🩺 MedGemma Clinical Triage System & Physician Dashboard

👉 [Click here to try the Live Web Demo on Hugging Face Spaces] ([(https://huggingface.co/spaces/TejaswiniRay/Clinical-Triage-Demo)])

👉 [Read the full technical article about this project] (https://medium.com/@tej.ray448/building-a-clinical-triage-ai-with-googles-medgemma-1-5-2ab1a49d03b2)

📌 Project Overview

Healthcare systems face massive administrative bottlenecks, with doctors spending hours reading intake forms. This project is an end-to-end, privacy-first AI Triage System built using Google's MedGemma 1.5 (4B) open-weight model.

It acts as a virtual clinic assistant that securely interviews patients, extracts medical symptoms using Chain-of-Thought reasoning, and populates a backend SQLite physician dashboard—all designed to run locally on-premise to simulate strict data privacy environments.

🔬 Implementing the MedGemma Technical Report

This project serves as an applied implementation of several core capabilities highlighted by Google DeepMind and Google Research in the MedGemma Technical Report:

• Medical Agentic Behavior: Implements the paper's claims on patient interviewing and triaging by allowing the model to autonomously guide a multi-turn symptom gathering process.
• Clinical Reasoning Extraction: Validates the model's advanced clinical reasoning by intercepting, parsing, and displaying the hidden <unused94> Chain-of-Thought tokens in the UI.
• Strict Constraint Compliance: Tests the model's instruction-tuning by enforcing negative prompt constraints ("Do not diagnose") and rigid output formatting.
• Privacy-Preserving Edge AI: Proves the paper's thesis that the compact 4B architecture can perform complex medical NLP tasks entirely locally, eliminating the need for cloud-based PHI (Protected Health Information) transmission.

🏗️ System Architecture

✨ Key Technical Features (The Engineering)

• Advanced Reasoning Parsing: MedGemma 1.5 uses hidden reasoning tokens. The app intercepts the AI's internal Chain-of-Thought before it renders to the UI, parses the strings, and hides the logic inside a collapsible "Reasoning Trace" expander for physician oversight.
• Strict Jinja Template Compliance: The Gemma family strictly requires alternating user/assistant chat structures. The app uses prompt engineering injection to dynamically combine system prompts and hidden directives into standard user turns to prevent pipeline crashes.
• Programmatic "Hard Stop": To prevent infinite chat loops, Python logic counts user turns. On the 5th turn, a secret system directive is appended to force the AI to output a 3-bullet CLINICAL SUMMARY, automatically disabling the chat interface.
• Full CRUD Architecture: The front-end is backed by a local SQLite database (clinic.db). Completed patient summaries are instantly saved and routed to a secondary "Doctor's Dashboard" view.
• Data Compliance Workflows: To simulate HIPAA/GDPR data-retention requirements, the dashboard features a Discharge Patient action that executes a SQL DELETE command, permanently and securely wiping the temporary patient record.

🛠️ Tech Stack

• AI Model: Google DeepMind's medgemma-1.5-4b-it
• LLM Pipeline: transformers, torch, accelerate
• Frontend Framework: Streamlit
• Database Backend: SQLite (sqlite3)

🚀 Installation & Setup (Local)

1. Clone the repository:

git clone https://github.com/YOUR_USERNAME/MedGemma-Triage-System.git
cd MedGemma-Triage-System

2. Install dependencies:

pip install -r requirements.txt

3. Get Hugging Face Access:

• Create a Hugging Face account.
• Go to google/medgemma-1.5-4b-it and agree to the Health AI Developer Foundations terms.
• Generate an Access Token in your Hugging Face settings.

4. Add your Token:

• Create a folder named .streamlit in the project root.
• Inside it, create a file named secrets.toml and add:

  HF_TOKEN = "your_actual_token_here"

5. Run the Application:

streamlit run Triage.py

🏥 How to Use It

1. Patient View: Start on the "Patient Triage" screen. Describe a medical symptom. Answer the AI's follow-up questions.
2. AI Reasoning: Click "🧠 View AI Reasoning Trace" on the AI's replies to see its internal medical logic.
3. Completion: On the 5th interaction, the AI will lock the chat and generate the summary. Click "Start Triage for Next Patient" to reset the kiosk.
4. Physician View: Use the sidebar to navigate to the "Doctor's Dashboard". Read the generated patient summaries.
5. Discharge: Click "✅ Discharge" to permanently wipe the record from the SQLite database.

🔮 Future Roadmap

• Migrate SQLite backend to a cloud-based PostgreSQL database.
• Format the summary output into FHIR (Fast Healthcare Interoperability Resources) JSON standards for direct EHR integration (e.g., Epic, Cerner).
• Add MedASR (Speech-to-Text) so patients can speak into the kiosk instead of typing.

---

Disclaimer: This application is for educational, portfolio, and demonstration purposes only. It is not intended for real-world medical diagnosis or actual patient use.