficacy signal with efficient stopping
task_2	~0.55-0.65	success	Balances efficacy with AE control
task_3	~0.2-0.4	futility	Futility is often the correct behavior because the task is designed to reward efficient rare-disease decision making, not forced over-enrollment

Notebook Demo

clinical_trial_env/Environment_Demo.ipynb is a runnable Colab notebook that connects to the live Space, runs a heuristic policy over WebSocket, plots p-value trajectories, and fetches the final grader breakdown.

Deploy to Hugging Face Spaces

openenv push --repo-id manasdutta04/clinicaltrialenv