技能 数据科学 不良事件报告结构化(E2B(R3))

不良事件报告结构化(E2B(R3))

v20260803
reporting-adverse-events
该技能用于将非结构化的不良事件描述文本,转化为符合ICH E2B(R3)标准的结构化数据。它能够系统性地提取和填充关键字段,包括可疑药物、反应(通过MedDRA编码)、严重性等级和结局等,是构建个体病例安全报告(ICSR)和进行药物警戒申报的专业工具。
获取技能
360 次下载
概览

Reporting adverse events into FAERS / ICH E2B(R3)

A pharmacovigilance case starts as free-text narrative ("68 yo on warfarin developed GI bleed, hospitalized"). To make it reportable you must structure it into the ICH E2B(R3) data elements that the FDA's FAERS (and EMA's EudraVigilance) expect: a suspect drug, one or more reactions coded to MedDRA Preferred Terms, seriousness criteria, and a reaction outcome.

OpenMed extracts the drug and condition spans on-device; this skill turns those spans plus the narrative into the E2B(R3) skeleton. The reaction coding step needs MedDRA, which is licensed by the MSSO and user-supplied — it is never bundled with OpenMed and must be loaded from the user's own subscription.

When to use

  • A narrative names a drug and an adverse reaction and you need an ICSR (Individual Case Safety Report) shell with the right E2B(R3) fields.
  • You must classify seriousness (E2B sections C.1.7 / E.i.3) — death, life-threatening, hospitalization/prolongation, disability, congenital anomaly, or "other medically important condition".
  • You need to characterize each drug as suspect / concomitant / interacting (the drugcharacterization axis FAERS uses).
  • You are pre-filling a 3500A / FAERS electronic submission or staging cases for a safety database.

This skill produces a structured draft for human safety review — it does not file reports or perform causality assessment autonomously.

Quick start

import openmed

narrative = (
    "68-year-old patient on warfarin 5 mg daily developed a gastrointestinal "
    "hemorrhage and was hospitalized. Warfarin was discontinued; the patient "
    "recovered."
)

# 1) Extract drug spans (Pharmaceutical category) on-device.
drugs = openmed.analyze_text(
    narrative,
    model_name="pharma_detection_superclinical",
    output_format="dict",
)["entities"]

# 2) Extract condition / reaction spans (Disease category).
conditions = openmed.analyze_text(
    narrative,
    model_name="disease_detection_superclinical",
    output_format="dict",
)["entities"]

# 3) Assemble an E2B(R3)-shaped ICSR skeleton (reaction PTs filled later via MedDRA).
icsr = {
    "patient": {"age": None, "sex": None},          # from de-identified demographics
    "drugs": [
        {
            "name": e["text"],
            "drugcharacterization": 1,              # 1=suspect 2=concomitant 3=interacting
            "action": None,                          # e.g. drug withdrawn / dose reduced
        }
        for e in drugs
    ],
    "reactions": [
        {
            "verbatim": e["text"],                   # narrative term, pre-MedDRA
            "meddra_pt": None,                       # coded with user's MedDRA dict
            "outcome": None,                         # E2B reaction outcome code
        }
        for e in conditions
    ],
    "seriousness": {
        "serious": None, "death": False, "lifeThreatening": False,
        "hospitalization": True, "disability": False, "congenitalAnomaly": False,
        "otherMedicallyImportant": False,
    },
}

E2B(R3) seriousness and outcome value sets

Seriousness is a set of boolean criteria (E2B E.i.3.2). A case is serious if any criterion is true:

Criterion E2B element FAERS field
Death E.i.3.2a seriousnessdeath
Life-threatening E.i.3.2b seriousnesslifethreatening
Hospitalization / prolonged E.i.3.2c seriousnesshospitalization
Disability / incapacity E.i.3.2d seriousnessdisabling
Congenital anomaly E.i.3.2e seriousnesscongenitalanomali
Other medically important E.i.3.2f seriousnessother

Reaction outcome (E2B E.i.7) is a coded value: 1 recovered/resolved, 2 recovering/resolving, 3 not recovered/not resolved, 4 recovered with sequelae, 5 fatal, 6 unknown.

Drug characterization (E2B G.k.1): 1 suspect, 2 concomitant, 3 interacting.

Workflow

  1. De-identify first. Run openmed.deidentify(narrative, policy=...) and work from result.deidentified_text. Patient name, MRN, and dates must be removed/shifted before the case leaves your environment.
  2. Extract drugs and reactions with the two analyze_text calls above. Keep each entity's start/end offsets for traceability.
  3. Characterize each drug as suspect (1), concomitant (2), or interacting (3). The drug that temporally precedes the reaction and was acted upon (withdrawn/reduced) is usually the suspect.
  4. Code reactions to MedDRA. Map each verbatim reaction term to a MedDRA Preferred Term (PT) and its System Organ Class using the user's licensed MedDRA dictionary (see "Edge cases"). Never invent PTs.
  5. Determine seriousness. Scan the narrative for the six criteria; set serious=True if any is met. "Hospitalized", "admitted", "ICU" → C.1.7c.
  6. Assign reaction outcome from the value set above.
  7. Hand the structured draft to a qualified safety reviewer for causality (e.g. WHO-UMC or Naranjo), expectedness, and final submission.

Hand-off to / from OpenMed

OpenMed's analyze_text returns a dict; result["entities"] is a list whose items carry text, label, confidence, start, end. Consume them:

  • From extracting-clinical-entities: Pharmaceutical entities → icsr["drugs"]; Disease entities → icsr["reactions"]. Keep offsets so each E2B field is traceable to the source span.
  • From normalizing-rxnorm: optionally attach an RxCUI to each suspect drug for product identification (E2B G.k.2.2) before coding.
  • De-identify with deidentifying-clinical-text (openmed.deidentify) before the case is exported or transmitted to any safety database.
  • To detecting-pv-signals: aggregated, coded cases feed disproportionality analysis. To querying-openfda-labels: confirm the reaction is/ isn't a labeled event (expectedness).

Edge cases & gotchas

  • MedDRA is licensed — never bundle it. MedDRA is distributed by the MSSO under subscription; OpenMed ships none of it. Load PTs/LLTs from the user's own MedDRA release (the version is itself a reportable field, E2B C.1.x). Verbatim reaction text stays in the case until a coder maps it.
  • One reaction term ≠ one PT. "GI bleed" maps to the PT Gastrointestinal haemorrhage; keep the verbatim term alongside the coded PT for the audit trail. Multi-word reactions span several OpenMed tokens — reassemble by offset.
  • Suspect vs concomitant matters. Disproportionality and labeling decisions hinge on drugcharacterization. Do not default every drug to suspect.
  • Seriousness is OR, not a severity scale. A mild rash that caused hospitalization is serious; a severe headache that resolved at home may not be. Classify by the six regulatory criteria, not by clinical severity words.
  • Causality is out of scope here. This skill structures the case; it does not assert the drug caused the event. Leave causality to the reviewer.
  • Local-first. NER and de-identification run on-device. Only de-identified, structured case data should reach an external safety database, and only under the appropriate regulatory agreement.

Standards & references

信息
Category 数据科学
Name reporting-adverse-events
版本 v20260803
大小 8.69KB
更新时间 2026-08-04
语言