Drug Interaction Evidence Scoring Prompt

This prompt is designed for clinical pharmacology and medication safety tools that must score and rank drug interaction evidence before surfacing alerts to clinicians or patients. It takes a drug pair, retrieved evidence passages, and patient-specific context, then outputs a structured interaction assessment with mechanism plausibility, study quality, severity grading, and patient-specific risk modifiers. Use this prompt when your application needs to distinguish between well-documented, clinically observed interactions and purely theoretical ones that lack supporting case reports or clinical studies.

This prompt belongs in the evidence-ranking stage of a RAG pipeline, after retrieval but before alert generation. It assumes you have already retrieved relevant passages from drug interaction databases, product labels, and biomedical literature. The prompt does not perform retrieval itself and should not be used as a standalone clinical decision maker without human review for high-severity interactions. The ideal user is an AI engineer or clinical informaticist building a medication safety module who needs structured, scored evidence that downstream alert logic can consume programmatically.

Do not use this prompt when you lack patient-specific context such as renal function, age, or concurrent medications, as the risk modifier scoring will be unreliable. Do not use it for drug-allergy interactions, food-drug interactions, or supplement interactions unless you have adapted the evidence hierarchy and severity grading to those domains. Avoid using this prompt as the sole gate for blocking medication orders; always route high-severity or uncertain outputs to a human reviewer, and log every scored interaction for audit and pharmacovigilance review.

This prompt template is designed to be pasted directly into your system prompt or user message template. It instructs the model to act as a clinical pharmacology reviewer, scoring each potential drug interaction across multiple evidence dimensions and producing a structured, ranked alert. The template uses square-bracket placeholders that your application must replace with live data from your retrieval pipeline, patient context, and drug knowledge base before sending the request to the model.

textCopy
You are a clinical pharmacology evidence reviewer. Your task is to score the evidence for a potential drug interaction between [DRUG_A] and [DRUG_B] in the context of a specific patient.

## PATIENT CONTEXT
- Age: [PATIENT_AGE]
- Sex: [PATIENT_SEX]
- Weight: [PATIENT_WEIGHT]
- Renal Function (CrCl): [PATIENT_CRCL]
- Hepatic Function (Child-Pugh): [PATIENT_LIVER]
- Current Medications: [PATIENT_MEDS]
- Relevant Diagnoses: [PATIENT_DX]
- Pharmacogenomic Notes: [PATIENT_PGX]

## RETRIEVED EVIDENCE PASSAGES
For each passage below, consider its source, study design, and recency.

[EVIDENCE_PASSAGES]

## OUTPUT SCHEMA
Return a single JSON object with the following structure:
{
  "interaction_pair": "[DRUG_A] + [DRUG_B]",
  "overall_risk_score": "high" | "moderate" | "low" | "theoretical_only" | "no_interaction_expected",
  "mechanism_plausibility": {
    "score": "established" | "plausible" | "theoretical" | "unlikely",
    "mechanism_description": "string",
    "supporting_references": ["citation_keys"]
  },
  "evidence_quality": {
    "score": "high" | "moderate" | "low" | "insufficient",
    "best_study_design": "RCT" | "observational_cohort" | "case_control" | "case_series" | "case_report" | "in_vitro" | "in_silico" | "expert_opinion",
    "total_patients_studied": number_or_null,
    "limitations": ["string"]
  },
  "severity_assessment": {
    "score": "contraindicated" | "major" | "moderate" | "minor" | "none",
    "potential_outcome": "string",
    "onset_timeframe": "string",
    "reversibility": "reversible" | "partially_reversible" | "irreversible" | "unknown"
  },
  "patient_specific_risk": {
    "score": "high" | "moderate" | "low" | "not_applicable",
    "risk_factors_present": ["string"],
    "mitigating_factors": ["string"],
    "recommended_monitoring": ["string"]
  },
  "clinical_observability": "clinically_observed" | "theoretical_only" | "unknown",
  "ranked_alerts": [
    {
      "rank": 1,
      "alert_type": "contraindication" | "dose_adjustment" | "monitoring" | "caution" | "informational",
      "description": "string",
      "urgency": "immediate_action" | "review_before_next_dose" | "routine_review",
      "supporting_citations": ["citation_keys"]
    }
  ],
  "citations": [
    {
      "citation_key": "string",
      "source": "string",
      "study_design": "string",
      "year": number,
      "sample_size": number_or_null,
      "summary": "string"
    }
  ],
  "confidence_statement": "string",
  "requires_human_review": true | false
}

