Clinical Finding Extraction Prompt for Discharge Summaries

This prompt is designed for clinical data engineers and healthcare NLP teams who need to convert unstructured discharge summaries into typed, queryable records. It extracts four core payloads: diagnoses, procedures, medications, and follow-up instructions. Each finding includes source text spans, SNOMED CT or ICD-10-CM codes where mappable, negation status, and a confidence indicator. The prompt is built for production pipelines where missing fields, provisional diagnoses, and negated findings must be handled explicitly rather than silently dropped.

Use this prompt when you have raw discharge summary text and need structured JSON that can be ingested into a clinical data warehouse, EHR analytics system, or research database. The prompt expects a complete discharge summary as input, not isolated sentences or partial notes. It works best with standard narrative sections—Hospital Course, Discharge Diagnoses, Discharge Medications, and Follow-Up Instructions—but can handle variations in section naming. Do not use this prompt for real-time clinical decision support, for extracting findings from imaging reports or lab results, or for generating billing codes without human review. The prompt is not a substitute for certified medical coding software and should not be used as the sole source of truth for regulatory submissions.

The prompt requires several inputs to function correctly: the full discharge summary text, a target output schema defining the four payload types, a terminology code system preference (SNOMED CT, ICD-10-CM, or both), and a confidence threshold for flagging uncertain extractions. You must also provide clear instructions for handling negated findings—such as 'no evidence of pneumonia'—and provisional diagnoses like 'suspected sepsis.' Without these constraints, the model may extract negated conditions as active problems or treat provisional diagnoses as confirmed. The prompt includes explicit handling for these edge cases, but you should validate outputs against your organization's clinical documentation improvement (CDI) guidelines.

Before deploying this prompt, prepare a golden dataset of 20–50 annotated discharge summaries with known diagnoses, procedures, medications, and follow-up items. Use this dataset to measure extraction recall, precision, and negation accuracy. Pay special attention to false positives from negated findings and false negatives from provisional or uncertain language. If your use case involves billing or quality reporting, add a human review step for all extractions with confidence below 0.85 or for any diagnosis mapped to a CMS quality measure. The prompt is a component in a larger pipeline, not a standalone solution—plan for validation, human review, and continuous monitoring of extraction quality over time.

This prompt template is designed to extract diagnoses, procedures, medications, and follow-up instructions from a hospital discharge summary. It enforces structured JSON output, maps findings to SNOMED CT and ICD-10-CM where possible, and explicitly handles negation, uncertainty, and provisional diagnoses. Replace each square-bracket placeholder with your specific data, schema definitions, and operational constraints before sending it to the model. The template is self-contained and can be used directly in an API call, a prompt management system, or a local evaluation harness.

textCopy
You are a clinical data extraction assistant. Your task is to read a hospital discharge summary and extract structured clinical findings.

## INPUT
Discharge Summary Text:
[DISCHARGE_SUMMARY_TEXT]

## OUTPUT SCHEMA
Return a single JSON object with the following structure:
{
  "patient_encounter": {
    "admission_date": "string or null (ISO 8601)",
    "discharge_date": "string or null (ISO 8601)",
    "encounter_type": "string (e.g., inpatient, observation, emergency)"
  },
  "diagnoses": [
    {
      "description": "string (the diagnosis as stated in the text)",
      "icd10cm_code": "string or null (best-match ICD-10-CM code)",
      "snomed_ct_code": "string or null (best-match SNOMED CT concept ID)",
      "status": "string (active, resolved, chronic, provisional, ruled-out)",
      "is_primary": boolean,
      "is_present_on_admission": boolean or null,
      "certainty": "string (confirmed, probable, possible, negated)",
      "evidence_span": "string (exact text supporting this extraction)"
    }
  ],
  "procedures": [
    {
      "description": "string",
      "cpt_code": "string or null",
      "snomed_ct_code": "string or null",
      "performed_date": "string or null (ISO 8601)",
      "status": "string (performed, planned, cancelled, contraindicated)",
      "evidence_span": "string"
    }
  ],
  "medications": [
    {
      "name": "string (generic name preferred)",
      "rxnorm_code": "string or null",
      "dose": "string or null",
      "route": "string or null",
      "frequency": "string or null",
      "status": "string (active, discontinued, prescribed, held, refused)",
      "is_discharge_medication": boolean,
      "evidence_span": "string"
    }
  ],
  "follow_up_instructions": [
    {
      "instruction": "string",
      "specialty": "string or null",
      "timeframe": "string or null (e.g., 1 week, 2 weeks, as needed)",
      "evidence_span": "string"
    }
  ],
  "extraction_metadata": {
    "negation_flags_detected": ["string (list of negated concepts found)"],
    "uncertainty_notes": ["string (list of ambiguous or conflicting statements)"],
    "missing_sections": ["string (expected sections not found in the text)"],
    "confidence_overall": "string (high, medium, low)"
  }
}

