title	DischargeAI
emoji	🏥
colorFrom	red
colorTo	blue
sdk	docker
app_file	app.py
pinned	false

DischargeAI: Safe Patient Transition Environment

DischargeAI is a hospital discharge planning simulation where an AI agent decides whether to discharge, keep (observation), or refer (SNF/Rehab) a patient based on their condition, recovery progress, and external factors like home support. This tests the agent's ability to balance patient safety, efficiency, and context-awareness.

Baseline Model

This environment uses meta-llama/Llama-3.1-8B-Instruct via the Hugging Face Inference Router as the baseline agent. No OpenAI key required — only a Hugging Face token with "Make calls to Inference Providers" permission.

Task Setup

Task	Patient Profile	Correct Action
EASY	Stable post-op patient, full home support, low risk	discharge
MEDIUM	Hip replacement patient, high fall risk, lives alone	refer (to rehab/SNF)
HARD	COPD patient, borderline SpO2 (93%), poor home setup	keep until Day 2, then discharge

Output Format

The agent must return a strict JSON payload:

{
  "decision": "discharge",
  "instructions": "Follow up with PCP in 2 weeks. Take medications as prescribed.",
  "follow_up_days": 14
}

Scoring

Each decision is scored on 3 criteria, clamped to [0.0, 1.0]:

• Safety (60%): Correct decision relative to patient condition.
• Instructions (20%): Length and quality of care instructions.
• Follow-up (20%): Reasonable follow-up window (1–30 days).

Agent Capabilities Evaluated

• Safety and protocol following vs. unwarranted discharge penalty.
• Instruction generation length and quality.
• Long-horizon planning (keeping a patient until vitals stabilize on day 2).

Example Baseline Run

[START] task=EASY env=DischargeAI model=meta-llama/Llama-3.1-8B-Instruct
[STEP] step=1 action={"decision": "discharge", "instructions": "Discharge the patient as they are recovering well, have full support at home...", "follow_up_days": 7} reward=1.00 done=true error=null
[END] success=true steps=1 score=1.000 rewards=1.00

Deployment Instructions

Prerequisites

pip install requests openai "httpx<0.28.0"

Run the Server Locally

uvicorn app:app --port 7860

Run the Baseline Agent

# Set environment variables
export ENV_URL="https://h8j10-dischargeai.hf.space"   # or http://localhost:7860
export HF_TOKEN="<your-huggingface-token>"

# Run
python inference.py

Windows (PowerShell):

$env:ENV_URL = "https://h8j10-dischargeai.hf.space"
$env:HF_TOKEN = "<your-huggingface-token>"
$env:TASK_NAME = "EASY"   # or MEDIUM, HARD
python inference.py

Deploying to Hugging Face Spaces (Docker)

1. Push this repository to a Hugging Face Space using the Docker template.
2. In your Space Settings → Variables and Secrets, add:
   • Secret: HF_TOKEN — your Hugging Face token (needs "Make calls to Inference Providers" permission)
   • Variable: MODEL_NAME — meta-llama/Llama-3.1-8B-Instruct (optional, this is the default)
3. Once the Space status shows Running, validate with:

   ./validate-submission.sh https://h8j10-dischargeai.hf.space