feat: add cardiology domain to NER catalog and zero-shot label maps

[pardeep-singh]

Implements #317 (OM-152): adds a cardiology domain to the zero-shot default label map plus cardiology keyword routing, so cardiology notes (ECG findings, ejection fraction, cardiac procedures/devices) stop falling back to generic clinical labels.

Per the discussion on the issue, this follows 1(a) (routing helper, get_model_suggestions behavior unchanged) and 2(a) (display labels consistent with existing domains; no new clinical canonical-label system).

Changes

• openmed/zero_shot/data/label_maps/defaults.json — new cardiology domain using the existing letters-only display-label style: CardiacFinding, ECGFinding, EjectionFraction, CardiacProcedure, CardiacDevice, Anatomy.
• openmed/core/model_registry.py
  • Extracted the keyword->category matching out of get_model_suggestions into a _match_categories(text) helper that returns (category, reason) pairs, so cardiology text routes to the Cardiology category independently of whether a Cardiology model is registered.
  • get_model_suggestions() user-facing behavior is unchanged: with no Cardiology model registered, cardiology text still falls back to general medical suggestions rather than surfacing unrelated results.
  • Added _CATEGORY_ENTITY_TYPES["Cardiology"] as forward registry metadata for future Cardiology models (none exists today).
• tests/unit/ner/test_label_map_consistency.py — tests for available_domains() / get_default_labels("cardiology"), non-empty labels, no duplicates, a label-style-consistency check across all domains, _match_categories routing for the echocardiogram example, and that get_model_suggestions behavior is unchanged.

openmed/core/labels.py is intentionally untouched — per the issue discussion, the strict CANONICAL_LABELS taxonomy there is the PII/PHI policy taxonomy, not a clinical zero-shot taxonomy, and no broad clinical canonical-label system is introduced in this issue.

Acceptance criteria

• get_default_labels("cardiology") returns a non-empty label set and available_domains() includes cardiology.
• Cardiology text (e.g. Echocardiogram shows reduced ejection fraction of 35%) routes to the Cardiology category via _match_categories.
• The cardiology domain reuses the existing display-label style (asserted by a style-consistency test across all domains).
• _CATEGORY_ENTITY_TYPES["Cardiology"] added as forward metadata; CATEGORIES still has no Cardiology key (no model exists yet).

Testing

• New cardiology tests pass; full unit suite passes (.venv/bin/python -m pytest tests/ -q -> 1619 passed, 12 skipped).
• The only 2 failures in the full run are pre-existing, network-gated integration tests (tests/integration/test_sentence_detection_real.py, which download dslim/bert-base-NER); they fail identically on master offline and are unrelated to this change.
• ruff check and ruff format clean on the changed files.

Out of scope

• Training a cardiology model (routing/labels only).
• ECG signal processing (text only).
• A clinical canonical-label system.

Closes #317

[maziyarpanahi]

Hi @pardeep-singh, thanks for the careful implementation and for keeping get_model_suggestions() behavior unchanged. CI is green and the scope matches the decision in #317.

I’ll review this one first since #552 is stacked on it. I’ll focus on the _match_categories helper, label-map consistency tests, and making sure Cardiology stays registry metadata rather than implying a shipped model.

[maziyarpanahi]

Thank you @pardeep-singh. I reviewed this against #317 / OM-152 and the follow-up clarification on the issue.

The implementation matches the intended shape: cardiology is added to the zero-shot defaults, _match_categories() exposes Cardiology keyword routing independently of registered models, get_model_suggestions() still falls back to the existing general suggestions because no Cardiology model is registered, and _CATEGORY_ENTITY_TYPES["Cardiology"] is only forward registry metadata.

I did not need to change the branch. Local verification on the PR checkout:

• PYTHONPATH=/tmp/openmed-pr-551.IO6dJS /Users/maziyar/Developer/openmed/.venv/bin/python -m pytest tests/unit/ner/test_label_map_consistency.py -q -> 54 passed
• /Users/maziyar/Developer/openmed/.venv/bin/ruff check openmed/core/model_registry.py tests/unit/ner/test_label_map_consistency.py -> passed
• /Users/maziyar/Developer/openmed/.venv/bin/ruff format --check openmed/core/model_registry.py tests/unit/ner/test_label_map_consistency.py -> passed

Hosted CI is green and the PR is mergeable. I also copied the labels from #317 onto the PR.

[pardeep-singh]

@maziyarpanahi are we good with this PR? if so, can we please merge this so that I can rebaae the other one?

[maziyarpanahi]

Approved. I did the final pass against #317 / OM-152 and the issue clarification.

This is ready to merge: the scope matches the agreed helper-based routing approach, labels are copied from the issue, hosted CI is green, and GitHub reports the branch as clean/mergeable.

Thanks again @pardeep-singh.

[maziyarpanahi]

Thanks @pardeep-singh this is merged now