Local-First Healthcare AI: Clinical NER and HIPAA-Compliant PII De-Identification
OpenMed is an open-source framework designed to process clinical text locally, ensuring patient data never leaves the controlled hardware environment. It specializes in Named Entity Recognition (NER) for medical contexts and HIPAA-compliant Protected Health Information (PHI) de-identification, all executed on-device without reliance on cloud services. The framework supports 2,266+ medical models across 35 languages, including English, Spanish, French, German, Arabic, Hindi, Japanese, Persian, and Kiswahili, among others. OpenMed operates under the Apache 2.0 license, ensuring permissive use and modification rights for developers and organizations.
The core runtime of OpenMed performs extraction and de-identification tasks once the required model artifacts are downloaded. While model downloads and certain integrations may require network access, the actual processing of patient data remains entirely local. This design ensures compliance with privacy regulations such as HIPAA, GDPR, and other data protection laws by eliminating the need for cloud-based processing or external data transmission.
OpenMed is built to integrate seamlessly with existing workflows, offering Python, Apple MLX, Android (via ONNX Runtime Mobile), and browser-based (Transformers.js) implementations. The framework is compatible with Python 3.10+, and its modular architecture allows for customization to meet specific organizational needs. For instance, the Nemotron Privacy Filter model can redact names, addresses, IDs, and billing data from clinical discharge notes in real-time, all while maintaining the integrity of the original text structure.
How It Works
OpenMed processes clinical text through a structured pipeline that extracts medical entities, detects PII, and generates de-identified output—all without sending data to a cloud API. The workflow begins with input clinical text, which is then analyzed by OpenMed’s local runtime. The system identifies medical entities such as diseases, drugs, and procedures while simultaneously detecting and redacting PII, including patient names, addresses, medical record numbers, and contact details. The final output is a structured, de-identified version of the original text, preserving clinical context while removing sensitive information.
The framework leverages a combination of pre-trained models and customizable configurations to achieve high accuracy in entity recognition and PII detection. For example, the Privacy Filter family of models—built on the OpenAI Privacy Filter architecture—includes variants such as the OpenAI Privacy Filter, Nemotron-PII fine-tune, and OpenMed Multilingual. All models are designed to operate locally, ensuring no data leaves the user’s infrastructure.
The framework also provides tools for offline and air-gapped environments, where models can be loaded from local directories without contacting external repositories.
Use Cases
OpenMed is primarily designed for healthcare providers, research institutions, and organizations handling sensitive clinical data. Its core functionality—clinical NER and PII de-identification—enables secure processing of patient records for purposes such as medical research, data analytics, and interoperability without compromising privacy.
The framework’s support for 21 languages makes it particularly valuable for multinational healthcare organizations or those serving diverse patient populations. OpenMed can process clinical notes in languages such as Spanish, Arabic, Hindi, and Japanese, ensuring that language barriers do not hinder privacy-preserving data processing. This multilingual capability is further enhanced by the inclusion of locale-specific PII patterns, such as Persian date handling, Iranian national ID surrogates, and Eastern-digit normalization for languages like Arabic and Hindi.
The framework’s ability to export structured data in formats like FHIR (Fast Healthcare Interoperability Resources) and OMOP (Observational Medical Outcomes Partnership) Common Data Model facilitates integration with clinical research databases and analytics platforms. Additionally, OpenMed’s compatibility with agents like Claude Code, OpenAI Codex, and OpenCode enables developers to build automated pipelines for de-identification and clinical data extraction, further streamlining workflows in both clinical and research settings.
Key Features & Capabilities
OpenMed’s architecture is optimized for performance, privacy, and flexibility. On Apple Silicon devices, the framework achieves 24–33× faster inference speeds compared to CPU-based PyTorch implementations when using the Privacy Filter model. This acceleration is particularly beneficial for real-time processing of clinical documents, such as de-identifying patient notes during admission or discharge workflows. Similarly, OpenMed’s support for ONNX Runtime Mobile on Android ensures efficient local processing on mobile devices, with models available in fp32, fp16, and INT8 formats to balance accuracy and performance.
The framework includes advanced features for clinical data extraction and structuring. For instance, OpenMed 2.5 introduces a ClinicalPrivacyProcessor that combines explicit language controls with document-level review status to preserve clinical context while redacting PII. This includes retaining diagnoses, medications, doses, and measurements in the de-identified output, which is critical for maintaining the utility of clinical data after privacy processing.
OpenMed provides robust support for healthcare data standards such as FHIR and OMOP. The to_fhir() method in OpenMed 2.5 allows users to export de-identified clinical entities into FHIR resources like Condition, Observation, MedicationStatement, and Procedure, assembling them into deterministic Bundles. This capability is particularly useful for integrating de-identified data into EHR systems or clinical data repositories. Additionally, OpenMed includes dependency-free FHIR R4 structural checks and US Core 9.0.0 validation to ensure compliance with interoperability standards.
The framework also addresses the needs of air-gapped or offline environments, where connectivity to external repositories is restricted. OpenMed allows users to specify local model directories, enabling the system to load models without contacting the Hugging Face Hub or other online sources.
OpenMed’s compatibility with multiple platforms and programming languages extends its utility across diverse technical environments. Additionally, OpenMed’s Agent Skills catalog allows developers to build portable, on-device pipelines for de-identification and clinical NER, which can be deployed in coding agents like Claude Code or OpenAI Codex without exposing sensitive data to cloud services.
Performance and Optimization
OpenMed is designed with performance in mind, particularly for resource-constrained environments. The framework’s support for Apple MLX acceleration on Apple Silicon devices reduces inference latency, making it suitable for real-time clinical workflows. Similarly, OpenMed’s ONNX Runtime Mobile integration on Android optimizes model execution for mobile devices, with support for quantized models (fp16, INT8) to further enhance performance without sacrificing accuracy.
This capability is particularly valuable for organizations processing large datasets, such as research institutions or healthcare systems conducting retrospective analyses. OpenMed also supports streaming digests and media-type detection, enabling efficient handling of diverse document formats, including PDFs, XLSX files, and PPTX presentations.
OpenMed’s model catalog includes over 2,266 entries, covering a wide range of clinical and PII detection tasks. The Privacy Filter family, for instance, includes three variants: the OpenAI Privacy Filter, Nemotron-PII fine-tune, and OpenMed Multilingual. Each variant is optimized for specific use cases, such as general PII detection, multilingual support, or fine-tuned performance on clinical datasets.