Clinical Decision Support Evidence Ranking Prompt

This prompt is for clinical decision support (CDS) and diagnostic support product teams that need to rank retrieved evidence by patient-specific match, outcome improvement evidence, and guideline alignment. Use it when your system retrieves multiple studies, guidelines, or clinical references and must produce a personalized, ranked evidence summary for a clinician. The ideal user is an engineering lead or AI architect building a retrieval-augmented generation (RAG) pipeline where the retrieval step returns a flat list of candidate passages, and the downstream clinician-facing UI requires a relevance-ordered, patient-contextualized evidence table before any answer generation occurs.

This prompt assumes you have already retrieved candidate evidence passages and have a structured patient profile containing demographics, diagnoses, medications, lab results, and relevant clinical history. It does not perform retrieval, diagnose conditions, or replace clinical judgment. The output is a ranked evidence table with explicit reasoning for each ranking decision, designed to be reviewed by a qualified healthcare professional. Do not use this prompt when the evidence set is empty, when the patient profile is incomplete for the clinical question at hand, or when the downstream task is generating a final clinical recommendation rather than an intermediate evidence ranking. For recommendation generation, use a separate synthesis prompt that consumes this ranked output.

Before deploying, build an evaluation harness that tests for confounding by indication (where treatment assignment correlates with outcome risk independently of treatment effect) and spectrum bias (where study populations differ systematically from your target patient population). These failure modes are common in observational evidence and can produce misleading rankings. Your eval dataset should include cases where the correct ranking is known to clinical reviewers, and your metrics should measure both rank correlation and clinically significant misrankings—a top-3 swap that changes a treatment decision matters more than a swap among low-relevance references. Always route the final ranked output through a human review step in any production clinical workflow, and log ranking decisions with patient de-identified context for audit and continuous improvement.

This template is designed to be pasted directly into your model's system or user message. It expects structured inputs for the clinical question, patient context, and retrieved evidence passages. Replace every square-bracket placeholder with real data before sending the request to the model. The prompt enforces a strict JSON output schema to ensure the ranked evidence list can be parsed and validated by downstream application code.

textCopy
You are a clinical evidence ranking assistant for a Clinical Decision Support (CDS) system. Your task is to rank retrieved evidence passages by their relevance, quality, and applicability to a specific patient scenario. You must produce a structured, ranked list with explicit reasoning for each ranking decision.

## INPUT

### Clinical Question
[CLINICAL_QUESTION]

### Patient Context
- Demographics: [PATIENT_DEMOGRAPHICS]
- Relevant History: [PATIENT_HISTORY]
- Current Presentation: [CURRENT_PRESENTATION]
- Active Medications: [ACTIVE_MEDICATIONS]
- Known Allergies: [KNOWN_ALLERGIES]
- Lab/Imaging Results: [LAB_IMAGING_RESULTS]

### Retrieved Evidence Passages
[EVIDENCE_PASSAGES]

### Clinical Guidelines Context
[GUIDELINES_CONTEXT]

## OUTPUT SCHEMA
Return a valid JSON object with the following structure. Do not include any text outside the JSON object.

{
  "clinical_question_summary": "A one-sentence restatement of the clinical question.",
  "patient_summary": "A one-sentence summary of the patient context relevant to evidence ranking.",
  "ranked_evidence": [
    {
      "rank": 1,
      "passage_id": "string identifier from the input passages",
      "relevance_score": 0.0-1.0,
      "relevance_rationale": "Why this passage is relevant to the specific patient and question.",
      "evidence_quality": "high" | "moderate" | "low" | "insufficient",
      "quality_rationale": "Assessment based on study design, sample size, recency, and guideline alignment.",
      "outcome_improvement_evidence": "Summary of outcome data, if present. State 'No outcome data reported' if absent.",
      "guideline_alignment": "aligned" | "conflicting" | "not_addressed" | "superseded",
      "guideline_alignment_detail": "Specific guideline reference and recommendation, if applicable.",
      "patient_specific_factors": "Factors that increase or decrease applicability to this patient.",
      "confounding_concerns": "Any confounding by indication, spectrum bias, or other methodological concerns specific to this patient.",
      "supporting_quote": "A verbatim quote from the passage that best supports the ranking."
    }
  ],
  "evidence_gaps": [
    "A specific gap in the retrieved evidence relevant to the clinical question."
  ],
  "overall_assessment": "A concise synthesis of the ranked evidence and its implications for the clinical question. Include a statement on the overall strength of the evidence.",
  "uncertainty_flags": [
    "A specific source of uncertainty that should be communicated to the clinician."
  ]
}

