Skip to main content

LLM-Based Diagnosis Code Extraction: The Architecture

Duration: 60 min · Level: Advanced · Module: 3. AI Medical Coding & CDI · Focus: ICD-10, LLM-coding, diagnosis-extraction, Claude, AI-coding

Learning objectives

By the end of this lesson you will be able to explain and apply:

  • LLM coding approach
  • Context engineering
  • Principal diagnosis selection
  • MCC/CC coding impact
  • AHIMA/AHA Coding Clinic

You will then consolidate these ideas in the hands-on lab below.

Why this matters​

Large language models understand clinical text better than any previous NLP technology.

Overview​

Large language models understand clinical text better than any previous NLP technology. A well-prompted Claude or GPT-4 can read a discharge summary and suggest appropriate ICD-10-CM codes with accuracy that approaches — and in some domains exceeds — trained human coders. The engineering challenge is not the model; it is the context, the validation layer, and the workflow integration.

Key concepts​

Key idea

LLM coding approach: feed the complete discharge summary + progress notes + operative report → ask model to identify all reportable diagnoses and procedures → model returns codes with supporting text citations → validation layer checks codes against ICD-10 tabular list

  • Context engineering: LLMs need the complete clinical picture; feeding only the discharge diagnosis list misses secondary diagnoses, CCs/MCCs (complication/comorbidity codes that increase DRG weight significantly)
  • Principal diagnosis selection: the condition "after study" chiefly responsible for admission; LLMs sometimes misidentify this as the patient's chief complaint; fine-tuning or few-shot examples on specialty-specific cases improves accuracy
  • MCC/CC coding impact: each additional complication or comorbidity code (sepsis, respiratory failure, acute kidney injury) can add $5,000-$15,000 in DRG reimbursement for the same admission; AI catches these more consistently than fatigued human coders
  • AHIMA/AHA Coding Clinic: official guidance on ICD-10 coding conventions; must be incorporated into AI system prompts as authoritative reference; published quarterly; subscription required
  • Model benchmark (2024 data): Claude 3 Opus achieves 91-94% code-level accuracy on Medicare discharge summaries in untuned zero-shot setting; fine-tuned models achieve 96-98% on in-distribution data
Hands-on lab

Build a minimal coding agent: (1) take a 500-word simulated discharge summary as input, (2) prompt Claude to extract principal diagnosis, secondary diagnoses, and procedure codes with textual evidence, (3) validate each code against a local ICD-10 code list, (4) output a structured JSON with codes, confidence, and supporting text citation.

Check your understanding​

Try to recall each answer before expanding it.

Q1. What do you know about LLM coding approach?

feed the complete discharge summary + progress notes + operative report → ask model to identify all reportable diagnoses and procedures → model returns codes with supporting text citations → validation layer checks codes against ICD-10 tabular list

Q2. What do you know about Context engineering?

LLMs need the complete clinical picture; feeding only the discharge diagnosis list misses secondary diagnoses, CCs/MCCs (complication/comorbidity codes that increase DRG weight significantly)

Q3. What do you know about Principal diagnosis selection?

the condition "after study" chiefly responsible for admission; LLMs sometimes misidentify this as the patient's chief complaint; fine-tuning or few-shot examples on specialty-specific cases improves accuracy

Q4. What do you know about MCC/CC coding impact?

each additional complication or comorbidity code (sepsis, respiratory failure, acute kidney injury) can add $5,000-$15,000 in DRG reimbursement for the same admission; AI catches these more consistently than fatigued human coders

Q5. What do you know about AHIMA/AHA Coding Clinic?

official guidance on ICD-10 coding conventions; must be incorporated into AI system prompts as authoritative reference; published quarterly; subscription required

References​

  • Automated ICD Coding with Large Language Models — Huang et al. (2024). JAMIA 2024

Next: H3.2 CDI: Generating Physician Queries Automatically →

Part of Module 3: AI Medical Coding & CDI.