openmed 1.4.0

OpenMed delivers state-of-the-art biomedical and clinical LLMs that rival proprietary enterprise stacks, unifying model discovery, advanced extractions, and one-line orchestration.

OpenMed

Production-ready medical NLP toolkit powered by state-of-the-art transformers

Transform clinical text into structured insights with a single line of code. OpenMed delivers enterprise-grade entity extraction, assertion detection, and medical reasoning—no vendor lock-in, no compromise on accuracy.

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:<35} {entity.confidence:.2f}")
# DISEASE      chronic myeloid leukemia            0.98
# DRUG         imatinib                            0.95

---

✨ Why OpenMed?

• Specialized Models: 12+ curated medical NER models outperforming proprietary solutions
• HIPAA-Compliant PII Detection: Smart de-identification with all 18 Safe Harbor identifiers
• One-Line Deployment: From prototype to production in minutes
• Dockerized REST API: FastAPI endpoints for service deployments
• Batch Processing: Multi-file workflows with progress tracking
• Production-Ready: Configuration profiles, profiling tools, and medical-aware tokenization
• Zero Lock-In: Apache 2.0 licensed, runs on your infrastructure

---

Quick Start

Installation

# From a local checkout of this repo:
# Install with Hugging Face support
uv pip install -e ".[hf]"

# Or include REST service dependencies
uv pip install -e ".[hf,service]"

Apple Silicon acceleration in Python:

uv pip install -e ".[mlx]"

Swift apps on macOS and iOS use OpenMedKit. As of 1.4.0, that means:

• MLX on Apple Silicon macOS and real iPhone/iPad hardware for supported OpenMed PII, OpenAI Privacy Filter, OpenAI Nemotron Privacy Filter, OpenMed Multilingual Privacy Filter, and experimental GLiNER-family artifacts
• CoreML when you already have a bundled Apple model package or want the fallback Apple path

Add the Swift package like this:

dependencies: [
    .package(url: "https://github.com/maziyarpanahi/openmed.git", from: "1.4.0"),
]

OpenMedKit is public and now supports native MLX runtime paths for PII token classification, Privacy Filter, and experimental GLiNER-family zero-shot tasks. The broader OpenMed model-packaging flow is still being hardened across the full collection, so treat conversion as active work rather than a fully universal public release surface.

For published releases, the editable install examples above can be replaced with plain uv pip install "openmed[...]".

Three Ways to Use OpenMed

1️⃣ Python API — One-liner for scripts and notebooks

from openmed import analyze_text

result = analyze_text(
    "Patient received 75mg clopidogrel for NSTEMI.",
    model_name="pharma_detection_superclinical"
)

2️⃣ REST API Service — FastAPI endpoints for app backends

uvicorn openmed.service.app:app --host 0.0.0.0 --port 8080

3️⃣ Batch Processing — Programmatic multi-document workflows

from openmed import BatchProcessor

processor = BatchProcessor(
    model_name="disease_detection_superclinical",
    confidence_threshold=0.55,
    group_entities=True,
)

result = processor.process_texts([
    "Patient started metformin for type 2 diabetes.",
    "Imatinib started for chronic myeloid leukemia.",
])

---

Key Features

Core Capabilities

• Curated Model Registry: Metadata-rich catalog with 12+ specialized medical NER models
• PII Detection & De-identification: HIPAA-compliant de-identification with smart entity merging
• Medical-Aware Tokenization: Clean handling of clinical patterns (COVID-19, CAR-T, IL-6)
• Advanced NER Processing: Confidence filtering, entity grouping, and span alignment
• Multiple Output Formats: Dict, JSON, HTML, CSV for any downstream system

Production Tools (v1.4.0)

• Batch Processing: Multi-text and multi-file workflows with progress tracking
• Configuration Profiles: dev/prod/test/fast presets with flexible overrides
• Performance Profiling: Built-in inference timing and bottleneck analysis
• Dockerized REST API: GET /health, POST /analyze, POST /pii/extract, POST /pii/deidentify
• Service Reliability Hardening: request validation, shared pipeline preload, and timeout/error envelopes

---

Documentation

Comprehensive guides available at openmed.life/docs

Quick links:

