Healthcare software colleagues of mine often describe a recurring discussion they have with their outsourced development teams. They recount that when they try to explain the difference between a billing diagnosis and a clinical diagnosis, the engineers on the other end deliver a common platitude: “a diagnosis is a diagnosis.”
That answer is defensible in ordinary English. In healthcare data, however, it is wrong in a way every team that accepts it inherits. The formal name of the “diagnosis” system is the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10). Statisticians created it in 1893 to count causes of death, and around 1938, leaders suggested that the codes might be useful to health insurers. Medicare then tied hospital reimbursement to those codes in 1983, and the rest is history.
Accordingly, I have started referring to ICD-10 as a billing category rather than a diagnosis code. Colleagues tell me the phrasing sounds pedantic, but we have all witnessed what has happened due to this lack of differentiation. Software analysts carry the loose usage into product requirements documents, data dictionaries and integration specifications. Developers then build software that treats a category as though it were a clinical finding. We are all suffering the consequences of asking a statistical instrument to carry the complexity and nuance within clinical meaning.
Grouper or Splitter?
The vocabulary distinction reminds me of when clinical informaticists once asked each other whether they were groupers or splitters. Groupers gather related conditions under a shared heading, while splitters separate them down to the specific entity. The people who built the ICD chose grouping because a statistical classification exists to produce comparable counts across whole populations. Grouping is what makes an epidemiological instrument work, but a clinical one fail.
Conversely, clinical care has continued to split because treatment now turns on distinctions a single code cannot carry. One example is ICD-10 code G60.0, hereditary motor and sensory neuropathy. Coders assign that single category to Charcot-Marie-Tooth disease (CMT) and to its near relatives alike. Yet researchers have now linked more than 100 genes to CMT alone, and dozens of clinically distinct diseases are filed under that heading, each with its own inheritance pattern and treatment path. A claim clears on G60.0 without difficulty, while a physician reading G60.0 learns almost nothing.
Maintainers cannot resolve this by adding codes. They added 487 new diagnosis codes in the fiscal year 2026 update, which all offer important detail. Even so, a larger classification system is still a classification system rather than a clinical vocabulary.
Where Therapy Already Split
Pharmaceutical manufacturers have already committed to the splitter’s position, a decision their product names record. Anyone watching evening television will hear a run of medications ending in -mab or -nib. Chemists apply the first stem to a monoclonal antibody engineered against a single protein, the second to a small molecule built to inhibit a specific kinase.
Eventually, however, even these naming conventions were not specific enough. In October 2021, the World Health Organization retired the -mab stem and replaced it with four stems because manufacturers had exhausted the available names. Concurrently, the FDA endorsed greater specificity through new drug approvals. Personalized medicines accounted for roughly 36% of new therapeutic molecular entities in 2025, the sixth consecutive year they exceeded a third of new approvals. Therapy is now chosen at the level of a protein, while the diagnosis pointing toward it is still recorded as a bucket.
Compiling the Confusion
AI-software vendors have supercharged the vocabulary problem operationally through ambient documentation. These solutions listen to a visit, infer a diagnosis and produce an ICD-10 code. That code is then treated as the diagnosis because that is the word everyone uses. Independent testers have measured the best general-purpose models at roughly 50% to 59% accuracy on medical coding. Consequently, a model that cannot determine the specific condition selects the broader code, presenting the clinician with a settled answer that is likely less accurate and certainly less informative.
Unfortunately, many vendors also group the alternate phrasings listed beneath a code as synonyms for it, which might be acceptable for billing, but is misleading for care. Several of those phrasings name distinct diseases with different genetics and different treatment. With only the category recorded, every downstream clinician may need to conduct their own time-consuming and costly investigation to arrive at a real diagnosis.
Writing the Distinction Into Systems
On a macro-level, fixing the problem must involve everyone in healthcare consistently differentiating ICD-10 codes as billing categories, both when speaking and writing. On a micro-level, software analysts should also distinguish the billing category and the clinical diagnosis as separate data elements. Similarly, evaluation teams should ensure a clinically specific diagnosis exists alongside the code captured in a way that a downstream platform can access and interpret it.
Until then, however, an evidence-based clinical knowledge layer can be integrated into electronic health record systems to import the specific diagnosis in the free text as a computable concept already mapped to its billing category, so both are recorded.
As a result, reimbursement is unaffected, but, more importantly, the next clinician then sees the clinical explanation alongside the code. So, in the past, when G60.0 (hereditary motor and sensory neuropathy) alone reported almost nothing, the additional detail now helps the clinician understand the type of hereditary neuropathy present to inform their decisions, potentially avoiding repetitive testing or ineffective therapies. The patient then receives the right treatment sooner, one that may never have been pursued if the condition remained obscured behind a vague label.
About Jay Anders, MD
Dr. Jay Anders serves as the Chief Medical Officer of Medicomp Systems, where he plays a pivotal role in product development and acts as a liaison to the healthcare community. He champions the use of technology to empower clinicians, rather than hinder them.
