药品不良事件信号检测

SKILL.md

中文导读

从真实不良事件报告中计算与解释信号指标，区分报告模式、统计关联和已确认的因果证据。

输入材料

• 药物、事件与查询条件
• 真实报告资料和分析环境

交付结果

• 不良事件信号分析
• 计算过程与证据说明

阅读与使用

本页保存指定版本的原始 Skill 和配套文件，可查看中文导读、原文与参考资料。获取完整包后，按原文准备所需工具与材料。执行需要 ToolUniverse SDK 或 MCP 服务，部分数据库需要独立访问权限。医疗资料供研究、资料整理和专业人员复核，试验资格、用药与诊疗结论需要核实当前来源。

来源仓库为 mims-harvard/ToolUniverse，采用 Apache-2.0 许可。Stars 表示仓库关注度，核验日期为 2026-10-15。分类表示适用的工作环节，实际执行条件以原文为准。

原始 Skill

COMPUTE, DON'T DESCRIBE

When analysis requires computation (statistics, data processing, scoring, enrichment), write and run Python code via Bash. Don't describe what you would do — execute it and report actual results. Use ToolUniverse tools to retrieve data, then Python (pandas, scipy, statsmodels, matplotlib) to analyze it.

Automated pipeline for detecting, quantifying, and contextualizing adverse drug event signals using FAERS disproportionality analysis, FDA label mining, mechanism-based prediction, and literature evidence. Produces a quantitative Safety Signal Score (0-100) for regulatory and clinical decision-making.

KEY PRINCIPLES:

1. Signal quantification first - Every adverse event must have PRR/ROR/IC with confidence intervals
2. Serious events priority - Deaths, hospitalizations, life-threatening events always analyzed first
3. Multi-source triangulation - FAERS + FDA labels + OpenTargets + DrugBank + literature
4. Context-aware assessment - Distinguish drug-specific vs class-wide vs confounding signals
5. Report-first approach - Create report file FIRST, update progressively
6. Evidence grading mandatory - T1 (regulatory/boxed warning) through T4 (computational)
7. English-first queries - Always use English drug names in tool calls, respond in user's language

REASONING STRATEGY — Start Here: Start with the signal: What adverse event was reported more than expected? (PRR >= 2.0, N >= 3, lower CI > 1.0 is the threshold). Then ask three questions in order:

1. Biologically plausible? Given the drug's mechanism of action and targets, does this adverse event make sense? An off-target kinase inhibitor causing cardiac events is plausible; a topical agent causing systemic toxicity needs more scrutiny. LOOK UP DON'T GUESS — use OpenTargets_get_drug_mechanisms_of_action_by_chemblId and drugbank_get_targets_by_drug_name_or_drugbank_id to check targets before asserting plausibility.
2. Timing consistent? Acute reactions (within hours/days) suggest immune or direct pharmacologic mechanism. Delayed reactions (weeks/months) suggest cumulative toxicity or idiosyncratic response. Check FAERS time-to-onset distribution.
3. Could confounders explain it? Patients taking this drug likely have the underlying disease — compare against background rate in that population, not the general population. Class-wide signals (appearing for all drugs in the class) suggest mechanism-based rather than molecule-specific toxicity.

Causality Assessment — Naranjo Algorithm Reasoning: When determining whether an adverse event is drug-caused (not just associated), apply these steps systematically. LOOK UP DON'T GUESS — search FAERS and FDA labels for each criterion:

1. Prior reports? Are there previous conclusive reports of this reaction? Check FDA label (FDA_get_adverse_reactions_by_drug_name) and literature (PubMed_search_articles). Yes = +1.
2. Temporal relationship? Did the AE appear after drug administration? Onset within expected pharmacokinetic window (1-5 half-lives) = +2. Use FAERS_stratify_by_demographics for time-to-onset data.
3. Dechallenge? Did the AE improve when the drug was stopped? Positive dechallenge = +1. Look for rechallenge/dechallenge case reports in literature.
4. Rechallenge? Did the AE reappear when the drug was restarted? Positive rechallenge = +2 (strongest single piece of evidence for causality).
5. Alternative causes? Could the underlying disease, concomitant drugs, or other factors explain the AE? Check drugbank_get_drug_interactions_by_drug_name_or_id for interacting drugs.
6. Dose-response? Did the reaction worsen with higher doses or improve with lower doses? Dose-dependent AEs suggest on-target toxicity.
7. Drug level confirmation? Was the drug detected in body fluids at toxic concentrations?