• Getting Started — Installation and first analysis
• Analyze Text Helper — Python API reference
• MLX Backend — Apple Silicon Python runtime
• OpenMedKit (Swift) — Use OpenMed models in macOS and iOS apps
• PII Detection Guide — Complete de-identification tutorial (v0.5.0)
• Batch Processing — Multi-text and multi-file workflows
• Configuration Profiles — Environment-specific presets
• REST Service — FastAPI and Docker usage
• Model Registry — Browse available models
• Configuration — Settings and environment variables

---

REST API

OpenMed includes a Docker-friendly FastAPI service with reliability hardening:

• GET /health
• POST /analyze
• POST /pii/extract
• POST /pii/deidentify

Run locally

uv pip install -e ".[hf,service]"
uvicorn openmed.service.app:app --host 0.0.0.0 --port 8080

Optional shared model warm-up:

OPENMED_SERVICE_PRELOAD_MODELS=disease_detection_superclinical,OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1 \
uvicorn openmed.service.app:app --host 0.0.0.0 --port 8080

Run with Docker

docker build -t openmed:1.4.0 .
docker run --rm -p 8080:8080 -e OPENMED_PROFILE=prod openmed:1.4.0

Example request

curl -X POST http://127.0.0.1:8080/pii/extract \
  -H "Content-Type: application/json" \
  -d '{"text":"Paciente: Maria Garcia, DNI: 12345678Z","lang":"es"}'

See the full service guide at REST Service docs.

Non-2xx responses now use a unified envelope:

{
  "error": {
    "code": "validation_error",
    "message": "Request validation failed",
    "details": [
      {
        "field": "body.text",
        "message": "Text must not be blank",
        "type": "value_error"
      }
    ]
  }
}

---

Models

OpenMed includes a curated registry of 12+ specialized medical NER models:

Model	Specialization	Entity Types	Size
disease_detection_superclinical	Disease & Conditions	DISEASE, CONDITION, DIAGNOSIS	434M
pharma_detection_superclinical	Drugs & Medications	DRUG, MEDICATION, TREATMENT	434M
pii_detection_superclinical	PII & De-identification	NAME, DATE, SSN, PHONE, EMAIL, ADDRESS	434M
anatomy_detection_electramed	Anatomy & Body Parts	ANATOMY, ORGAN, BODY_PART	109M
gene_detection_genecorpus	Genes & Proteins	GENE, PROTEIN	109M

📖 Full Model Catalog

---

Advanced Usage

PII Detection & De-identification (v0.5.0)

from openmed import extract_pii, deidentify

# Extract PII entities with smart merging (default)
result = extract_pii(
    "Patient: John Doe, DOB: 01/15/1970, SSN: 123-45-6789",
    model_name="pii_detection_superclinical",
    use_smart_merging=True  # Prevents entity fragmentation
)

# De-identify with multiple methods
masked = deidentify(text, method="mask")        # [NAME], [DATE]
removed = deidentify(text, method="remove")     # Complete removal
replaced = deidentify(text, method="replace")   # Faker-backed locale-aware fakes
hashed = deidentify(text, method="hash")        # Cryptographic hashing
shifted = deidentify(text, method="shift_dates", date_shift_days=180)

# Deterministic obfuscation: same input -> same surrogate within doc
deidentify(text, method="replace", lang="pt", locale="pt_BR",
           consistent=True, seed=42)

Smart Entity Merging (NEW in v0.5.0): Fixes tokenization fragmentation by merging split entities like dates (01/15/1970 instead of 01 + /15/1970), ensuring production-ready de-identification.

Faker-backed obfuscation (v1.3.0): method="replace" uses Faker with custom providers for clinical IDs (CPF, CNPJ, BSN, NIR, Codice Fiscale, NIE, Aadhaar, Steuer-ID, NPI). Surrogates are locale-aware, format-preserving, and optionally deterministic. See Anonymization Guide.

HIPAA Compliance: Covers all 18 Safe Harbor identifiers with configurable confidence thresholds.

