pleio-hpo

Offline extraction of Human Phenotype Ontology (HPO) codes from free-text clinical notes. Runs fully offline on CPU — no data leaves the machine, which makes it usable with PHI.

A three-stage pipeline: lexical matching (Aho-Corasick over HPO labels + synonyms, with morphology-aware and order-free token-set matching for plural and reordered phrasings) → biomedical-embedding nearest-neighbour (SapBERT) for paraphrases → a PubMedBERT cross-encoder validator that filters the fuzzy candidates.

Install

pip install pleio-hpo            # Python >=3.10
python -m spacy download en_core_web_sm
pleio-hpo download              # fetch the embedding index + validator (~570 MB, one time)

The large model assets are fetched on first use, not bundled in the wheel. Runtime needs roughly 1.2 GB RAM (SapBERT + the PubMedBERT validator on CPU).

Quickstart — CLI

$ pleio-hpo "The patient has macrocephaly and hypotonia."
HP:0000256  Macrocephaly                             1.00
HP:0001252  Hypotonia                                1.00

# TSV / JSON, from a file or stdin, with context control:
pleio-hpo "Bilateral hearing loss; no microcephaly." --format tsv --include-negated
pleio-hpo --file note.txt --format json --output codes.json
echo "Short stature and seizures." | pleio-hpo --format tsv

pleio-hpo info shows the version, pinned HPO release, and asset paths.

Quickstart — Python

from pleio_hpo import Annotator

annotator = Annotator()  # construct once, reuse (models load lazily on first call)
result = annotator.annotate("Global developmental delay and seizures.")

for code in result.codes:
    print(code.hpo_id, code.label, code.score, code.score_source)

print(result.to_json())

code.score is an opaque per-source strength signal (1.0 for an exact lexical match, cosine similarity for an embedding match) — not a calibrated probability. See examples/ for batch processing, context filtering, and threshold tuning.

Evaluation

On GSC+ (228 PubMed abstracts, community human gold), under a uniform document-level micro-F1 protocol (HPO v2026-02-16):

¹ The Doc2HPO numbers use its public web API (text leaves the machine); a local, PHI-safe install needs a UMLS license.

pleio-hpo tops the full benchmark, ahead of PhenoTagger by 0.055 F1 (p<0.001). Part of that margin is inheritance/onset coverage the benchmark scores; on phenotype-only recognition the two are level (0.651 vs 0.647, p=0.67). Across three human-gold corpora — GSC+, BC8, and 112 out-of-distribution genetics case reports — pleio-hpo and PhenoTagger are statistically tied (the nominal leader varies by corpus), and the strongest frontier-LLM extractor performs comparably. The corpora are abstracts, exam observations, and case reports — not free-text clinical notes.

Condensed results, figures, ablation, and caveats: docs/RESULTS.md; per-tool comparison: COMPARISON.md.

HPO version

The library runtime tracks the latest HPO release (the index + validator are built against the current hp.obo). The reported evaluation is pinned to v2026-02-16 for comparability, so a user's runtime ontology may be newer than the evaluation snapshot (docs/RESULTS.md).

Scope and disclaimer

This tool is decision-support, not clinical decision-making. It is not FDA-cleared and not for autonomous use in patient care; inferred codes may be incorrect and should be verified by a qualified user. English-only. HPO is a Western/English-language ontology — apply cautiously to other populations.

License & citation

Apache-2.0 (see LICENSE and NOTICE). If you use this work, please cite it — see CITATION.cff.