Healthcare / Research prototype
From a patient note to an inspectable coding decision.
A small ICD-10-CM research checker connects proposed codes to explicit evidence and keeps uncertainty visible.
Start with the documentation
A symptom is evidence about a patient’s presentation. It is not, by itself, proof of a particular disease. A coding workflow must preserve that distinction while recording what the clinician actually documented, which encounter setting applies and which code-set revision is in scope.
This prototype starts after a candidate interpretation has been proposed. It checks structured evidence against a synthetic note and a small code catalogue. The model’s task, if one is used upstream, is to identify possible evidence. Acceptance belongs to the explicit checker.
A narrow first model
The repository uses a deliberately small set of examples so the evidence path is inspectable. A proposal carries its code and supporting documentation. Checks bind the proposal to the expected note and supported policy context. An unsupported code or missing evidence cannot silently become a successful result.
The official outpatient guidance distinguishes uncertain diagnoses from documented signs, symptoms and reasons for the encounter. The research design takes that distinction seriously: a speculative diagnosis is a case to withhold or review, and a symptom must be represented at the level the documentation supports.
What the proof means
A theorem can show that a successful result requires the conditions encoded in the checker. Exact evidence matching can detect a changed quotation or mismatched note. It cannot establish that a quoted sentence is medically true, complete or sufficient under every rule in the codebook.
Review the source, the named theorem statements and the executable negative examples together. The meaningful boundary cases include negation, uncertainty, unsupported scope and missing documentation. The whitepaper explains which checks are implemented and where human or model interpretation remains trusted.
How to evaluate it
The included synthetic examples are regression tests for the encoded rules. They make rejection behavior reproducible and give reviewers a small surface to inspect. They are not patient data and do not establish clinical coding accuracy.
A larger evaluation needs independently labeled notes, expert review, an exact code-set release and clear rules for encounter setting. It should report supported code coverage, evidence-span accuracy, accepted-case errors and review rates. Code selection and reimbursement must be evaluated separately.
The next research step
Expand the catalogue only alongside sourced rules, examples that challenge those rules and an explicit review of the statement being proved. A larger list of codes does not automatically make a stronger checker. The aim is to preserve the chain from documented fact to a particular, justified coding decision.