elisaterumi-ai/clinical-agent-skills/coding-icd

When mapping clinical text to ICD codes, follow these steps:

1. Identify codable conditions

Extract only diagnoses or conditions that are:

• Explicitly stated in the text
• Clinically meaningful and codable

Do not infer diagnoses from symptoms alone.

2. Normalize the diagnosis

• Interpret the clinical term as written
• Map it to the closest standard diagnostic concept
• Preserve specificity when possible (e.g., type, location, severity)

3. Assign ICD code

• Map each diagnosis to the most appropriate ICD code
• Prefer the most specific code available
• If specificity is insufficient, use a more general code

Do not fabricate or guess codes.

4. Handle multiple diagnoses

• Extract and code each diagnosis independently
• Avoid duplication
• Maintain clear mapping between diagnosis and code

5. Output format

Return the result as JSON:

{
  "codes": [
    {
      "diagnosis": "...",
      "icd_code": "...",
      "description": "..."
    }
  ]
}

6. Constraints

• Only code explicitly mentioned diagnoses
• Do not infer diagnoses from symptoms
• Do not hallucinate ICD codes
• If uncertain, prefer a higher-level (less specific) code
• Do not include explanations outside the JSON output