📓 Complete PII Notebook | 📖 Smart Merging | 📖 Anonymization

Privacy Filter Family (Public)

OpenMed ships three Privacy Filter families on the OpenAI Privacy Filter architecture — same model code (gpt-oss-style sparse-MoE transformer with local attention, sink tokens, RoPE+YaRN, tiktoken o200k_base tokenization), different training data:

Variant	Trained on	PyTorch (CPU + CUDA)	MLX full (OpenMedKit + Apple Silicon)	MLX 8-bit (OpenMedKit + Apple Silicon)
OpenAI Privacy Filter	OpenAI's PII training set	openai/privacy-filter	OpenMed/privacy-filter-mlx	OpenMed/privacy-filter-mlx-8bit
Nemotron-PII fine-tune	Nemotron PII dataset	OpenMed/privacy-filter-nemotron	OpenMed/privacy-filter-nemotron-mlx	OpenMed/privacy-filter-nemotron-mlx-8bit
OpenMed Multilingual Privacy Filter	OpenMed multilingual PII corpus with official support for 16 languages	OpenMed/privacy-filter-multilingual	OpenMed/privacy-filter-multilingual-mlx	OpenMed/privacy-filter-multilingual-mlx-8bit

All model IDs above route through the same extract_pii() / deidentify() API — only the model_name= argument changes.

The MLX artifacts above use the OpenMed MLX artifact layout consumed by OpenMedKit for native macOS, iOS, and iPadOS apps.

Install

The PyTorch path runs anywhere (Linux, macOS, Windows; CPU or CUDA):

pip install "openmed[hf]"

The MLX path adds Apple Silicon acceleration on top of [hf]:

pip install "openmed[mlx]"          # MLX runtime + tiktoken + huggingface-hub

tiktoken (the OpenAI Privacy Filter tokenizer) ships in the [mlx] extra.

Use it (PyTorch — Linux / Windows / non-Apple-Silicon macOS)

from openmed import extract_pii, deidentify

text = ("Patient Sarah Connor (DOB: 03/15/1985) at MRN 4471882. "
        "Email: [e-mail], phone [phone].")

# OpenAI baseline — runs on CPU/CUDA via Transformers
result = extract_pii(text, model_name="openai/privacy-filter")

# Nemotron-PII fine-tune — same code path, different weights
result = extract_pii(text, model_name="OpenMed/privacy-filter-nemotron")

# De-identify with any method
deidentify(text, model_name="OpenMed/privacy-filter-nemotron", method="mask")
deidentify(text, model_name="OpenMed/privacy-filter-nemotron",
           method="replace", consistent=True, seed=42)

Use it (MLX — Apple Silicon)

from openmed import extract_pii, deidentify

text = "Patient Sarah Connor born on 03/15/1985, MRN 4471882."

# OpenAI baseline (full / 8-bit MLX artifacts)
extract_pii(text, model_name="OpenMed/privacy-filter-mlx")
extract_pii(text, model_name="OpenMed/privacy-filter-mlx-8bit")

# Nemotron-PII fine-tune (full / 8-bit MLX artifacts)
extract_pii(text, model_name="OpenMed/privacy-filter-nemotron-mlx")
extract_pii(text, model_name="OpenMed/privacy-filter-nemotron-mlx-8bit")

# OpenMed Multilingual Privacy Filter (full / 8-bit MLX artifacts)
extract_pii(text, model_name="OpenMed/privacy-filter-multilingual-mlx")
extract_pii(text, model_name="OpenMed/privacy-filter-multilingual-mlx-8bit")

Cross-platform note

The MLX artifact names work everywhere — on a non-Apple-Silicon host (or anywhere MLX isn't installed) the request is automatically substituted with the matching PyTorch model and a one-time UserWarning names the substitution. The substitution is family-aware:

• OpenMed/privacy-filter-mlx* ⇒ falls back to openai/privacy-filter
• OpenMed/privacy-filter-nemotron-mlx* ⇒ falls back to OpenMed/privacy-filter-nemotron
• OpenMed/privacy-filter-multilingual-mlx* ⇒ falls back to OpenMed/privacy-filter-multilingual

So your code can ship an MLX model name and run on any host without changes — Apple Silicon users get MLX speed, everyone else gets the same family's PyTorch checkpoint.

📖 Privacy Filter Architecture & Backend Routing | ▶ Side-by-side example | ▶ Faker obfuscation demo

Multilingual PII (9 Languages)

OpenMed now supports multilingual PII extraction and de-identification across en, fr, de, it, es, nl, hi, te, and pt. French, German, Italian, and Spanish expose the full 35-model family; Portuguese ships 31 public API-visible models; Dutch, Hindi, and Telugu currently ship one flagship public model each, bringing the total PII catalog to 210 models.

