qsong-github/drugclaw/phee

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.

API

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)