Using AI for ICD-10 Code Lookup

ICD-10-CM is the U.S. diagnosis code set. The National Center for Health Statistics maintains it with CMS, and it is updated every year, effective October 1. Codes run from three to seven characters and always start with a letter. Some need a placeholder X, and many injury codes need a seventh character that shows the encounter type. The official guidelines say to look up a term in the Alphabetic Index first and then confirm it in the Tabular List. The Tabular List holds the instructional notes that the Index leaves out, such as Excludes1, Excludes2, "code first" and "use additional code." AI lookup tools work in two main ways. Retrieval tools search the official code files and index terms, then rank matches. Generative chatbots write an answer based on patterns they learned in training. That is why chatbots invent codes. ICD-10-CM codes follow regular patterns, so a model can produce a string with the right shape and a believable description without checking that the code exists. A model's training data may also be older than the latest October update. Independent research has found that general-purpose language models often get medical codes wrong when asked to produce them directly. A few misconceptions are common. The first is that a real code must be the right code. A code can exist and still be less specific than the documentation, or clash with an Excludes1 note. The second is that unspecified codes are always wrong. They are acceptable when the record does not support anything more specific. The third is that setting does not matter. In outpatient coding, uncertain diagnoses such as "probable" or "rule out" are not coded as if confirmed. The coder reports the documented signs and symptoms instead. Inpatient rules treat uncertain diagnoses differently. Whatever the tool, codes come from the provider's documentation, never from what the AI infers.