## CONSTRAINTS
- Distinguish between clinically observed interactions and theoretical interactions based only on mechanism speculation.
- If the best evidence is a case report or in-vitro study, reflect that uncertainty in the evidence_quality score.
- Do not invent citations. Only use citation_keys from the provided evidence passages.
- If evidence is insufficient to score, set overall_risk_score to "theoretical_only" and explain why in the confidence_statement.
- Flag any patient-specific risk factors that amplify or mitigate the interaction.
- If the interaction is rated "contraindicated" or "major," set requires_human_review to true.

To adapt this template, replace each square-bracket placeholder with data from your application context. [DRUG_A] and [DRUG_B] should be populated from the user's medication list or order entry. [PATIENT_*] fields must be drawn from the patient's electronic health record or session context. [EVIDENCE_PASSAGES] should be the raw text of retrieved passages from your drug interaction knowledge base, each annotated with a unique citation_key that the model can reference. The output schema is designed to be parsed by a downstream validation function; ensure your application validates the JSON structure, checks that all citation_keys in the output exist in the input passages, and escalates any response where requires_human_review is true but was not flagged for clinician attention.

This prompt is designed to operate inside a retrieval-augmented generation (RAG) pipeline that first fetches candidate evidence from drug databases (e.g., DailyMed, DrugBank, PubMed), then passes the top-N passages to the model for scoring. The harness must supply the prompt with structured inputs: a patient medication list, relevant clinical context (age, renal/hepatic function, pharmacogenomic variants), and the raw text of retrieved interaction claims. The model should never be asked to recall interaction facts from its training data alone—all evidence must be provided in the [EVIDENCE_PASSAGES] placeholder to ensure grounding and auditability.

Validation and output enforcement are critical because downstream medication safety alerts can cause alert fatigue or missed interactions. The harness should parse the model's JSON output and validate that every scored interaction includes a mechanism_plausibility field, a study_quality tier (e.g., clinical_trial, case_report, theoretical, in_vitro), a severity_rating, and at least one supporting_citation with a passage index linking back to the provided evidence. If the model outputs an interaction not present in the input evidence, the harness must discard it and log a hallucination event. Use a JSON Schema validator post-generation; on failure, retry once with a repair prompt that includes the schema and the specific validation error. If the retry also fails, route the case to a human pharmacist review queue with the original evidence and the failed output attached.

Model choice and latency matter in clinical workflows. For real-time medication order entry screens, use a fast instruction-tuned model (e.g., GPT-4o, Claude 3.5 Sonnet) with a strict timeout of 3–5 seconds. For batch retrospective drug utilization review, a larger reasoning model (e.g., o1, Claude Opus) can be used with longer timeouts. Always set temperature=0 to maximize output determinism for safety-critical scoring. The harness should log every prompt invocation with a trace ID, the model version, the input evidence set, the scored output, and any validation failures. This audit trail is essential for pharmacovigilance reporting and for defending the system's recommendations during regulatory review.

Human-in-the-loop gates are required for high-severity findings. If the model scores any interaction as severity: contraindicated or severity: major with study_quality: theoretical, the harness must not surface the alert directly to a prescriber. Instead, route it to a clinical pharmacist for confirmation, along with the scored evidence and a pre-drafted alert language that the pharmacist can approve, edit, or suppress. For moderate and minor interactions with strong clinical evidence, the harness may surface the alert directly but must include the supporting citations and a link to the original source passages so the clinician can verify the basis. Never allow the model to generate patient-facing advice; the output is always a structured alert payload for clinician-facing systems.

Eval and monitoring must go beyond schema validation. Build a golden dataset of 50–100 known drug interaction pairs with expert-annotated severity, mechanism, and evidence quality labels. Run this eval suite on every prompt or model version change. Key failure modes to monitor in production: (1) the model scoring a theoretical interaction as clinically observed, (2) missing a well-documented interaction because the retrieval step failed to supply the evidence, (3) hallucinating a citation that does not exist in the input passages, and (4) over-alerting on low-severity interactions and contributing to alert fatigue. Track precision and recall against the golden set, and set a minimum recall threshold of 0.95 for contraindicated interactions before any prompt update can ship.