Choosing Openmed Models

Choosing OpenMed Models

OpenMed ships a registry of clinical and biomedical NER models grouped into 12 categories. Never hardcode a model list — query the registry at runtime so your code stays correct as models are added. This skill helps you go from "I need to find diseases in Spanish discharge notes" to a concrete model key.

When to use

• The user knows the task (find diseases / tumors / PHI) but not the model.
• You need the right PII model for a language (es, fr, de, …).
• You want to filter models by size, task, or tier before loading.
• You want to inspect a model's labels, params, and license first.

Once you have a key, hand off to loading-openmed-models to load it.

Install

bashCopy code
pip install openmed         # registry queries work without the [hf] extra

Quick start: browse categories, then pick

pythonCopy code
import openmed

# 1) The 12 categories
openmed.list_model_categories()
# ['Medical', 'Privacy', 'Anatomy', 'Hematology', 'Chemical', 'Disease',
#  'Genomics', 'Oncology', 'Species', 'Pathology', 'Pharmaceutical', 'Protein']

# 2) Models in a category -> list[ModelInfo]
for m in openmed.get_models_by_category("Disease"):
    print(m.model_id, "|", m.size_category, "|", m.entity_types)

# 3) Inspect one model before loading
info = openmed.get_model_info("OpenMed/OpenMed-NER-DiseaseDetect-BigMed-278M")
print(info.display_name, info.task, info.param_count, info.license)

get_models_by_category and get_all_models return ModelInfo objects. get_all_models() returns a dict[str, ModelInfo] keyed by registry key.

What ModelInfo tells you

Every model exposes (real attributes):

textCopy code
model_id          # HF repo id, e.g. "OpenMed/OpenMed-NER-DiseaseDetect-BigMed-278M"
display_name      # human-friendly name
category          # one of the 12 categories
specialization    # e.g. "disease entity detection"
entity_types      # list[str] of labels the model emits, e.g. ["DISEASE", ...]
size_category     # "Tiny" | "Small" | "Medium" | "Large" | "XLarge"
recommended_confidence   # suggested confidence_threshold for this model
family            # "NER" | "PII" | ...
task              # "token-classification"
languages         # e.g. ["en"], ["es"]
param_count       # e.g. 278000000
license           # e.g. "apache-2.0"

Use entity_types to confirm the model emits the labels you need, and recommended_confidence as a sensible default confidence_threshold.

Disease vs Oncology vs Privacy: worked choices

pythonCopy code
import openmed

# Disease conditions in a general clinical note:
disease = openmed.get_models_by_category("Disease")
# e.g. "OpenMed/OpenMed-NER-DiseaseDetect-BigMed-278M"
#      "OpenMed/OpenMed-NER-DiseaseDetect-BioClinical-108M" (smaller/faster)

# Tumors, staging, oncologic findings -> Oncology, not Disease:
onco = openmed.get_models_by_category("Oncology")
# e.g. "OpenMed/OpenMed-NER-OncologyDetect-BigMed-278M"

# PHI / PII detection -> Privacy category:
privacy = openmed.get_models_by_category("Privacy")

Rule of thumb: bigger (278M/560M) = more accurate, slower; smaller (108M, "Small"/"Tiny") = faster, edge-friendly. Start with a mid-size model and size up only if recall is short.

Pick a PII model by language

pythonCopy code
import openmed

# All PII models for Spanish -> dict[str, ModelInfo]
es_models = openmed.get_pii_models_by_language("es")

# The recommended default PII model id for a language:
default_es = openmed.get_default_pii_model("es")
print(default_es)   # HF repo id, or None if unsupported

deidentify(..., lang="es") and extract_pii(..., lang="es") already select an appropriate default — use these helpers when you need to override or to confirm coverage. Supported de-id languages live in openmed.SUPPORTED_LANGUAGES (en es pt fr de it nl hi te ar tr ja).

Structured search with ModelQuery

For filtering by task, language, size, or tier, use the typed search:

pythonCopy code
from openmed import search_models, ModelQuery

results = search_models(ModelQuery(
    task="token-classification",
    language="en",
    max_params=200_000_000,   # keep it small for on-device
    license="apache-2.0",
))
for r in results:
    print(r.repo_id, r.param_count, r.languages, r.formats)

Each result is a ModelSearchResult with fields like repo_id, family, task, languages, tier, param_count, architecture, base_model, formats, canonical_labels, license, and released. ModelQuery filters include task, language, tier, max_params, min_params, format, license, and a free-text query.

Let OpenMed suggest a model from text

pythonCopy code
import openmed

for key, info, reason in openmed.get_model_suggestions(
    "Stage III adenocarcinoma with metastasis to regional lymph nodes."
):
    print(key, "->", reason)

get_model_suggestions(text) returns (registry_key, ModelInfo, reason) tuples — handy when the domain is unclear from the request.

CLI

bashCopy code
openmed models list                 # registry keys (add --include-remote to query the Hub)
openmed models info <registry-key>  # max sequence length for a key
openmed analyze --text "Stage III adenocarcinoma." --model oncology_detection_bigmed_278m

Hand-off to / from OpenMed

• To loading-openmed-models: pass the chosen model_id/registry key as model_name= to ModelLoader.load_model(...) or openmed.analyze_text(...).
• To extracting-clinical-entities: use the model's recommended_confidence as your confidence_threshold and verify entity_types matches your schema.
• To de-identification: feed get_default_pii_model(lang) into openmed.deidentify(model_name=..., lang=...).

pythonCopy code
import openmed
key = "oncology_detection_bigmed_278m"
info = openmed.get_model_info(key)
result = openmed.analyze_text(
    "Stage III adenocarcinoma with nodal metastasis.",
    model_name=key,
    confidence_threshold=info.recommended_confidence,
)

Edge cases & gotchas

• Category, not keyword. "cancer" is the Oncology category; "diabetes" is Disease. Check entity_types if unsure which fits.
• get_default_pii_model(lang) can return None for an unsupported language — fall back to a supported one and warn, do not silently use English on non-English text.
• search_models reads a committed manifest, so it only returns models that have been catalogued — combine with get_all_models() for the full registry.
• Match labels before committing. A model in the right category may still not emit the exact label you need; confirm via entity_types / canonical_labels.
• Licensing. All OpenMed registry models are permissively licensed; do not swap in models that bundle restricted terminologies (UMLS/SNOMED/CPT).

Standards & references

• OpenMed model org & cards: https://huggingface.co/OpenMed
• Canonical PII label taxonomy: openmed.CANONICAL_LABELS (see extracting-pii-entities).