Patient Router

Live Demo →

This is a student research prototype, not a validated clinical tool. It has not been tested on real patients, so please don't use it to make actual triage decisions.

---

Overview

In hospital emergency departments, patients are often sent to the wrong department at first, and that wastes time that matters. For this project I tried to see if a simple ML model could help with that first triage step: a Gradient Boosting classifier trained on structured patient data, plus some rule-based logic on top for priority and emergency cases, and a feedback loop so corrections go back into the training data.

I also added Gemini API and Hybrid prediction methods to experiment with other approaches while keeping the locally trained model as the main part of the project.

The local model is trained only on synthetic data I generated myself, so please treat the predictions as a proof of concept, not something clinically reliable.

Patient Router is available as a web application and as an Electron desktop application for Windows and Linux.

The project has three parts:

Part	Location	Responsibility
ML core	backend/ml/	Synthetic data generation, model comparison, training, evaluation, inference
Flask API	backend/app.py, routes/, services/	Exposes the prediction pipeline and other backend functionality over HTTP
React dashboard	frontend/	Patient intake, prediction, feedback collection, dataset management, training, evaluation, and logs

For the full pipeline, API contract, environment setup, and everything else, see the documentation below.

---

Documentation

Doc	Covers
docs/architecture.md	ML pipeline, the three prediction methods, priority scoring, emergency detection, normalization, evaluation, model comparison, feedback loop
docs/api.md	Full request/response examples for every route
docs/setup.md	Environment variables, local setup, troubleshooting
docs/frontend.md	React dashboard structure: pages, hooks, shared components
docs/data-schema.md	Symptom/vital/history vocabulary and weight tables
docs/deployment.md	Vercel frontend, backend hosting requirements, desktop builds

---

System Flow

flowchart LR
    A[Patient Input] --> B{Prediction Method}
    B --> C[Patient Router]
    B --> D[Gemini API]
    B --> E[Hybrid]

    E --> F[Local Model First]
    F --> G{Confidence >= 0.60?}
    G -- Yes --> H[Prediction Result]
    G -- No --> D

    C --> I[Priority & Emergency Rules]
    I --> H
    D --> H

    H --> J[Recommendation + Logged Prediction]

Loading

For the full breakdown of each stage, see docs/architecture.md.

---

Screenshots

Click to expand

Patient Router	Train Model

Data Manager	Evaluation

System Logs

---

API

Patient Router exposes REST API endpoints for prediction, feedback, dataset management, model training, evaluation, model comparison, and logs.

For the complete API reference with request and response examples, see docs/api.md.

---

Running Locally

# backend
cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
python -m ml.generate_data
python -m ml.train
python app.py

# frontend
cd frontend
npm install
cp .env.example .env
npm run dev

For environment variables, config, and troubleshooting, see docs/setup.md.

---

Limitations

• The local model is trained on synthetic data I generated, not real patient records, so I can't say how it would actually perform in a hospital
• Only 6 departments, 20 symptoms, 7 vitals, and 6 history conditions: this was a scope decision to keep the project manageable, not something I ran out of time to add

---

Tech Stack

Backend: Python, Flask, scikit-learn, pandas, numpy, joblib Frontend: React, TypeScript, Vite, lucide-react Desktop: Electron, electron-builder ML: GradientBoostingClassifier, CountVectorizer, OneHotEncoder External API: Gemini 2.5 Flash