Patient records can't leave the intranet, so OpenMed runs medical NLP on a local workstation

1. Why you need this tool

Director Wang at a hospital IT department recently ran into a nasty problem: the department wanted to deploy a clinical NLP system to extract diseases, drugs and dosages from electronic medical records. Security killed it with one sentence — "patient records cannot leave the hospital intranet." He tried three cloud vendors in a row, each requiring a security review and a data processing agreement, with quotes starting at ¥360k a year.

This isn't an isolated case. Compliance requirements around medical data leave a huge number of teams stuck at the first step of "let's use AI for something": HIPAA, GDPR and internal hospital network policies all say the same thing — don't send PHI to an external API. But once you demand 100% localization, the open-source options that can actually run clinical NER are few and far between, and most only support English.

OpenMed targets exactly this contradiction: runs locally, a huge catalog of medical models, multilingual PII coverage, and de-identification built in. 4,976 stars on GitHub, over 3.4 million monthly downloads on PyPI, and already running steadily on intranet workstations in a number of hospitals.

2. What it is, in one sentence

OpenMed is a local-first medical AI toolkit built around two main threads: clinical named entity recognition (NER) and HIPAA-aligned PII de-identification. 2,200+ medical models, 33 PII languages, Apache-2.0. The core runtime is fully offline by default — data never leaves the intranet — and deploys to CPU, CUDA, Apple MLX, Android ONNX and the browser. The SDK is Python, with Swift/Kotlin on mobile, plus a REST service, FHIR/HL7 interoperability and Transformers.js export already in place.

3. Core features in depth

1. Clinical NER: diseases, drugs, anatomy, genes

OpenMed's model catalog covers five verticals: disease_detection_superclinical (disease/diagnosis, 434M), pharma_detection_superclinical (drug/treatment, 434M), pii_superclinical_large (PII, 434M), anatomy_detection_electramed (anatomy/organs, 109M) and gene_detection_genecorpus (genes/proteins, 109M). One line of code calls them:

from openmed import analyze_text
result = analyze_text(
    "Patient started on imatinib for chronic myeloid leukemia.",
    model_name="disease_detection_superclinical",
)
for entity in result.entities:
    print(f"{entity.label:<12} {entity.text:<28} {entity.confidence:.2f}")

Output: DISEASE chronic myeloid leukemia 0.98; DRUG imatinib 0.95. This out-of-the-box API shape means clinical NLP no longer needs weeks of annotation and fine-tuning.

2. PII de-identification: HIPAA Safe Harbor aligned

extract_pii and deidentify form the privacy core. It supports 18 identifier types, covering names, birthdays, SSNs, phone numbers, emails, addresses, medical record numbers and more. Four masking methods: mask, replace (Faker swaps in same-language fake data), hash, and shift_dates (date shifting that preserves the time series).

Worth noting is the smart entity merging mechanism: traditional NER often slices 01/15/1970 into three fragments — 01, /15, /1970 — while OpenMed's merge strategy keeps a full date, phone number or ID as one entity, so downstream rule engines don't miss it.

3. 33 PII languages + national ID validation

From Arabic 29801011234567 to Japan's My Number 1234 5678 9012, the Dutch BSN, Brazil's CPF and Spain's DNI, OpenMed ships a matching Faker locale and validator for each. A single lang="ja" switches it; no manual loading of multiple models.

4. Multi-platform deployment: Python / Swift / Kotlin / browser

The same ONNX model runs on a Python CPU, on Android ONNX Runtime Mobile, on Apple MLX on iOS, and in the browser via Transformers.js + WebGPU. For field scenarios that need offline record entry on a phone, OpenMedKit (Swift) provides a scan-and-de-identify pipeline.

4. Technical highlights and performance numbers

MLX acceleration shines on Apple Silicon: Privacy Filter inference latency is 24–33× faster than CPU PyTorch (median inference latency). Batch PII processing improves throughput 3.3× on CPU and 2.2× on MLX compared with per-document processing.

