Skip to content

MedExpertMatch Documentation

Welcome to the MedExpertMatch documentation!

About MedExpertMatch

MedExpertMatch is an AI-powered medical expert recommendation system developed for the MedGemma Impact Challenge . The system matches medical cases with appropriate doctors based on:

  • Medical case analysis (symptoms, diagnoses, ICD-10 codes)
  • Doctor expertise and clinical experience
  • Clinical guidelines and evidence
  • Similar case outcomes
  • Specialty matching

MedExpertMatch leverages a Hybrid GraphRAG architecture, combining:

  • Vector Similarity Search (PgVector) - Semantic matching based on clinical experiences (40% weight in scoring)
  • Graph Traversal (Apache AGE) - Relationship-based discovery of doctor-case connections (30% weight in scoring)
  • Historical Performance - Past outcomes, ratings, and success rates (30% weight in scoring)
  • Keyword Search - Traditional text matching for medical terms
  • Semantic Reranking - Precision optimization using MedGemma
  • LLM Orchestration - Natural language answer generation with MedGemma models

Note: The Find Specialist flow actively uses Apache AGE graph for relationship scoring. See Find Specialist Flow for details.

MedGemma Impact Challenge

MedExpertMatch is being developed for the MedGemma Impact Challenge, a hackathon organized by Google Research on Kaggle.

Challenge Requirements

  • Use HAI-DEF Models: MedGemma 1.5 4B, MedGemma 27B, MedASR, CXR Foundation, Derm Foundation, Path Foundation, HeAR
  • Human-Centered Design: Focus on real clinical workflows and user experience
  • Privacy-First Architecture: Local deployment capability, HIPAA compliance
  • Working Solution: Full-featured prototype or MVP, not just a demo

Submission Deadline

February 24, 2026 - ~6 weeks development timeline

Getting Started

Architecture

Development

Configuration

Features

Key Features

  • Case Analysis: Analyze medical cases using MedGemma to extract ICD-10 codes, urgency, and required specialty
  • Doctor Matching: Match doctors to cases based on specialty, experience, and similar case outcomes
  • Consultation Queue: Prioritize consult requests by clinical urgency so specialists see the sickest patients first (Consultation Queue)
  • Evidence Retrieval: Search clinical guidelines and PubMed for evidence-based recommendations (all tools implemented)
  • Clinical Recommendations: Generate evidence-based clinical recommendations using MedGemma (all tools implemented)
  • Network Analytics: Query graph for top experts and aggregate metrics (all tools implemented)
  • Regional Routing: Score facility-case routing matches using Semantic Graph Retrieval (all tools implemented)
  • Agent Skills: 7 medical-specific Agent Skills for modular knowledge management (Medical Agent Tools)
  • Hybrid GraphRAG: Combines vector, graph, and keyword search for optimal matching
  • Privacy-First: Local deployment capability, HIPAA-compliant data handling
  • Simulated security: User selector (Regular User / Administrator); Synthetic Data and Graph Visualization are admin-only. See Architecture - Simulated Security.

Core Use Cases

MedExpertMatch addresses six primary use cases:

  1. Specialist Matching for Complex Inpatient Cases - Reduce consultation delays and length of stay
  2. Online Second Opinion / Telehealth - Fast, accurate second opinion matching
  3. Prioritizing the Consultation Queue - Ensure urgent cases are seen first
  4. Network Analytics - Data-driven expertise mapping and routing policies
  5. Human-in-the-Loop Decision Support - AI copilot with expert recommendations for specialists
  6. Cross-Organization / Regional Routing - Optimal facility and specialist routing across networks

See Use Cases for detailed workflows, API endpoints, and technical implementation.

Architecture Overview

MedExpertMatch uses a modern, scalable architecture:

  • Backend: Spring Boot 4.0.2, Java 21
  • Database: PostgreSQL 17 with PgVector and Apache AGE 1.6.0
  • AI/ML: Spring AI 2.0.0-M2 with MedGemma models
  • Vector Search: PgVector with HNSW indexing
  • Graph Database: Apache AGE for relationship traversal
  • Agent Skills: Spring AI Agent Skills for medical domain knowledge

Architecture Components

MedExpertMatch includes:

  • Infrastructure: Database, vector search, graph, LLM integration
  • Services: Retrieval, fusion, reranking with medical-specific adaptations
  • Domain Models: Doctor, MedicalCase, ClinicalExperience, ICD10Code
  • Medical Components: Medical domain modules, case analysis, evidence retrieval, consultation queue

Project Status

Current Phase: MVP complete, feature refinement

  • Challenge analysis completed
  • Architecture design complete
  • Core implementation complete (case analysis, matching, queue, evidence, network, routing)
  • Consultation queue prioritization uses real cases from DB and deterministic urgency ordering

Documentation Structure

This documentation is organized into several sections:

  1. Overview - Project overview and challenge details
  2. Architecture - System architecture and design
  3. Development - Setup and development guides
  4. Configuration - Configuration and optimization guides

Important Disclaimers

⚠️ MedGemma is NOT a Medical Device:

  • Models are not certified for clinical use
  • Additional validation required for real-world deployment
  • Not intended for diagnostic decisions without human-in-the-loop
  • All applications are for research and educational purposes

⚠️ HIPAA Compliance:

  • All patient data must be anonymized
  • Local deployment option for privacy
  • No transmission of PHI without proper safeguards

Contributing

For development guidelines, see:

Support

For questions or issues related to MedExpertMatch or the MedGemma Impact Challenge, please refer to the relevant documentation section.


Last updated: 2026-02-03