Hospital Triage AI

A hybrid Reinforcement Learning and LLM multi-agent system for intelligent patient triage and ward management. Built as an AI Agent course project.

Live Demo

The application is deployed and accessible without any local setup: https://delightful-learning-production-7606.up.railway.app

---

Architecture

Phase 1 — Single Triage Agent (DQN)

• Deep Q-Network trained over 2000 episodes
• Manages a single patient across 5 severity states (S0 Healthy to S4 Emergency)
• Legal action masking prevents clinically invalid actions
• Single step automatically generates a Claude explanation
• Episode run logs all steps, explanations available on demand via Ask button

Phase 2 — Multi-Agent System (Q-Learning)

• Supervisor Agent (Q-Learning, 3000 episodes) observes the entire ward
• Coordinates 3 independent DQN Triage Agents, one per patient
• Ward state W0-W4 derived by aggregating all three patient states
• Supervisor can override the most critical triage agent when ward is in crisis
• Single step automatically generates a Claude ward report
• Episode run logs all steps, ward reports available on demand via Ask button
• Patient states are assigned directly via dropdown (no NLP input in Phase 2)

LLM Layer (Claude API)

• Phase 1 only: parses natural language patient descriptions into structured vitals
• Generates clinical triage explanations after every single step
• Generates ward-level supervisor reports covering all three patients
• Strictly separated from RL logic — LLM never selects actions or sees Q-values

---

## Project Structure
hospital-triage-agent/
├── backend/
│   ├── agents/
│   │   ├── triage_agent.py        # DQN agent — Phase 1 and Phase 2
│   │   ├── supervisor_agent.py    # Q-Learning supervisor — Phase 2
│   │   └── llm_layer.py           # Claude API integration
│   ├── api/
│   │   └── main.py                # FastAPI REST endpoints
│   ├── models/
│   │   ├── dqn_model.py           # Neural network + action mask
│   │   ├── q_table.py             # Q-table + ward action mask
│   │   └── saved/
│   │       ├── triage_agent.pth
│   │       └── supervisor_qtable.npy
│   ├── training/
│   │   ├── train_triage.py
│   │   └── train_supervisor.py
│   ├── utils/
│   │   └── state_mapper.py        # Vitals scoring + ward derivation
│   └── requirements.txt
├── frontend/
│   ├── src/
│   │   ├── pages/
│   │   │   ├── About.jsx
│   │   │   ├── Phase1.jsx
│   │   │   └── Phase2.jsx
│   │   ├── components/
│   │   │   ├── StateBadge.jsx
│   │   │   ├── ActionBadge.jsx
│   │   │   ├── QValueBar.jsx
│   │   │   └── RewardChart.jsx
│   │   └── App.jsx
│   └── package.json
└── README.md

---

MDP Design

Triage States S0-S4

State	Label	Clinical Description
S0	Healthy	All vitals within normal range
S1	At Risk	One vital mildly out of range
S2	Unstable	Multiple vitals elevated
S3	Critical	Dangerous readings, urgent response needed
S4	Emergency	Life-threatening, immediate intervention required

Vitals to State Scoring

Each vital is scored 0 to 3 independently. Scores are summed together with a risk modifier. Final state is determined by total score thresholds.

Vital	Score 0	Score 1	Score 2	Score 3
Heart Rate	60 to 100	100-120 or 50-60	120-140 or 40-50	Above 140 or below 40
BP Systolic	100 to 140	140-160 or 90-100	160-180 or 80-90	Above 180 or below 80
SpO2	95% or above	92 to 95%	88 to 92%	Below 88%
Temperature	36 to 38C	38 to 39C	39 to 40C or below 36	Above 40 or below 35

Risk modifiers: age above 70 adds 1, two or more pre-existing conditions adds 1.

State thresholds: total 0 = S0, 1-2 = S1, 3-5 = S2, 6-8 = S3, 9 or above = S4.