A few architectural points worth noting:

• OpenAI Privacy Filter architecture as the base (gpt-oss-style sparse MoE + local attention + sink tokens + RoPE+YaRN + tiktoken o200k_base), with OpenMed providing three variants (baseline, Nemotron fine-tune, multilingual) behind the same API;
• Automatic backend routing: an MLX model name falls back to the PyTorch checkpoint on non-Apple hardware;
• Offline/air-gapped deployment: point model_id at a local directory and the runtime never contacts the Hugging Face Hub again;
• Ecosystem compatibility: PyTorch, CUDA, MLX, CoreML, ONNX Runtime Mobile, Transformers.js and WebGPU all covered;
• REST service exposes four endpoint groups — /health, /analyze, /pii/extract, /pii/deidentify — with keep_alive, hot model unloading, API key/JWT, dynamic batching, rate limiting and observability metrics.

5. Use cases and who it's for

Typical scenarios:

• A hospital IT department needs to structure ten years of electronic records in bulk and extract entities, without data leaving the intranet;
• An insurtech company de-identifies claims text on a local workstation before feeding it downstream;
• Cross-border multilingual record research that needs to extract from Portuguese, Arabic, Japanese and Turkish mixed together;
• A mobile follow-up app lets doctors enter records offline on an iPad and automatically masks patient names;
• Academic research pipes de-identified clinical data into cohort analysis via FHIR/HL7.

Who it's for: healthcare IT engineers, clinical researchers, medical NLP teams, medical SaaS startups, and integration vendors that need compliance audits. Who it's not for: pure algorithm hobbyists who just want to run a few public dataset demos and don't care about local deployment — cloud inference on Hugging Face is less hassle.

6. How it compares

There's really only one core reason to choose OpenMed: it's one of the few open-source options that simultaneously satisfies "rich vertical medical models," "fully local" and "multilingual PII built in" — and the SDK is Apache-2.0, so there's no licensing friction for commercial use.

7. Getting started fast

1. Install (30 seconds)

# CPU / CUDA
pip install --upgrade "openmed[hf]"

# Apple Silicon acceleration
pip install --upgrade "openmed[mlx]"

# REST service
pip install --upgrade "openmed[hf,service]"

2. Minimal working example

from openmed import analyze_text

result = analyze_text(
    "Patient received 75mg clopidogrel for NSTEMI.",
    model_name="pharma_detection_superclinical",
)
print([(e.label, e.text) for e in result.entities])

3. Further tips

• In an intranet environment, point model_id at a pre-downloaded local directory so it runs even when the network is down;
• Use BatchProcessor(operation="deidentify", batch_size=16) for tens of thousands of documents;
• For production, start the REST service from the Docker image with OPENMED_PROFILE=prod;
• For Android integration see android/README.md, for iOS see swift/OpenMedKit;
• Agent developers can run ./install-skills.sh to inject the skill into Claude Code / Codex / OpenCode.

8. Takeaways and next steps

OpenMed breaks the default assumption that "medical NLP has to go to the cloud." 2,200+ models, 33 PII languages, 24× speedup on Apple MLX, Apache-2.0 commercial friendliness, plus full iOS/Android/browser coverage — this is no longer a demo project but an engineered solution you can drop onto a hospital intranet workstation.

How to use it: first run disease_detection_superclinical against real samples from your department and confirm entity accuracy; then add pii_superclinical_large as a compliance guardrail; finally connect it to your existing ETL through the REST service or Docker image. If you're working on a multilingual project, run the lang="pt", lang="ja" and lang="ar" end-to-end demos — see for yourself whether the Faker-generated fake data matches local conventions.

• GitHub: https://github.com/maziyarpanahi/openmed
• Website: https://openmed.life/
• Docs: https://openmed.life/docs/
• PyPI: https://pypi.org/project/openmed/
• Paper: https://arxiv.org/abs/2508.01630

One-line pitch: if you're tired of the tug-of-war between "records can't leave the intranet" and "NLP accuracy," OpenMed is the open-source medical AI toolkit most worth running on your local workstation right now.