uv pip install "openmed[hf]" && python -c "from openmed import extract_pii; print([(e.label,e.text) for e in extract_pii('Dr. Pedro Almeida, CPF: 123.456.789-09, email: [e-mail], tel: [phone]', lang='pt').entities])"

from openmed import extract_pii

portuguese = extract_pii(
    "Paciente: Pedro Almeida, CPF: 123.456.789-09, email: [e-mail], telefone: [phone]",
    lang="pt",
    model_name="OpenMed/OpenMed-PII-Portuguese-SnowflakeMed-Large-568M-v1",
    use_smart_merging=True,
)

dutch = extract_pii(
    "Patiënt: Eva de Vries, geboortedatum: 15 januari 1984, BSN: 123456782, telefoon: [phone]",
    lang="nl",
    model_name="OpenMed/OpenMed-PII-Dutch-SuperClinical-Large-434M-v1",
    use_smart_merging=True,
)

hindi = extract_pii(
    "रोगी: अनीता शर्मा, जन्मतिथि: 15 जनवरी 1984, फोन: [phone], पता: 12 गली संख्या 5, नई दिल्ली 110001",
    lang="hi",
    model_name="OpenMed/OpenMed-PII-Hindi-SuperClinical-Large-434M-v1",
    use_smart_merging=True,
)

telugu = extract_pii(
    "రోగి: సితా రెడ్డి, జన్మ తేదీ: 15 జనవరి 1984, ఫోన్: [phone], చిరునామా: 12 వీధి 5, హైదరాబాద్ 500001",
    lang="te",
    model_name="OpenMed/OpenMed-PII-Telugu-SuperClinical-Large-434M-v1",
    use_smart_merging=True,
)

print([(e.label, e.text) for e in portuguese.entities])
print([(e.label, e.text) for e in dutch.entities])
print([(e.label, e.text) for e in hindi.entities])
print([(e.label, e.text) for e in telugu.entities])

Batch Processing

from openmed import BatchProcessor, OpenMedConfig

config = OpenMedConfig.from_profile("prod")
processor = BatchProcessor(
    model_name="disease_detection_superclinical",
    config=config,
    group_entities=True,
)

result = processor.process_texts([
    "Metastatic breast cancer treated with trastuzumab.",
    "Acute lymphoblastic leukemia diagnosed.",
])

Configuration Profiles

from openmed import analyze_text

# Apply a profile programmatically
result = analyze_text(
    text,
    model_name="disease_detection_superclinical",
    config_profile="prod"  # High confidence, grouped entities
)

Performance Profiling

from openmed import analyze_text, profile_inference

with profile_inference() as profiler:
    result = analyze_text(text, model_name="disease_detection_superclinical")

print(profiler.summary())  # Inference time, bottlenecks, recommendations

📖 More Examples

---

Contributing

We welcome contributions! Whether it's bug reports, feature requests, or pull requests.

• 🐛 Found a bug? Open an issue

---

Credits & Acknowledgements

OpenMed builds on excellent open-source work from the community. Particular thanks to:

• OpenAI for open-sourcing the Privacy Filter model. The OpenAI Privacy Filter architecture (gpt-oss-style sparse-MoE transformer with local attention, sink tokens, RoPE+YaRN, tiktoken o200k_base tokenization) is the foundation of OpenMed's privacy-filter family — both the upstream baseline and our Nemotron-PII fine-tunes share this architecture.
• NVIDIA for releasing the Nemotron PII dataset, which we used to fine-tune the OpenAI Privacy Filter weights into the OpenMed/privacy-filter-nemotron family (MLX, MLX 8-bit).
• HuggingFace for transformers, tokenizers, huggingface_hub, and the broader model-distribution ecosystem.
• Apple for MLX, which powers OpenMed's Apple Silicon acceleration.
• The Faker maintainers for Faker and its community-contributed locale providers, which power OpenMed's locale-aware obfuscation engine.

---

License

OpenMed is released under the Apache-2.0 License.

---

Citation

If you use OpenMed in your research, please cite:

@misc{panahi2025openmedneropensourcedomainadapted,
      title={OpenMed NER: Open-Source, Domain-Adapted State-of-the-Art Transformers for Biomedical NER Across 12 Public Datasets},
      author={Maziyar Panahi},
      year={2025},
      eprint={2508.01630},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.01630},
}

---

Star History

If you find OpenMed useful, consider giving it a star ⭐ to help others discover it!

---

Built with ❤️ by the OpenMed team

🌐 Website • 📚 Documentation • 🐦 X/Twitter • 💬 LinkedIn