Triage Actions with Legal Mask

Action	Label	Legal States
A0	Monitor	S0, S1, S2
A1	Treat	S1, S2
A2	Escalate	S1, S2, S3, S4
A3	Emergency Response	S2, S3, S4

Triage Reward Function

Condition	Reward
Patient improves N states	+N x 10
No change	0
Patient worsens N states	-N x 10
Stuck at S4 to S4	-30
Over-escalation (S2 + Emergency)	-15
Under-action at S3 or S4 (A0, A1)	-20
S4 saved to S2 or better	+25

Supervisor Ward States W0-W4

Ward	Label	Derivation Rule
W0	Calm	All patients at S0 or S1
W1	Active	count(S2) >= 1
W2	Busy	count(S2) >= 3 or count(S3) = 1
W3	Overloaded	count(S3) >= 2 or count(S4) = 1
W4	Crisis	count(S4) >= 2

Supervisor Actions with Legal Mask

Action	Label	Legal Wards
B0	Standby	W0, W1
B1	Reallocate	W1, W2, W3
B2	Override Triage	W2, W3, W4
B3	Request Backup	W2, W3, W4

Trained Q-Table (after 3000 episodes)

Ward State	B0 Standby	B1 Reallocate	B2 Override	B3 Backup
W0 Calm	-6.03	---	---	---
W1 Active	3.27	3.71	---	---
W2 Busy	12.49	12.28	13.28	12.29
W3 Overloaded	---	-1.56	16.85	21.21
W4 Crisis	---	---	6.03	18.59

Key insight: Request Backup at W3 (21.21) vastly outperforms Reallocate (-1.56). The agent learned that bringing new resources outperforms reshuffling existing ones. Discovered with zero hardcoded rules.

---

API Endpoints

Method	Endpoint	Description
POST	/api/vitals-to-state	Map raw vitals to a state integer
POST	/api/parse-patient	Parse free text into vitals and state
POST	/api/triage-step	Run one Phase 1 agent step
POST	/api/supervisor-step	Run one Phase 2 multi-agent step
POST	/api/explain-step	Generate LLM explanation for a transition
POST	/api/ward-explain	Generate ward report for an episode step
GET	/api/metrics	Get live session performance metrics

---

Setup and Run

Prerequisites

• Python 3.12
• Node.js 22
• Anthropic API key from console.anthropic.com

1. Clone

git clone https://github.com/Nsk246/redesigned-funicular
cd redesigned-funicular

2. Backend setup

cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

3. Add API key

echo "ANTHROPIC_API_KEY=sk-ant-your-key-here" > .env

4. Run backend (Terminal 1)

cd backend/api
uvicorn main:app --reload --host 0.0.0.0 --port 8000

5. Run frontend (Terminal 2)

cd frontend
npm install
npm run dev

6. Open

http://localhost:5173

Apple Silicon (M1/M2/M3)

pip install torch==2.3.0 --index-url https://download.pytorch.org/whl/cpu

---

Running in GitHub Codespaces

The browser in a codespace cannot reach localhost:8000 directly. Follow these extra steps:

1. In the VS Code Ports tab at the bottom, right-click port 8000 and set visibility to Public
2. Find your codespace name by running echo $CODESPACE_NAME in the terminal
3. Create frontend/.env.local with the following content, replacing YOUR-CODESPACE-NAME with your actual codespace name: VITE_API_URL=https://YOUR-CODESPACE-NAME-8000.app.github.dev
4. Restart the frontend dev server

This file is gitignored and does not affect the Railway deployment.

---

Retraining

Retraining is only needed if the reward function or MDP structure is changed.

# Retrain Triage Agent (Phase 1) — approximately 2 minutes
python backend/training/train_triage.py

# Retrain Supervisor Agent (Phase 2) — approximately 3 minutes
python backend/training/train_supervisor.py

---

Tech Stack

Category	Technology
RL	PyTorch DQN, Experience Replay, Tabular Q-Learning, Epsilon-Greedy, Legal Action Masking
LLM	Anthropic Claude API (claude-haiku-4-5)
Backend	Python 3.12, FastAPI, Uvicorn
Frontend	React 19, Vite 8, TailwindCSS, Recharts, Axios
Fonts	Inter, JetBrains Mono
Deploy	Railway (backend + frontend)

---

License

MIT