## CONSTRAINTS
- Use the exact JSON schema above. Do not add or remove fields.
- For every extracted finding, include the exact evidence_span from the text.
- If a finding is explicitly negated (e.g., "no evidence of pneumonia"), set certainty to "negated" and include it in the diagnoses list with status "ruled-out".
- If a diagnosis is described as provisional, possible, or suspected, set certainty to "probable" or "possible" and status to "provisional".
- Map to SNOMED CT and ICD-10-CM codes only when you have high confidence. Use null when uncertain.
- For medications, prefer generic names and map to RxNorm when possible.
- If a section of the discharge summary is missing (e.g., no follow-up instructions), list it in missing_sections.
- Do not fabricate findings. If the text does not support a field, use null or an empty array as appropriate.
- Set confidence_overall to "low" if the text is ambiguous, incomplete, or contains conflicting information.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK LEVEL
[HIGH_RISK_WARNING: This extraction is for clinical decision support and must be reviewed by a qualified clinician before use in patient care. Do not rely on automated extraction for billing, treatment decisions, or regulatory reporting without human validation.]

After copying the template, replace [DISCHARGE_SUMMARY_TEXT] with the raw text of the discharge summary you are processing. Replace [FEW_SHOT_EXAMPLES] with 2–4 annotated examples that demonstrate correct handling of negation, provisional diagnoses, and missing sections. Replace [HIGH_RISK_WARNING] with your organization's specific risk language and escalation policy. If you are using a model that supports structured output or function calling, you can replace the JSON schema block with a tool definition or response format parameter. Always validate the output against your schema before ingestion, and route extractions with confidence_overall: "low" or certainty: "possible" to a human review queue.

This prompt is designed to be called as a single step within a larger clinical data ingestion pipeline, not as a standalone chat interface. The discharge summary text should be pre-processed to remove any PHI that is not required for clinical extraction (per your organization's de-identification protocol) before being passed into the [DISCHARGE_SUMMARY_TEXT] placeholder. The [TARGET_SCHEMA] placeholder should be populated with a strict JSON Schema definition that matches your downstream data warehouse or FHIR store contract. The [TERMINOLOGY_BINDINGS] placeholder should contain a mapping of extracted concept types to their required coding systems (e.g., diagnosis: SNOMED CT, ICD-10-CM).

Wire the prompt into an application function that enforces a strict retry-and-review loop. On each invocation, parse the model's JSON output and validate it against the provided [TARGET_SCHEMA]. If validation fails, catch the specific schema errors and feed them back into the [PREVIOUS_OUTPUT] and [VALIDATION_ERRORS] placeholders of a retry prompt. Implement a maximum of two automatic retries. If the output still fails schema validation after two retries, or if the model's confidence field for any critical finding falls below a configurable threshold (e.g., 0.85), route the entire record to a human review queue. Log every attempt, the raw model output, validation errors, and the final disposition (auto-accepted, auto-rejected after retries, or escalated) for auditability.

For model choice, use a model with strong JSON mode and a large context window to accommodate lengthy discharge summaries and detailed schemas. Enable structured output mode (JSON mode) at the API level as a first line of defense against malformed responses, but do not rely on it alone—always apply the post-hoc schema validation. Do not use this prompt for real-time clinical decision support; it is designed for retrospective batch processing of finalized discharge summaries. The extracted data should be treated as a preliminary coding suggestion and must be reviewed by a qualified clinical coder before being used for billing, quality reporting, or downstream clinical analytics.