## CONSTRAINTS
1. Rank passages by patient-specific applicability first, then by evidence quality.
2. Explicitly flag when a passage describes a population that does not match the patient.
3. Identify and explain any conflicts between passages or between a passage and the provided guidelines.
4. If a passage reports only surrogate outcomes, note this limitation.
5. Do not infer patient data not provided in the Patient Context.
6. If no evidence is sufficient to answer the clinical question, set the overall_assessment to reflect this and populate evidence_gaps thoroughly.
7. For any recommendation that would change the patient's management, include an uncertainty_flags entry recommending human review.

To adapt this template, start by mapping your retrieval pipeline's output into the [EVIDENCE_PASSAGES] placeholder. Each passage must include a unique passage_id and the full text. The [GUIDELINES_CONTEXT] placeholder should be populated with the most recent, authoritative guidelines relevant to the clinical domain. If no guidelines are available, replace the content with 'No specific guidelines provided.' The output schema is designed to be parsed by a JSON validator; implement a post-processing step that checks for the presence of all required fields and rejects malformed responses, triggering a retry or fallback. For high-risk deployments, route any output where uncertainty_flags is non-empty or evidence_quality is 'low' or 'insufficient' to a human review queue before displaying it to a clinician.

This prompt is designed to operate inside a clinical decision support pipeline, not as a standalone chat interface. The harness must enforce strict input validation before the prompt is assembled: patient context must be de-identified or handled under a BAA, the evidence list must be deduplicated and truncated to a manageable token window (typically 8–15 passages), and the output schema must be pre-registered so the model response can be parsed and validated deterministically. Because CDS outputs can influence care decisions, the harness must treat every response as provisional until it passes downstream checks for confounding by indication, spectrum bias, and guideline alignment.

Wire the prompt into an application flow that follows this sequence: (1) retrieve candidate evidence passages from a clinical knowledge base, guideline repository, or literature index using a patient-context-aware query; (2) deduplicate and filter passages by recency, study design minimums, and patient-population overlap; (3) assemble the prompt with the ranked [PATIENT_CONTEXT], [EVIDENCE_LIST], [OUTPUT_SCHEMA], and [CONSTRAINTS] fields; (4) call the model with a low temperature (0.1–0.2) and request structured JSON output matching the pre-defined schema; (5) validate the response against the schema, check that every ranked passage includes a patient_match_rationale and outcome_evidence_strength field, and flag any ranking that cites a study population clearly mismatched to the patient; (6) log the full prompt, response, and validation results for audit; (7) if confidence scores fall below a configurable threshold or if top-ranked evidence conflicts with active guidelines, route to a human reviewer before surfacing in the CDS UI. Use a model with strong medical reasoning capabilities and consider fine-tuning on domain-specific ranking examples if generic models struggle with evidence hierarchy distinctions.

Build eval suites that test for known failure modes before deployment. Include test cases where the highest-quality study (RCT, large sample) conflicts with a lower-quality but more population-matched observational study—the prompt should prioritize population match when instructed but not ignore study design entirely. Test for confounding by indication by feeding cases where a treatment appears harmful only because sicker patients received it. Test for spectrum bias with evidence drawn from narrow trial populations applied to broader real-world patients. Add regression tests that verify the output schema remains stable across prompt version changes. For high-risk deployments, implement a shadow-mode evaluation where the prompt runs alongside existing CDS logic for a defined period, with outputs compared by clinical reviewers before the prompt influences live decision support. Never deploy this prompt without a human-review escape hatch for low-confidence, conflicting, or guideline-discordant rankings.