• Score: Definite (>=9), Probable (5-8), Possible (1-4), Doubtful (<=0).
• Even without individual patient data, you can estimate causality from aggregate FAERS signals + label evidence + mechanistic plausibility.

Reference files (in this directory):

• PHASE_DETAILS.md - Detailed tool calls, code examples, and output templates per phase
• REPORT_TEMPLATE.md - Full report template and completeness checklist
• TOOL_REFERENCE.md - Tool parameter reference and fallback chains
• QUICK_START.md - Quick examples and common drug names

---

When to Use

Apply when user asks:

• "What are the safety signals for [drug]?"
• "Detect adverse events for [drug]"
• "Is [drug] associated with [adverse event]?"
• "What are the FAERS signals for [drug]?"
• "Compare safety of [drug A] vs [drug B] for [adverse event]"
• "What are the serious adverse events for [drug]?"
• "Are there emerging safety signals for [drug]?"
• "Post-market surveillance report for [drug]"
• "Pharmacovigilance signal detection for [drug]"

Differentiation from tooluniverse-pharmacovigilance: This skill focuses specifically on signal detection and quantification using disproportionality analysis (PRR, ROR, IC) with statistical rigor, produces a quantitative Safety Signal Score (0-100), and performs comparative safety analysis across drug classes.

---

Workflow Overview

Phase 0: Input Parsing & Drug Disambiguation
  Parse drug name, resolve to ChEMBL ID, DrugBank ID
  Identify drug class, mechanism, and approved indications
    |
Phase 1: FAERS Adverse Event Profiling
  Top adverse events by frequency
  Seriousness and outcome distributions
  Demographics (age, sex, country)
    |
Phase 2: Disproportionality Analysis (Signal Detection)
  Calculate PRR, ROR, IC with 95% CI for each AE
  Apply signal detection criteria
  Classify signal strength (Strong/Moderate/Weak/None)
    |
Phase 3: FDA Label Safety Information
  Boxed warnings, contraindications
  Warnings and precautions, adverse reactions
  Drug interactions, special populations
    |
Phase 4: Mechanism-Based Adverse Event Context
  Target-based AE prediction (OpenTargets safety)
  Off-target effects, ADMET predictions
  Drug class effects comparison
    |
Phase 5: Comparative Safety Analysis
  Compare to drugs in same class
  Identify unique vs class-wide signals
  Head-to-head disproportionality comparison
    |
Phase 6: Drug-Drug Interactions & Risk Factors
  Known DDIs causing AEs
  Pharmacogenomic risk factors (PharmGKB)
  FDA PGx biomarkers
    |
Phase 7: Literature Evidence
  PubMed safety studies, case reports
  OpenAlex citation analysis
  Preprint emerging signals (EuropePMC)
    |
Phase 8: Risk Assessment & Safety Signal Score
  Calculate Safety Signal Score (0-100)
  Evidence grading (T1-T4) for each signal
  Clinical significance assessment
    |
Phase 9: Report Synthesis & Recommendations
  Monitoring recommendations
  Risk mitigation strategies
  Completeness checklist

---

Phase Summaries

Phase 0: Input Parsing & Drug Disambiguation

Resolve drug name to ChEMBL ID, DrugBank ID. Get mechanism of action, blackbox warning status, targets, and approved indications.

• Tools: OpenTargets_get_drug_chembId_by_generic_name, OpenTargets_get_drug_mechanisms_of_action_by_chemblId, OpenTargets_get_drug_blackbox_status_by_chembl_ID, drugbank_get_safety_by_drug_name_or_drugbank_id, drugbank_get_targets_by_drug_name_or_drugbank_id, OpenTargets_get_drug_indications_by_chemblId

Phase 1: FAERS Adverse Event Profiling

Query FAERS for top adverse events, seriousness distribution, outcomes, demographics, and death-related events. Filter serious events by type (death, hospitalization, life-threatening). Get MedDRA hierarchy rollup.

