An AI coding agent for medical billing just learned to keep a list of its own mistakes, and it works, sort of.
Researchers built a system called LearnActCoder that splits medical coding into two AI roles: a Coder that hunts for missed diagnoses and a Judge that vetoes unsupported ones. Errors from a small labeled training batch get logged into a structured error memory, called MistakeKDB, that the agents consult before making calls on new charts. Tested on 150 matched clinical notes from the MIMIC-III dataset, the structured memory boosted CPT procedure-code F1 scores by 5.9 percentage points over a no-memory baseline; the corresponding ICD-9 diagnosis-code improvement was not statistically significant. On a separate MIMIC-IV cohort, the same memory pushed ICD-10 coding toward higher precision but at the cost of recall, leaving overall F1 essentially flat.
Medical coding errors are expensive: they drive claim denials and reimbursement fights, and most fixes so far have meant retraining or rebuilding the underlying model. This approach skips that step. It patches behavior by editing a lookup memory, no weight updates required, which matters for anyone trying to keep a coding tool current as billing rules shift without a full retraining cycle.
The gains are real but modest, and the paper is careful to say so: CPT and HCPCS accuracy stayed low overall, and the whole study is retrospective, run on archived hospital notes rather than a live billing department. A memory bank of past mistakes is a reasonable patch for an imperfect coder, but it is not the same as an AI ready to sign off on an actual claim.