Multi-Agent by raywong2121 - MCP Server

Multi-Agent Deliberative Intelligence Chinese Medical Imaging Report Text Standardization MCP Server

Project Overview

This project implements a multi-agent deliberative intelligence approach for Chinese medical imaging report text standardization, delivered through the Model Context Protocol (MCP). The system employs a complete processing pipeline of "LLM Tokenization & POS Tagging → Medical Entity Recognition → Six-Agent Collaboration → Quality Gates" to efficiently and accurately convert Chinese imaging reports to OMOP CDM 5.4 standard format.

Core Features

🤖 Six-Agent Collaborative Architecture

Linguist Agent: Performs text normalization, syntactic analysis, and ambiguity resolution
Structurer Agent: Strictly follows OHDSI separation of concerns principles, decomposing concepts into independent standard components
Mapper Agent: Uses ReAct (Reasoning-Acting) mode to rapidly generate concept mapping proposals
Skeptic Agent: Specializes in challenging and validating mapper proposals to ensure quality control
Auditor Agent: Performs final validation based on complete deliberation history
OMOP Builder: Converts validated concept mappings to OMOP CDM 5.4 standard format according to OHDSI community best practices

🧠 Deliberative Intelligence Algorithm

Proposer-Skeptic Adversarial Mechanism: Ensures high-quality and reliable concept mapping through multi-round deliberation
Dynamic Threshold Adjustment: Adaptive threshold adjustment based on terminology complexity and context
Multi-Strategy Concept Search: Four-tier matching strategies (exact match, synonym match, fuzzy match, semantic match)

🛡️ Quality Control System

Four-Dimensional Quality Assessment: Logical consistency, deliberation quality, evidence strength, OHDSI compliance
Multi-Level Quality Gate Mechanism: High/medium/low confidence level processing
Real-Time Quality Monitoring: Provides detailed quality assessment reports and improvement recommendations

🏥 OMOP CDM 5.4 Integration

SNOMED CT Terminology Mapping: Precise concept alignment and standardization
Multi-Table Record Generation: Reasonable allocation to OBSERVATION, MEASUREMENT, FACT_RELATIONSHIP tables
OHDSI Best Practices: Follows Observational Health Data Sciences and Informatics collaborative organization standards

Technical Architecture

Chinese Imaging Report → [LLM Sentence-by-Sentence Tokenization + POS Tagging] → POS-Tagged Entities → [Medical Entity Validation] → Enhanced Entity Annotation
    ↓                    (ANATOMY/FINDING/ATTRIBUTE/MODIFIER/NEGATION/MEASUREMENT/PROCEDURE)
[Linguist Agent] → Normalized Text + Semantic Segments → [Structurer Agent] → Structured Entity List
    ↓
[Mapper ↔ Skeptic Deliberation] → POS Tag Validated Mapping → [Auditor Agent] → Quality Validation
    ↓
[OMOP Builder] → OBSERVATION/MEASUREMENT/FACT_RELATIONSHIP Table Records → Local Database Storage

Installation and Deployment

Environment Requirements

Node.js >= 18.0.0
TypeScript >= 5.3.0
OpenAI-compatible LLM Service
PostgreSQL Database

Installation Steps

Clone Repository

git clone <repository-url>
cd medical-text-standardization-mcp

Configure Environment Variables

# Copy environment configuration file
cp .env.example .env

# Edit configuration file (pre-configured with real services)
# LLM Service:
# Database:

Install Dependencies

npm install

Build Project

npm run build

Health Check

# Check LLM service


# Check database connection

Start Server

npm start

Development Mode

npm run dev

MCP Tools Usage

1. Process Medical Report

{
  "name": "process_medical_report",
  "arguments": {
    "reportText": "Patient chest CT examination shows a nodule in the left upper lobe with a diameter of approximately 12mm, clear boundaries, and uniform density.",
    "reportId": "R001",
    "patientId": "P12345",
    "reportType": "CT"
  }
}

2. Analyze Medical Concept

{
  "name": "analyze_medical_concept",
  "arguments": {
    "concept": "pulmonary nodule",
    "context": "chest CT examination finding"
  }
}

3. Validate Concept Mapping

{
  "name": "validate_concept_mapping",
  "arguments": {
    "sourceText": "pulmonary nodule",
    "targetConceptId": 4230254
  }
}

