Mimic MCP Server

🚀 MCP Server × MIMIC-IV Based on Azure Database for PostgreSQL

Evaluate whether the Model Context Protocol (MCP) Server deployed on Azure Database for PostgreSQL can match or exceed the performance of direct SQL in large-scale clinical analysis while reducing the workload of developers.

🚀 Quick Start

The project aims to assess whether the Model Context Protocol (MCP) Server deployed on Azure Database for PostgreSQL can match or outperform direct SQL in large-scale clinical analysis and reduce developers' workload.

✨ Features

• Compare the performance of direct SQL and MCP Server → PostgreSQL on the same workload.
• Provide scripts for creating schemas and importing MIMIC-IV CSV files.
• Offer a forked version of Azure PostgreSQL MCP with environment configuration.

📦 Installation

Prerequisites

• Python 3.10+
• Node 18+ (optional, for additional MCP tools)
• Azure Database for PostgreSQL – Flexible Server (Standard D4s v5 or larger)
• MIMIC-IV credentials and data download permissions

# Core libraries
pip install psycopg[binary] rich click pandas matplotlib
# Benchmarking
pip install httpx pytest pytest‑benchmark
# MCP Server
pip install mcp[cli] azure-identity azure-mgmt-postgresqlflexibleservers

Dataset Loading

cd data-loading
psql "$PGURI" -f 00_create_mimiciv_note_schema.sql
psql "$PGURI" -f 01_create_emar_tables.sql
python 02_bulk_copy.py --csv-root /path/to/mimic-iv

Estimated loading time on a standard D4s: Approximately 45 minutes for 72 GB (11.7 M note rows + 87 M EMAR rows).

Starting the MCP Server

cd mcp-server
python azure_postgresql_mcp.py

Environment variables (PGHOST, PGUSER, PGPASSWORD, PGDATABASE) must be set or provided through your Windows desktop / VS Code configuration.

If you need to use Microsoft Entra authentication instead of a password, set:

export AZURE_USE_AAD=True
export AZURE_SUBSCRIPTION_ID=...
export AZURE_RESOURCE_GROUP=...

💻 Usage Examples

Running the Baseline

Direct SQL

python benchmarks/run_sql.py --query count_emar_detail

Output example:

count=87,371,064 elapsed=101550.2 ms

MCP Server

python benchmarks/run_mcp.py --query count_emar_detail

Example output:

count=87,371,064 elapsed=101550.2 ms

📚 Documentation

Problem Statement

Clinical datasets like MIMIC-IV contain tens of millions of rows and hundreds of attributes. Writing and maintaining raw SQL for such extensive data operations is error-prone and inefficient. We pose the following question:

Can the abstractions based on MCP simultaneously improve developer productivity and maintain (or enhance) runtime efficiency compared to traditional SQL-only workflows?

To answer this question, we compared direct SQL and the MCP Server → PostgreSQL path on the same workloads - from simple COUNT(*) on mimiciv_hosp.emar_detail (approximately 87 M rows) to multi-table note aggregations.

Repository Layout

.
├── mimic-postgres/          # Scripts to create schema and import MIMIC-IV CSV files
├── src/            # Fork of Azure PostgreSQL MCP with environment configuration
└── README.md              # This file

Future Work

• Further optimize the performance of the MCP Server on Azure Database for PostgreSQL.
• Explore new integration methods with the MIMIC-IV dataset to support more complex queries and analyses.
• Develop a user-friendly interface for easier management and monitoring of the MCP Server.

📄 License

The project code follows the [MIT License], and the dataset uses the [MIMIC-IV Terms of Use]. For citation, please refer to [Related Literature].

👏 Acknowledgments

Thanks to Microsoft Azure for providing computing resources and the MIMIC-IV community for their support.