Medical Literature Evidence Scoring Prompt Template

This prompt is for clinical decision-support engineers who need to rank retrieved medical studies by evidence quality before generating answers. Use it when your RAG pipeline returns multiple studies for a clinical question and you must surface the strongest evidence first. The prompt applies evidence-based medicine (EBM) hierarchies, scoring each study by design, sample size, recency, and alignment with clinical practice guidelines. It outputs a ranked evidence table with strength-of-recommendation tags.

Do not use this prompt for primary diagnosis, for replacing systematic review methodologies, or when retrieved passages lack study metadata. This prompt assumes your retrieval step already returns study abstracts or structured metadata. It does not perform retrieval itself. The prompt is designed for a second-pass ranking step in a clinical RAG pipeline, not for generating final clinical recommendations without human review. If your system handles patient-specific data, you must add a human-in-the-loop gate after evidence ranking and before any answer generation.

Before deploying, validate that your retrieval pipeline consistently returns study design, sample size, publication date, and guideline references in a parseable format. The prompt will degrade silently if fed unstructured or incomplete metadata. Plan an eval suite that checks for conflating correlation with causation, over-weighting large observational studies over smaller RCTs, and missing retracted or superseded studies. If your use case involves real-time clinical decision-making, pair this prompt with a human approval step and log every ranked output for audit.

This prompt template is designed to be pasted directly into your system instructions or user message for a medical literature evidence scoring task. It instructs the model to act as a clinical evidence analyst, ranking retrieved studies according to evidence-based medicine (EBM) hierarchies. The template uses square-bracket placeholders for all dynamic inputs, ensuring you can adapt it to different clinical questions, retrieved document sets, and output format requirements without modifying the core logic.

textCopy
You are a clinical evidence analyst. Your task is to evaluate and rank the provided medical literature passages based on evidence-based medicine principles.

## INPUT
- Clinical Question: [CLINICAL_QUESTION]
- Patient Context (if any): [PATIENT_CONTEXT]
- Retrieved Studies: [RETRIEVED_STUDIES]

## RANKING CRITERIA
For each study, assign a score from 1 (weakest) to 5 (strongest) for each of the following dimensions. Provide a brief justification for each score.
1. Study Design: Score based on the EBM hierarchy (Systematic Review/Meta-Analysis of RCTs > RCT > Cohort Study > Case-Control > Case Series/Expert Opinion).
2. Sample Size and Power: Score based on the adequacy of the sample size to detect a clinically meaningful effect.
3. Recency: Score based on the publication date relative to the current standard of care. Penalize studies older than [RECENCY_THRESHOLD_YEARS] years unless they are seminal works.
4. Guideline Alignment: Score based on alignment with major clinical practice guidelines from [RELEVANT_GUIDELINE_BODIES].
5. Directness: Score based on how directly the study population, intervention, and outcome match the clinical question and patient context.

## OUTPUT FORMAT
Return a single JSON object with the following schema:
{
  "clinical_question": "string",
  "evidence_table": [
    {
      "study_id": "string",
      "title": "string",
      "study_design": "string",
      "scores": {
        "study_design": {"score": integer, "justification": "string"},
        "sample_size_power": {"score": integer, "justification": "string"},
        "recency": {"score": integer, "justification": "string"},
        "guideline_alignment": {"score": integer, "justification": "string"},
        "directness": {"score": integer, "justification": "string"}
      },
      "composite_score": float,
      "strength_of_recommendation": "Strong" | "Moderate" | "Weak" | "Insufficient",
      "key_limitations": ["string"]
    }
  ],
  "evidence_gaps": ["string"],
  "overall_assessment": "string"
}

## CONSTRAINTS
- Do not fabricate study details. If information for a score is missing from the retrieved text, note it as "Not reported" in the justification and assign a score of 1.
- The composite_score is the mean of the five dimension scores.
- Strength_of_recommendation mapping: >=4.0 is Strong, 3.0-3.9 is Moderate, 2.0-2.9 is Weak, <2.0 is Insufficient.
- Flag any study with a clear conflict of interest or industry funding in the key_limitations field.
- If no studies are provided or none are relevant, return an empty evidence_table and explain the gap.