4. Search OMOP Concepts

{
  "name": "search_omop_concepts",
  "arguments": {
    "query": "pulmonary nodule",
    "domain": "Condition",
    "limit": 10
  }
}

5. Get Processing Statistics

{
  "name": "get_processing_statistics",
  "arguments": {}
}

Configuration Guide

Agent Configuration

agents: {
  linguist: {
    name: 'Linguist',
    model: '',
    temperature: 0.3,
    maxTokens: 2048
  },
  // ... other agent configurations
}

Quality Threshold Configuration

qualityThresholds: {
  logicalConsistency: 0.8,
  deliberationQuality: 0.75,
  evidenceStrength: 0.7,
  ohdsiCompliance: 0.85,
  overallScore: 0.8
}

Performance Metrics

Based on research paper experimental results:

Concept Mapping Accuracy: 95.2%
Processing Speed: 1,200 reports/hour
F1 Score: 0.943
Quality Gate Pass Rate: 93.7%
Deliberation Rounds: Average 2.3 rounds
Consensus Achievement Rate: 94.7%

Project Structure

src/
├── agents/           # Six-agent implementation
│   ├── system.ts     # Multi-agent system
│   ├── linguist.ts   # Linguist agent
│   ├── structurer.ts # Structurer agent
│   ├── mapper.ts     # Mapper agent
│   ├── skeptic.ts    # Skeptic agent
│   ├── auditor.ts    # Auditor agent
│   └── omop-builder.ts # OMOP Builder agent
├── core/             # Core processing logic
│   ├── processor.ts  # Medical text processor
│   └── quality-gate.ts # Quality gate mechanism
├── services/         # Service layer
│   └── omop-builder.ts # OMOP builder service
├── types/            # Type definitions
│   └── index.ts
├── utils/            # Utility classes
│   └── logger.ts
├── config/           # Configuration management
│   └── index.ts
└── index.ts          # MCP server entry point

Data Privacy and Security

Real Service Integration

Production-Grade LLM Service: Integrated OpenAI-compatible LLM2.5 model providing high-quality medical text understanding
Enterprise Database: Connected to PostgreSQL OHDSI database supporting real OMOP concept queries
Health Check Mechanism: Automatic service availability detection ensuring stable system operation
Configuration Management: Flexible service connection parameter configuration through environment variables

Security Features

Data Encryption: Support for data transmission and storage encryption
Access Control: Role-based access control mechanism
Audit Logging: Complete operational audit tracking

Quality Assurance

Test Coverage

Unit Testing: Coverage of all core components
Integration Testing: End-to-end process validation
Performance Testing: Large-scale data processing validation

Quality Monitoring

Real-time Monitoring: Real-time processing quality monitoring
Anomaly Detection: Automatic identification of abnormal mappings
Quality Reports: Regular generation of quality analysis reports

Extension and Customization

Agent Extension

// Custom agent example
class CustomAgent extends BaseAgent {
  async process(input: any): Promise<any> {
    // Custom processing logic
  }
}

Mapping Strategy Extension

// Custom mapping strategy
const customStrategy: MappingStrategyConfig = {
  strategy: 'custom_match',
  weight: 0.9,
  threshold: 0.85,
  enabled: true
};

Troubleshooting

Common Issues

Server Startup Failure
- Check if Node.js version >= 18.0.0
- Confirm all dependencies are correctly installed
- Check if port is occupied
Low Mapping Quality
- Adjust quality threshold configuration
- Check medical dictionary completeness
- Optimize agent prompts
Slow Processing Speed
- Check local LLM performance
- Adjust concurrent processing count
- Optimize database queries

Log Analysis

# View server logs
tail -f logs/mcp-server.log

# View quality gate logs
tail -f logs/quality-gate.log

# View agent logs
tail -f logs/linguist.log

Contributing Guide

Development Environment Setup

Fork the project repository
Create feature branch
Commit code changes
Create Pull Request

Code Standards

Use TypeScript strict mode
Follow ESLint configuration
Add appropriate unit tests
Update relevant documentation

License

This project is licensed under the MIT License.

Contact

Email: raywong2121@gmail.com

Acknowledgments

Thanks to the following organizations and projects for their support:

OHDSI (Observational Health Data Sciences and Informatics)
SNOMED International
Model Context Protocol (MCP) Community

Note: This project is for research and educational purposes only. Before using in production environments, please ensure compliance with relevant medical regulations and data protection requirements.