Hallucinated Medical Term Detection Prompt for Clinical Summaries

This prompt is a post-generation safety net for clinical documentation teams. It takes a model-generated clinical summary and the original source notes, then identifies any medical terms, diagnoses, medications, or procedures in the summary that cannot be traced to the source. The output is a structured audit report that flags each hallucinated term, assigns a severity level, and recommends clinician review before the summary enters any patient record or downstream system. Use this when your product generates clinical text and you need a verifiable, auditable barrier between model output and clinical use.

Do not use this as a replacement for source-grounded generation or retrieval-augmented generation. It is a repair and validation step, not a primary generation strategy. The prompt requires both the model-generated summary and the complete source notes as inputs; without both, it cannot perform reliable cross-referencing. Implement this as a mandatory gate in your clinical documentation pipeline—every generated summary passes through this audit before any downstream system ingests it. The severity levels (Critical, High, Medium, Low) should map to your organization's clinical risk tolerance, with Critical findings automatically blocking the summary from entering the patient record until a clinician resolves them.

Before deploying, calibrate the prompt against a golden dataset of summaries with known hallucinated terms to establish your acceptable false-positive and false-negative rates. Pair this prompt with human-in-the-loop review for all Critical and High-severity flags, and log every audit result for compliance traceability. Never use this prompt to validate summaries where the source notes are incomplete or ambiguous—the model cannot distinguish between a term that is genuinely missing from the source and one that is implied but not explicitly stated. In those cases, flag the ambiguity for clinician review rather than assuming hallucination.

The prompt below is designed to be pasted directly into your prompt layer. It instructs the model to act as a clinical documentation auditor, comparing a generated summary against the original source notes. Every medical term, diagnosis, medication, procedure, or lab value in the summary must be traced back to the source. The prompt uses square-bracket placeholders—replace each with live data from your application before sending the request. No placeholder should remain unresolved in production.

textCopy
You are a clinical documentation integrity auditor. Your task is to compare a model-generated clinical summary against the original source notes and identify every medical term, diagnosis, medication, procedure, lab value, or clinical finding in the summary that cannot be traced to the source.

## INPUT

### Source Notes
[SOURCE_NOTES]

### Generated Clinical Summary
[CLINICAL_SUMMARY]

## INSTRUCTIONS

1. Extract every discrete medical entity from the summary. Include:
   - Diagnoses and conditions
   - Medications, dosages, and routes
   - Procedures and interventions
   - Lab tests, values, and units
   - Anatomical references
   - Clinical findings and observations
2. For each entity, search the source notes for supporting evidence. Consider:
   - Exact matches
   - Synonym matches (e.g., "HTN" matches "hypertension")
   - Abbreviation expansion (e.g., "MI" matches "myocardial infarction")
   - Implicit clinical equivalents (e.g., "elevated troponin" supports "myocardial injury")
3. Classify each entity as:
   - **GROUNDED**: Clear evidence exists in source notes.
   - **PARTIALLY_GROUNDED**: Some but not all aspects are supported (e.g., condition mentioned but not severity).
   - **HALLUCINATED**: No evidence found in source notes.
4. For HALLUCINATED and PARTIALLY_GROUNDED entities, provide:
   - The exact text from the summary
   - The entity type
   - A brief explanation of why it could not be fully grounded
   - A severity rating: CRITICAL (would affect clinical decisions), MODERATE (adds unsupported detail), or MINOR (stylistic or non-clinical)

## OUTPUT FORMAT

Return a JSON object with this exact structure:

{
  "summary_id": "[SUMMARY_ID]",
  "audit_timestamp": "[AUDIT_TIMESTAMP]",
  "total_entities_extracted": <integer>,
  "grounded_count": <integer>,
  "partially_grounded_count": <integer>,
  "hallucinated_count": <integer>,
  "findings": [
    {
      "entity_text": "<exact text from summary>",
      "entity_type": "<diagnosis|medication|procedure|lab_value|anatomy|finding>",
      "grounding_status": "<GROUNDED|PARTIALLY_GROUNDED|HALLUCINATED>",
      "source_evidence": "<quoted source text or null if none>",
      "explanation": "<brief explanation of grounding decision>",
      "severity": "<CRITICAL|MODERATE|MINOR>",
      "recommended_action": "<REMOVE|FLAG_FOR_REVIEW|RETAIN>"
    }
  ],
  "requires_clinician_review": <boolean>,
  "review_urgency": "<ROUTINE|PRIORITY|IMMEDIATE>"
}

## CONSTRAINTS

- Do not invent medical knowledge. Ground every decision in the provided source notes only.
- If the source notes are ambiguous, classify as PARTIALLY_GROUNDED and explain the ambiguity.
- Never downgrade severity to avoid flagging. Err toward flagging when uncertain.
- If the summary contains no hallucinated entities, return an empty findings array with grounded_count equal to total_entities_extracted.
- Do not include entities from the source notes that are absent from the summary. Audit the summary only.
- If [RISK_LEVEL] is "high", set requires_clinician_review to true for any HALLUCINATED or PARTIALLY_GROUNDED finding.