To adapt this template for your application, replace the square-bracket placeholders with your data. The [CLINICAL_QUESTION] should be a focused, answerable question, ideally in PICO format. [PATIENT_CONTEXT] can include demographics, comorbidities, and preferences that affect directness scoring. [RETRIEVED_STUDIES] should be a pre-fetched set of study abstracts or summaries from your retrieval pipeline. Set [RECENCY_THRESHOLD_YEARS] based on the rate of change in the clinical domain—5 years is a common default, but fast-moving fields like oncology may require 2 years. [RELEVANT_GUIDELINE_BODIES] should list the specific organizations whose guidelines you want the model to reference, such as 'AHA/ACC' or 'NICE'. Before deploying, validate the JSON output against the schema and run evals to ensure the model does not conflate correlation with causation or overstate the strength of observational studies.

This prompt is designed to operate as a scoring and ranking step within a larger clinical evidence pipeline, not as a standalone chatbot. The typical integration pattern is: retrieve candidate studies from a biomedical search API (PubMed, Semantic Scholar, or an internal vector store), pass the top-N results into this prompt along with the clinical question and patient context, receive a scored evidence table, and then feed the top-ranked studies into a downstream synthesis or answer-generation step. The prompt expects structured inputs and produces structured outputs, which makes it suitable for API-driven workflows where the model response is parsed by application code before any user-facing content is rendered.

Input assembly requires four components mapped to the prompt's placeholders: [CLINICAL_QUESTION] is the PICO-formatted query (Population, Intervention, Comparison, Outcome); [PATIENT_CONTEXT] contains relevant demographics, comorbidities, and medications as structured text; [RETRIEVED_STUDIES] is a JSON array of study objects with fields for title, abstract, study design, sample size, publication date, journal, and DOI; and [EVIDENCE_HIERARCHY] defines the ranking weights for study designs (e.g., meta-analysis > RCT > cohort > case-control > case series). The application layer should validate that each study object contains the required fields before assembly, rejecting or flagging records with missing abstracts or unknown study designs. For production systems, implement a pre-filter that removes retracted papers (checking against the PubMed retraction index or Retraction Watch database) and studies older than a configurable recency threshold unless the clinical domain explicitly requires historical evidence.

Output parsing and validation must be strict because downstream synthesis depends on correctly ranked evidence. The prompt requests a JSON output with a ranked_evidence array where each entry includes rank, study_id, composite_score, strength_of_recommendation, and scoring_rationale. Implement a post-processing validator that checks: (1) the response is valid JSON, (2) every study_id in the output matches an input study, (3) composite_score values are numeric and within the expected range, (4) strength_of_recommendation tags use only the allowed enum values (Strong, Moderate, Weak, Insufficient), and (5) no study appears more than once. If validation fails, use a retry loop with the error details appended to the prompt context—but cap retries at two attempts before logging the failure and routing to a human review queue. Log every scoring run with the input query, model version, output scores, and validation status for auditability.

Model selection and latency considerations matter here because evidence scoring requires careful reasoning over study methodology. Use a model with strong instruction-following and structured output capabilities (GPT-4o, Claude 3.5 Sonnet, or equivalent). For high-throughput clinical workflows, consider batching multiple clinical questions against the same retrieval set, but be aware that context windows fill quickly with full abstracts. If latency exceeds 5 seconds, implement a streaming parse that extracts the JSON array incrementally, or pre-score studies with a lightweight classifier model and use this prompt only for the top-20 candidates. For regulated clinical decision-support systems (FDA Class II or EU MDR software), this prompt output must be treated as decision support, not decision replacement—always route scored evidence through a clinician review interface that displays the rationale and allows re-ranking or dismissal before any recommendation reaches the patient.

Failure modes to monitor include: the model conflating correlation with causation when scoring observational studies (mitigate by requiring explicit causal language checks in eval), overweighting recent but small studies over older landmark trials (mitigate by tuning recency weights in the hierarchy definition), hallucinating study details when abstracts are truncated (mitigate by enforcing minimum abstract length in pre-processing), and producing inconsistent scores for the same study across similar clinical questions (mitigate by running periodic eval suites with golden-scored reference sets and measuring rank correlation metrics like Kendall's tau). The most dangerous failure is the model assigning a high strength-of-recommendation tag to a single small study that contradicts established guidelines—always cross-check top-ranked studies against a guideline reference database when available, and flag discrepancies for human review.