phee-query

Query the PHEE pharmacovigilance event extraction dataset. Use whenever the user asks about annotated adverse drug events, pharmacovigilance case reports, drug–effect associations from medical literat

SKILL.md

GitHub

PHEE Query Skill

Search PHEE annotated pharmacovigilance events by drug name, adverse effect, or free text. Substring matching is case-insensitive across all event fields.

Input Pattern Match Logic
drug name (e.g. phenytoin) substring on drug, treatment, combination_drug, and text
adverse effect (e.g. hepatotoxicity) substring on effect, treatment_disorder, and text
any other text substring on all event fields and full sentence

API

Function Input Returns
load_phee(data_dir) directory path list[dict] (cached)
search(records, entity) single entity string list[dict]
search_batch(records, entities) list of entity strings dict[str, list[dict]]
summarize(hits, entity) hit list + label compact text (sentence count, ADE/PTE counts, top drugs & effects)
to_json(hits) hit list list[dict]

Record Schema

Each record returned by search:

{
  "id": "8908396_3",
  "text": "A 52-year-old Black woman on phenytoin therapy ...",
  "split": "train",
  "events": [
    {
      "event_type": "Adverse_event | Potential_therapeutic_event",
      "trigger": ["developed"],
      "subject": ["A 52-year-old Black woman ..."],
      "subject_age": ["52-year-old"],
      "subject_gender": ["woman"],
      "subject_race": ["Black"],
      "subject_population": [],
      "subject_disorder": ["post-traumatic epilepsy"],
      "treatment": ["phenytoin therapy ..."],
      "drug": ["phenytoin"],
      "dosage": [],
      "freq": [],
      "route": [],
      "duration": [],
      "treatment_disorder": ["post-traumatic epilepsy"],
      "combination_drug": [],
      "effect": ["transient hemiparesis"]
    }
  ]
}

Usage

See if __name__ == "__main__" block in 38_PHEE.py for runnable examples covering: single drug search, single effect search, batch search, and JSON output.

Data

  • Source: PHEE dataset (Zenodo record 7689970), EMNLP 2022
  • Files: train.json (2,900 sentences), dev.json (960), test.json (~960)
  • Path: DATA_DIR variable in 38_PHEE.py (resources_metadata/drug_nlp/PHEE/data/json)
  • Event types: Adverse_event, Potential_therapeutic_event
  • Main arguments: Subject, Treatment, Effect
  • Sub-arguments: Subject (age, gender, race, population, disorder); Treatment (drug, dosage, freq, route, duration, disorder, combination.drug)