Drug-Drug Interaction (DDI) Evidence-Based Assessment

Overview

This skill implements a Retrieval-Augmented Generation (RAG) approach to drug-drug interaction assessment, modeled after the Micromedex Drug-Reax System methodology. Rather than relying on LLM training data alone -- which may be outdated, incomplete, or hallucinated -- every claim in the output is grounded in real-time literature retrieval from peer-reviewed sources with full citation traceability. The structured report mirrors Micromedex Drug-Reax classification grades and is designed to complement (not replace) certified drug information systems and clinical pharmacist expertise.

Trigger

• /drug-drug DrugA DrugB
• Any query mentioning two drug names and asking about interaction, safety, or concomitant use
• Keywords: DDI, interaction, contraindication, concomitant, 交互作用, 併用

Workflow (execute strictly in order)

Step 0: Drug Name Resolution

Extract two drug names from user input. If the user provides brand names, resolve to INN/generic names. If only one drug is provided or names are ambiguous, ask the user to clarify.

Step 1: Literature Search

Execute at least 5 distinct searches to ensure adequate coverage:

1. PubMed Search (via web_search)

- site:pubmed.ncbi.nlm.nih.gov "DrugA" "DrugB" interaction - site:pubmed.ncbi.nlm.nih.gov "DrugA" "DrugB" pharmacokinetic - site:pubmed.ncbi.nlm.nih.gov "DrugA" "DrugB" CYP450

1. CrossRef Search (via web_fetch)

- https://api.crossref.org/works?query=DrugA+DrugB+drug+interaction&rows=10&sort=relevance

1. General WebSearch

- DrugA DrugB drug interaction clinical significance - DrugA DrugB interaction mechanism CYP enzyme - DrugA DrugB interaction case report adverse event

1. FDA Label / DailyMed

- site:dailymed.nlm.nih.gov DrugA interaction - DrugA DrugB FDA drug interaction warning

1. PubMed MCP Tool (if PubMed MCP server is connected)

- Use PubMed:searcharticles with "DrugA" AND "DrugB" AND "drug interaction" - Use PubMed:getarticle_metadata for key articles

Step 2: Evidence Appraisal

Read references/evidence-grading.md for complete grading criteria.

For each retrieved article, extract and document:

• Study design (RCT / cohort / case-control / case report / in vitro / review)
• Sample size
• Key findings (AUC fold-change, Cmax change, clinical events)
• Interaction mechanism described
• Credibility assessment

Traceability requirement: Every factual claim in the final report must be attributable to a specific source with PMID or DOI. If a claim is inferred from pharmacologic reasoning rather than direct evidence, it must be explicitly labeled as such (e.g., "based on known CYP3A4 inhibition profile" rather than stated as established fact).

Uncertainty signaling: When evidence is limited or conflicting, the report must explicitly state the level of uncertainty. Use phrases such as "limited evidence suggests," "based on case reports only," or "no direct clinical studies available; assessment based on pharmacologic reasoning." Never present uncertain inferences with the same confidence as RCT-level evidence.

Step 3: Structured Classification

Classify using the Micromedex Drug-Reax framework. See references/evidence-grading.md for detailed decision trees.

3a. Severity

3b. Documentation

3c. Onset

3d. Mechanism

Classify the interaction mechanism into:

Pharmacokinetic (PK):

• CYP450 enzyme inhibition (specify isoform: CYP3A4, CYP2D6, CYP2C19, CYP2C9, CYP1A2, etc.)
• CYP450 enzyme induction
• P-glycoprotein (P-gp) / transporter-related
• Protein binding displacement
• Renal tubular secretion competition
• Absorption-level (pH alteration, chelation, etc.)

Pharmacodynamic (PD):

• Additive effect (e.g., QTc prolongation, bleeding risk, CNS depression)
• Synergistic effect
• Antagonistic effect

Step 4: Report Generation

Produce the final report in the following format:

# Drug-Drug Interaction Assessment Report

## Drug Pair

- **Drug A:** [Generic Name] ([Brand Names])
- **Drug B:** [Generic Name] ([Brand Names])

## Structured Classification

| Parameter       | Grade          | Note              |
|-----------------|----------------|-------------------|
| Severity        | [Grade]        | [note]            |
| Onset           | [Grade]        | [note]            |
| Documentation   | [Grade]        | [note]            |

## Interaction Effect

[Description of the clinical effect of the interaction — what happens when these drugs are used together]

## Clinical Management

[Specific clinical management recommendations:]

- Whether to avoid concomitant use
- Dose adjustments required
- Monitoring parameters
- Alternative drug suggestions

## Probable Mechanism

[Pharmacological mechanism — PK and/or PD, including specific enzymes, transporters, or receptors involved. Include PK data (AUC/Cmax changes) if available from studies.]

## Evidence Sources

[Key references: Author, Journal, Year, PMID/DOI]

## Disclaimer

> **This report is generated by AI using a retrieval-augmented generation (RAG) approach and
> is intended as a clinical decision support aid only.** It does not constitute medical advice
> and cannot replace certified drug information systems (Micromedex, Lexicomp, Clinical
> Pharmacology) or the judgment of qualified healthcare professionals. AI-generated content
> carries inherent limitations including potential for incomplete literature retrieval,
> misinterpretation of source data, and inability to account for individual patient factors.
> All clinical decisions must be made by licensed practitioners with access to complete
> patient information, institutional formulary policies, and current prescribing guidelines.
> When in doubt, consult a clinical pharmacist or drug information specialist.

Key Principles

1. Retrieval before generation: Never rely on LLM training data alone. Always perform real-time literature search before generating any assessment. Training data may be outdated, incomplete, or reflect unverified sources (social media, blogs). Real-time retrieval from peer-reviewed databases ensures currency and reliability.
2. Multi-source cross-validation: Use at least PubMed + CrossRef + WebSearch (three sources). Cross-validate findings across sources to reduce the risk of hallucination or bias from any single retrieval.
3. Full citation traceability: Every factual claim must cite its source (PMID, DOI, or URL). This enables clinicians to fact-check AI-generated answers against the original sources -- a core requirement for explainable AI (XAI) in clinical settings.
4. Conservative grading: When evidence is insufficient, err on the side of higher severity and lower documentation grade (err on the side of caution). Patient safety takes precedence over precision.
5. Mechanism is king: Even without clinical studies, if the pharmacologic mechanism is clear (e.g., known CYP inhibitor + known CYP substrate), document it and assign at least Fair documentation.
6. Transparency of uncertainty: The report must clearly distinguish between evidence-supported facts, pharmacologic inferences, and areas of genuine uncertainty. Never present inferred information with the same confidence as controlled-study evidence.
7. Context matters: Note dose-dependent or population-specific differences when evidence supports them (e.g., high-dose vs. low-dose regimens may carry different risk profiles).
8. Complement, not replace: This tool supports clinical decision-making but does not substitute for professional judgment. The output is designed to be reviewed by a clinician, not acted upon autonomously.
9. Language: The report body defaults to English. If the user writes in another language, respond in that language but keep drug names, grade labels, and pharmacological terms in English.

Reference Files

• references/evidence-grading.md: Complete evidence grading criteria, decision trees, CYP450 quick reference, and common PD interaction patterns