// the find
maziyarpanahi/openmed
Local-first healthcare AI: clinical NER & HIPAA PII de-identification that runs 100% on-device. 1,000+ medical models, 12 languages, Apple MLX + Python, no cloud, no patient data leaving your network. Apache-2.0
OpenMed is a Python library for clinical NER and HIPAA PII de-identification that runs entirely on local hardware — no cloud calls, no data egress. It wraps 1,000+ specialized biomedical models from Hugging Face and ships a FastAPI service, batch processor, and a Swift/MLX path for native iOS/macOS apps. The target is any team that needs to process patient records but can't send them to a cloud API.
The on-device constraint is real, not marketing: the MLX backend and CoreML export path mean you can actually ship this in an iOS app with no network permission required. The PII coverage is unusually thorough — all 18 HIPAA Safe Harbor identifiers, smart entity merging to avoid tokenization fragmentation, format-preserving Faker-backed replacement with locale-aware clinical ID providers (NPI, CPF, BSN, etc.). The model name fallback across backends is a genuine quality-of-life win: one name works on CPU, CUDA, and Apple Silicon without code changes. The golden-fixture eval harness and quality gates show someone is actually measuring regression rather than hoping the model is good.
The 'HIPAA-aware' claim needs scrutiny before you rely on it legally — HIPAA compliance is an operational and organizational posture, not something a library can certify; the README never says this and most users will misread it. The 1,000+ model count is real but the registry is opaque: no benchmark numbers on the page, no indication of which models are actively maintained vs. experimental one-offs. The REST service has no auth layer at all — it's documented as something you just expose on 0.0.0.0:8080, which is fine for internal deployments but a footgun if someone puts it behind a load balancer without thinking. Swift/OpenMedKit is still using GLiNER-family zero-shot as 'experimental', meaning the iOS story for clinical NER beyond PII is not production-ready yet.