• Tools: FAERS_count_reactions_by_drug_event, FAERS_count_seriousness_by_drug_event, FAERS_count_outcomes_by_drug_event, FAERS_count_patient_age_distribution, FAERS_count_death_related_by_drug, FAERS_count_reportercountry_by_drug_event, FAERS_filter_serious_events, FAERS_rollup_meddra_hierarchy

Phase 2: Disproportionality Analysis (Signal Detection)

CRITICAL PHASE. For each top adverse event (at least 15-20), calculate PRR, ROR, IC with 95% CI. Classify signal strength. Stratify strong signals by demographics.

• Tools: FAERS_calculate_disproportionality, FAERS_stratify_by_demographics
• MedDRA term level note: FAERS_count_reactions_by_drug_event filters by MedDRA Lowest Level Term (reactionmeddraverse) while FAERS_calculate_disproportionality uses Preferred Terms. Case counts can differ dramatically — always use disproportionality analysis as the primary signal metric, not raw counts.
• Signal criteria: PRR >= 2.0 AND lower CI > 1.0 AND N >= 3
• Strength: Strong (PRR >= 5), Moderate (PRR 3-5), Weak (PRR 2-3)
• See PHASE_DETAILS.md for full signal classification table

Phase 3: FDA Label Safety Information

Extract boxed warnings, contraindications, warnings/precautions, adverse reactions, drug interactions, and special population info. Note: {error: {code: "NOT_FOUND"}} is normal when a section does not exist.

• Tools: FDA_get_boxed_warning_info_by_drug_name, FDA_get_contraindications_by_drug_name, FDA_get_warnings_by_drug_name, FDA_get_adverse_reactions_by_drug_name, FDA_get_drug_interactions_by_drug_name, FDA_get_pregnancy_or_breastfeeding_info_by_drug_name, FDA_get_geriatric_use_info_by_drug_name, FDA_get_pediatric_use_info_by_drug_name, FDA_get_pharmacogenomics_info_by_drug_name

Phase 4: Mechanism-Based Adverse Event Context

Get target safety profile, OpenTargets adverse events, ADMET toxicity predictions (if SMILES available), and drug warnings.

• Tools: OpenTargets_get_target_safety_profile_by_ensemblID, OpenTargets_get_drug_adverse_events_by_chemblId, ADMETAI_predict_toxicity, ADMETAI_predict_CYP_interactions, OpenTargets_get_drug_warnings_by_chemblId

Phase 5: Comparative Safety Analysis

Head-to-head comparison with class members using FAERS_compare_drugs. Aggregate class AEs. Identify class-wide vs drug-specific signals.

• Tools: FAERS_compare_drugs, FAERS_count_additive_adverse_reactions, FAERS_count_additive_seriousness_classification

Phase 6: Drug-Drug Interactions & Risk Factors

Extract DDIs from FDA label, DrugBank, and DailyMed. Query PharmGKB for pharmacogenomic risk factors and dosing guidelines. Check FDA PGx biomarkers.

• Tools: FDA_get_drug_interactions_by_drug_name, drugbank_get_drug_interactions_by_drug_name_or_id, DailyMed_parse_drug_interactions, PharmGKB_search_drugs, PharmGKB_get_drug_details, PharmGKB_get_dosing_guidelines, fda_pharmacogenomic_biomarkers

Phase 7: Literature Evidence

Search PubMed, OpenAlex, and EuropePMC for safety studies, case reports, and preprints.

• Tools: PubMed_search_articles, openalex_search_works, EuropePMC_search_articles

Phase 8: Risk Assessment & Safety Signal Score

Calculate Safety Signal Score (0-100) from four components: FAERS signal strength (0-35), serious AEs (0-30), FDA label warnings (0-25), literature evidence (0-10). Grade each signal T1-T4. See PHASE_DETAILS.md for scoring rubric.

Phase 9: Report Synthesis

Generate comprehensive markdown report with executive summary, all phase outputs, monitoring recommendations, risk mitigation strategies, patient counseling points, and completeness checklist. See REPORT_TEMPLATE.md for full template.

---

Edge Cases

• No FAERS reports: Skip Phases 1-2; rely on FDA label, mechanism predictions, literature
• Generic vs Brand name: Try both in FAERS; use OpenTargets_get_drug_chembId_by_generic_name to resolve
• Drug combinations: Use FAERS_count_additive_adverse_reactions for aggregate class analysis
• Confounding by indication: Compare AE profile to the disease being treated; note limitation in report
• Drugs with boxed warnings: Score component automatically 25/25 for label warnings; prioritize boxed warning events