Contribution · Application — Healthcare

AI-Assisted Medical Coding (ICD-10 and CPT)

Medical coding converts physician narratives into ICD-10-CM, ICD-10-PCS, CPT, and HCPCS codes for billing. Experienced coders take 10-20 minutes per inpatient encounter; AI coding assist can cut that by 50-70% while improving coding accuracy. The caveat: upcoding, downcoding, or hallucinated codes create both revenue loss and fraud exposure under the False Claims Act in the US and under GST / IRDA rules in India.

Application facts

Domain
Healthcare
Subdomain
Revenue Cycle Management
Example stack
Claude Opus 4.7 or GPT-5 for code suggestion · LangChain retrieval over ICD-10-CM / CPT embeddings in pgvector · UMLS Metathesaurus for concept normalization · Pydantic structured outputs for billable-code contracts · Audit store in append-only Postgres with coder override tracking

Data & infrastructure needs

  • Clinical notes — H&Ps, progress notes, operative reports
  • ICD-10-CM and ICD-10-PCS codebooks (licensed)
  • CPT codebook (AMA-licensed)
  • HCC risk-adjustment mappings
  • Certified coder gold standard for evaluation

Risks & considerations

  • Upcoding or unbundling triggering False Claims Act exposure
  • Hallucinated codes not supported by documentation
  • PHI leakage through model providers without BAAs
  • Outdated codes after annual ICD-10 / CPT updates
  • Bias in code selection for under-documented populations

Frequently asked questions

Which LLM works best for medical coding?

Both general-purpose models (Claude Opus 4.7, GPT-5) and specialized models (3M M*Modal, Solventum CAC) are viable as of 2026. Accuracy depends more on RAG over ICD-10-CM / CPT knowledge graphs and on fine-tuning against certified coder decisions than on base model choice.

Is AI medical coding compliant with HIPAA and CMS rules?

Yes, when deployed under a HIPAA Business Associate Agreement with encryption, access controls, and audit logs. CMS requires all billed codes to be substantiated by the medical record; AI-suggested codes must be reviewable and auditable by humans before submission.

What is the biggest risk?

Upcoding — AI adding codes not substantiated by documentation — creates False Claims Act liability in the US. Deployments must use low-temperature extraction, require human confirmation on billable codes, and maintain a full decision trail for every claim.

Sources

  1. CMS — ICD-10 resources — accessed 2026-04-20
  2. AAPC — AI in medical coding — accessed 2026-04-20
  3. AHIMA — Coding standards — accessed 2026-04-20

Related

  • AI for Clinical Documentation Summarization — Clinical summarization uses LLMs to condense patient records, consult notes, and discharge summaries — a high-value, high-risk application requiring RAG, evaluation, and audit trails.
  • AI for Prior Authorization Automation — Prior authorization — insurer approval before a service is rendered — is a slow, paperwork-heavy bottleneck. AI automates eligibility checks, policy lookup, and clinical evidence extraction to speed approval decisions and reduce denials.
  • AI for Support Knowledge Base Generation — Support teams use LLMs to turn resolved tickets, engineering docs, and SME conversations into searchable knowledge base articles — keeping KB up to date without a dedicated writer.
  • All Applications — Browse the complete applications library
  • VSET — Programmes offered — B.Tech programs (CSE, AI&ML, AI&DS, Electronics, IIoT)