To adapt this prompt for your system, replace [SOURCE_NOTES] with the raw clinical notes (history, physical exam, labs, orders), [CLINICAL_SUMMARY] with the model-generated summary text, and [SUMMARY_ID] with your internal record identifier. Set [AUDIT_TIMESTAMP] to the current ISO 8601 time. The [RISK_LEVEL] parameter should be driven by your application context—use "high" for summaries that will be signed, submitted to payers, or inserted into the legal medical record. For draft summaries or internal review workflows, "standard" may be appropriate. The output schema is designed for direct ingestion into a review queue; map requires_clinician_review and review_urgency to your clinical workflow routing logic.

Before deploying, validate that your prompt layer resolves every placeholder. An unresolved [SOURCE_NOTES] or [CLINICAL_SUMMARY] will produce a useless audit. Wire the output into a structured validation step that checks hallucinated_count > 0 and triggers the appropriate review workflow. Do not auto-accept summaries where requires_clinician_review is true—route them to a human clinician. Log every audit result with the summary ID and timestamp for retrospective quality analysis.

This prompt is designed to sit in a post-generation validation layer within a clinical summarization pipeline. After a model generates a clinical summary from source notes, this prompt acts as a safety net: it compares every medical term, diagnosis, medication, and procedure in the summary against the original source text and flags any term that cannot be grounded. The output is not a corrected summary but a structured audit report that downstream systems or human reviewers can act on. Do not use this prompt as a real-time filter that silently strips terms—fabricated clinical content must be surfaced for review, not hidden.

Integration pattern: Wire this prompt as a synchronous validation step after summary generation and before the summary enters any clinical record system. The input harness requires two fields: [SOURCE_NOTES] (the original clinical notes, verbatim) and [GENERATED_SUMMARY] (the model-produced summary to audit). The output schema should be a JSON array of flagged terms, each with the term itself, its category (diagnosis, medication, procedure, anatomical reference, or other), a grounding status (grounded, ungrounded, or partially_grounded), the source evidence span if grounded, and a recommended action (review, remove, or verify). Implement a strict JSON schema validator on the response before processing it—if the model returns malformed JSON, retry once with an explicit formatting reminder. If the second attempt fails, escalate the entire summary for full human review rather than risking silent passage of fabricated terms.

Model choice and latency: Use a model with strong instruction-following and medical terminology familiarity—GPT-4, Claude 3.5 Sonnet, or equivalent. This is a high-stakes validation step where precision matters more than latency. Set temperature=0 to minimize variance across runs. If your pipeline processes summaries in batch, run this validation asynchronously and queue flagged summaries for clinician review. Log every validation result, including the prompt version, model version, input hash, and the full flagged-terms payload, for auditability. If a term is flagged as ungrounded, the system should block automated insertion into the EHR or clinical record and route the summary to a review queue with the audit report attached.

Human review integration: The flagged terms output should feed directly into a review UI where a clinician can see the original source notes side-by-side with the generated summary, with ungrounded terms visually highlighted. The clinician can confirm removal, override the flag if the term is clinically appropriate, or edit the summary. Track override rates by term category to identify patterns—if certain medication names or procedure codes are consistently flagged but clinically correct, you may need to expand your source-grounding logic or add a terminology allowlist. Never auto-accept ungrounded terms without human confirmation in a clinical context. The prompt's job is to reduce the review surface area, not eliminate human judgment.

Failure modes to monitor: Watch for false negatives where the model fails to flag a genuinely fabricated term because it appears plausible in context. Implement periodic spot-checking where a clinical reviewer audits a random sample of summaries that passed validation with zero flags. Also monitor for false positives where common medical abbreviations or implied diagnoses (e.g., 'hypertension' inferred from 'HTN' in source notes) are flagged as ungrounded. You can reduce these by including an abbreviation expansion step before validation or by adding a terminology normalization layer that maps common clinical abbreviations to their standard forms before grounding checks. If your pipeline includes RAG retrieval from external knowledge bases, ensure the source notes field contains only the original clinical notes, not retrieved content, to prevent the grounding check from being contaminated by external sources.