How AI clinical coding works, and where general-purpose AI gets it wrong
Short answer
AI clinical coding reads a clinical note, finds the documented diagnoses and procedures, and suggests the matching codes. Reliable systems don't stop at a prediction: they look each code up in the official code set, check it against coding guidelines, and show the text that supports it so a coder can approve it. General-purpose chatbots skip those steps, which is why their codes can look right and still be wrong.
Key facts
- Clinical coding is a look-up and rules problem, not only a language problem.
- Common chatbot errors: codes that don’t exist, codes from an outdated year, unspecified codes when the note supports more detail, and pairs of codes the guidelines don’t allow together.
- ICD-10-CM and ICD-10-PCS change every October 1, so a model’s training data goes out of date.
- In a safe workflow the AI drafts and a qualified coder approves every code.
What does a clinical coder actually do?
A coder reads the whole record, decides which conditions and procedures are reportable, finds each one in the code set's index and checks it in the tabular list, then applies the official guidelines: which diagnosis is principal, which codes combine, which must be sequenced first, and what the documentation does and doesn't support. When the record is unclear, the coder sends the physician a query.
Every one of those steps can go wrong if a model only predicts what a code probably is.
Where does general-purpose AI go wrong?
Ask a chatbot to code a note and the output usually looks convincing. The errors are in the details:
- Invented or invalid codes: plausible codes that are not in the code set, or category codes that need more characters to be billable.
- Outdated codes: codes deleted or expanded in a later annual update.
- Lost specificity: an unspecified code when the note documents the type, site or laterality.
- Guideline conflicts: Excludes1 pairs reported together, or a manifestation code without the underlying condition sequenced first.
- Coding what isn’t a diagnosis: abnormal lab values without a provider’s diagnosis, or, in outpatient care, conditions documented only as “rule out” or “probable”.
- No evidence: nothing shows which sentence led to which code, so every code has to be checked against the whole note.
How is an AI coding pipeline built?
Accurate AI coding treats the model as one step in a pipeline, with checks that do not depend on the model being right:
- Read: find every diagnosis and procedure in the note, with its context, such as negated, historical or uncertain.
- Look up: match each one to a code in the right code set and version.
- Validate: check every code against the rules, such as billable level, Excludes1, code-first and use-additional notes, 7th characters, and age and sex edits.
- Review: show each code with its supporting passage, for a coder to accept, change or reject.
What should you ask an AI coding vendor?
Before trusting any tool with production coding, ask:
- Does every code link to the text that supports it?
- Which code-set versions does it use, and how fast are annual updates applied?
- Which guidelines are enforced as rules rather than left to the model?
- Is patient data hidden from the AI, and is it ever used for training?
- Who approves the final codes?
- Can we test it on our own de-identified charts before buying?