DEV Community

floworkos
floworkos

Posted on

OpenMed — Local-First Clinical NER & PII Detection, Genuinely Offline

What It Is

OpenMed is a Python library (with Swift/iOS bindings via OpenMedKit) for named-entity recognition, PII de-identification, and clinical text extraction that runs entirely on-device. No cloud calls, no API keys. The README claims 1,000+ specialized biomedical models from Hugging Face, support for 12+ languages, 247 PII checkpoints, and Apple MLX acceleration (24–33× faster on Silicon).

Who It's For

  • Healthcare teams building HIPAA-aware pipelines where patient data cannot leave the network
  • Clinicians & researchers who need structured extraction (disease, drug, procedure) from unstructured notes
  • Apple-first shops wanting native iOS/macOS on-device inference via OpenMedKit
  • Air-gapped / sovereign-AI environments where cloud connectivity is not an option

What's Genuinely Strong

  1. No-bullshit on-device commitment — the README doesn't claim "mostly local" or "hybrid." It shows actual demos (iPhone, PII redaction) and is explicit: clinical text never leaves the device.
  2. Practical multi-platform story — Python one-liner, REST service, batch processor, AND native Swift package in one codebase. That's rare.
  3. Real-world privacy angle — covers all 18 Safe Harbor PII identifiers, format-preserving anonymization, entity merging. Not just "remove [NAME]."
  4. Apache 2.0, no lock-in — zero vendor risk. You own the models, the inference, the data.
  5. Transparent performance claims — MLX speedup chart is specific (24–33×) and labeled "median latency per inference step," not marketing hyperbole.

One Real Trade-Off

Model quality & coverage unknown. The README says "1,000+ specialized models" and "many outperforming proprietary stacks" but provides no benchmarks, no F1 scores, no comparison studies. You don't know if "disease_detection_superclinical" will work for your notes vs. a general BERT. The arXiv link (2508.01630) may have details, but the README itself doesn't prove the models are clinically validated or production-ready. You're trusting Panahi's curation and the HF Hub labels — which may be fine, but it's a real bet.

One-Line Verdict

A genuinely privacy-first clinical AI toolkit with no cloud lock-in, but you'll need to validate model quality yourself before production — the README doesn't prove the 1,000+ models work at scale.


REPO: maziyarpanahi/openmed


🔗 Repo: https://github.com/maziyarpanahi/openmed

An honest review by the Flowork team — we read the README so you don't have to. We build open-source tooling too; this isn't a sponsored post.

Top comments (0)