The directory has a place to record what a doctor does at each hospital. It is empty 68% of the time.

by Eugene Vestel · 2026-08-28 · exploratory, not a pre-registered finding

A question came up in the CMS National Directory of Healthcare community channel this week. A physician can be an internist at one organization and practise a subspecialty at another, and NPPES flags only one taxonomy as primary. So which specialty should a directory publish, and whose definition of primary governs the record a patient eventually sees?

The question is usually argued from first principles. It is answerable from the data instead, because FHIR already models the thing being asked for and the 2026-08-20 NDH release is sitting in a warehouse. PractitionerRole.specialty is scoped to one practitioner at one organization, and its cardinality is 0..*. A practitioner can carry a different specialty at each place they work, and more than one at the same place.

Three numbers, in the order they matter.

This is exploratory and it is not a finding. Every AINPI finding registers its hypothesis before the numbers exist. This one followed a question that was already on the table. Publishing it as a finding would spend the credibility that discipline exists to earn. The compute script is analysis/explore_specialty_context.py and the data is at /api/v1/exploratory/specialty-context.json.

Multi-specialty is real, and NPPES is not the part that is broken

1,334,111 of 7,139,831 individual NPIs in NPPES carry more than one taxonomy. That is 18.7%. 336,504 carry three or more, and the maximum is the full 15 slots.

Every one of those NPIs has exactly one slot flagged primary. Not zero, not two. The primary switch is internally consistent across the whole file, so whatever goes wrong downstream is not NPPES contradicting itself. It is consumers reading the file, and one way of reading it is worth naming: slot 1 is not the primary slot in 14.97% of rows. Anyone treating the first taxonomy as primary is wrong for about one provider in seven.

The NDH has the same property, which I only established by re-running this against the whole array rather than the first entry. The directory carries the NPPES taxonomy set rather than a choice from it: 100.0% of practitioners hold a qualification matching the NPPES true primary somewhere in the array. Read qualification[0] alone and that falls to 88.7%. So the NDH is not disagreeing with NPPES about primacy. It is declining to encode primacy at all, because position does not mean anything in either file. Both figures are in the NPI and taxonomy correctness finding.

That is worth sitting with, because it is the whole difficulty of expressing this data in a flat one-taxonomy-per-provider format. The set is there. Which member of it is primary is not.

The NDH does use the context-scoped field, for a minority

Of the practitioners holding at least one active role that carries both an organization and a specialty, 131,539 carry more than one specialty at a single organization. The record maximum is 16.

Of the 445,527 practitioners the directory places at two or more organizations, 82,977 carry a different specialty at differently-named ones. That is 18.6%. Counting organization records rather than names it is 89,061, and the 6,084 between the two figures are practitioners whose specialty differs only across duplicate records for the same organization. The stricter number is the one to quote. The rest carry the same specialty everywhere.

You can look at them. A thousand of these cases are browsable at /exploratory/specialty-by-organization, searchable by name, NPI, organization or specialty.

So the model is not collapsing everyone to one provider-level code. One in five multi-organization practitioners has a genuinely context-scoped record today. What that 18.6% cannot tell you is whether it is the true rate of context variation or only the rate at which somebody bothered to record it. Those are different claims and this measurement does not separate them.

The field is empty far more often than it disagrees

11,269,522 of 16,545,158 PractitionerRole records carry no specialty element at all. That is 68.1%.

Counted per person rather than per record it is starker. 7,373,232 practitioners are active in the release. 2,315,046 of them carry any role at all. Only 1,193,395, or 16.2%, have a role-level specialty anywhere in the directory.

For the other 84% there is no context-scoped answer to interpret, so a consumer falls back to the provider-level code. Not because it chose to flatten the record, but because it is the only code present. The argument about which taxonomy should be primary is being had about a field that is populated for one practitioner in six.

What follows

The lever with the most reach here is not settling which code is primary. It is requiring a specialty on any published role.

Be precise about how far that goes. A floor on PractitionerRole.specialty reaches the 2,315,046 practitioners the directory already affiliates, which is 31% of the active set. It would roughly double the population with a context-scoped specialty. It does nothing for the 69% carrying no role at all, and that gap is larger and is a different problem, measured in the role gap composition.

Three notes on the numbers above. They are read from the stored FHIR records rather than from the flattened _specialty_code column this project also keeps, which holds the first specialty per record only. An earlier pass through that column returned 52,898 for the varies-by-organization count instead of 89,061, and the same shortcut sat behind three figures in the taxonomy finding, which have now been recomputed across the full arrays and republished. Roles with no organization reference are excluded from the variation counts, because whether a specialty varies by organization is undefined without one. And the NPPES multiplicity figures come from the BigQuery public snapshot, whose newest enumeration is 2026-02-07, so they describe the shape of taxonomy multiplicity rather than its current level.