README.md

metadata

title: Appoint Ready - MedGemma Demo
emoji: 📋
colorFrom: blue
colorTo: gray
sdk: docker
models:
  - google/medgemma-27b-text-it
pinned: false
license: apache-2.0
short_description: Simulated Pre-visit Intake Demo built using MedGemma

Table of Contents

• Demo Description
• Technical Architecture
• Running the Demo Locally
• Models used
• Caching
• Disclaimer
• Other Models and Demos

AppointReady: Simulated Pre-visit Intake Demo built using MedGemma

Healthcare providers often seek efficient ways to gather comprehensive patient information before appointments. This demo illustrates how MedGemma could be used in an application to streamline pre-visit information collection and utilization.

The demonstration first asks questions to gather pre-visit information. After it has identified and collected relevant information, the demo application generates a pre-visit report using both collected and health record information (stored as FHIR resources for this demonstration). This type of intelligent pre-visit report can help providers be more efficient and effective while also providing an improved experience for patients compared to traditional intake forms.

At the conclusion of the demo, you can view an evaluation of the pre-visit report which provides additional insights into the quality of the demonstrated capabilities. For this evaluation, MedGemma is provided the patient's reference diagnosis, allowing MedGemma to create a self-evaluation report highlighting strengths as well as areas for improvement.

Technical Architecture

This application is composed of several key components:

• Frontend: A web interface built with React that provides the user interface for the chat and report visualization.
• Backend: A Python server built with Gunicorn/Flask that handles the application logic. It communicates with the LLMs, manages the conversation flow, and generates the final pre-visit report.
• API called:
  • MedGemma: Acts as the clinical assistant, asking relevant questions and summarizing information.
  • Gemini: Role-plays as the patient, providing responses based on a predefined scenario.
  • Gemini TTS: Generative text-to-speech model.
• Deployment: The entire application is containerized using Docker for easy deployment and scalability.

A high-level overview of the architecture:

Running the Demo Locally

To run this demo on your own machine, you'll need to have Docker installed.

Prerequisites

• Docker
• Git
• A Google Cloud project with the Vertex AI API enabled.

Setup & Configuration

1. Clone the repository:

   git clone https://huggingface.co/spaces/google/appoint-ready
   cd appoint-ready
   

2. Configure environment variables: This project uses an env.list file for configuration, which is passed to Docker. Create this file in the root directory.

   # env.list
   GEMINI_API_KEY="your-gemini-token"
   GENERATE_SPEECH=false or true
   GCP_MEDGEMMA_ENDPOINT=medgemma vertex ai endpoint 
   GCP_MEDGEMMA_SERVICE_ACCOUNT_KEY="service-account-key json"  
   

   GEMINI_API_KEY: Key can be generated via aistudio. GENERATE_SPEECH: Should the demo generate speech not found in cache. Default is false. GCP_MEDGEMMA_ENDPOINT: Deploy MedGemma via Model Garden.

Execution

1. Build and run the Docker containers:

   run_local.sh
   

2. Access the application: Once the containers are running, you can access the demo in your web browser at http://localhost:[PORT]. (e.g., http://localhost:7860).

Models used

This demo uses four models:

• MedGemma 27b-text-it: https://huggingface.co/google/medgemma-27b-text-it
  For this demo MedGemma-27b was deployed via Model Garden (https://console.cloud.google.com/vertex-ai/publishers/google/model-garden/medgemma).
• Gemini: https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash
  We use Gemini to role play the patient while MedGemma plays the clinical assistant.
• Gemini TTS: https://ai.google.dev/gemini-api/docs/models#gemini-2.5-flash-preview-tts
  As an alternative consider using Clould Text-to-Speech: https://cloud.google.com/text-to-speech
• Veo 3: Was used to generate patient avatar animation https://gemini.google/overview/video-generation

Caching

This demo is functional, and results are persistently cached to reduce environmental impact.

Disclaimer

This demonstration is for illustrative purposes only and does not represent a finished or approved product. It is not representative of compliance to any regulations or standards for quality, safety or efficacy. Any real-world application would require additional development, training, and adaptation. The experience highlighted in this demo shows MedGemma's baseline capability for the displayed task and is intended to help developers and users explore possible applications and inspire further development.

This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.

Other Models and Demos

See other demos here: https://huggingface.co/collections/google/hai-def-concept-apps-6837acfccce400abe6ec26c1

MedGemma is finetunable - see colab here: https://github.com/Google-Health/medgemma/blob/main/notebooks/fine_tune_with_hugging_face.ipynb

Contacts

• This demo is part of Google's Health AI Developer Foundations (HAI-DEF)
• Technical info - contact @lirony
• Press only: [email protected]