maziyarpanahi/openmed

What it solves

OpenMed provides a local-first framework for processing clinical text, specifically focusing on the extraction of medical entities (NER) and the de-identification of Personally Identifiable Information (PII). It allows healthcare providers and researchers to turn unstructured clinical notes into structured data while keeping sensitive patient data on their own hardware to maintain privacy and security.

How it works

OpenMed uses a curated catalog of specialized medical NER models and PII filters. It offers a local runtime that can be deployed across various environments, including Python, Swift (via OpenMedKit), Android (via ONNX Runtime Mobile), and web browsers (via Transformers.js). It supports multiple execution backends such as CPU, CUDA, and Apple's MLX for acceleration on Apple Silicon.

Who it’s for

It is designed for developers and healthcare organizations that need to process clinical text locally to avoid sending sensitive PHI (Protected Health Information) to cloud APIs, as well as those building medical AI agents or mobile health applications.

Highlights

• Local-First Execution: Processes data locally on the user's hardware after model artifacts are downloaded.
• Broad Hardware Support: Optimized for Apple Silicon (MLX), Android (ONNX), and browser-based inference (WebGPU).
• Comprehensive PII Tools: Supports masking, replacing, hashing, and date-shifting for de-identification across 33+ model-backed languages.
• Specialized Medical Models: Includes curated models for detecting diseases, medications, anatomy, and genes.
• Agent-Ready: Provides an MCP server and a skills catalog for integration with coding agents like Claude Code and OpenAI Codex.