Skip to main content

​
What this model answers

Patients are often referred to providers based on which clinic locations are closest to the patient’s home. Provider directories are notoriously stale. Clinicians move, retire, consolidate into new groups, or shift to telehealth, and the directories that list them lag behind by months or years. And even a current listing is not always a place a patient can go: many physicians practice at addresses where no one can book them. This model answers two deliberately narrow questions about each practitioner–address pair:
  • Stale affiliation risk (staleAffiliationRisk): is this practitioner actually seeing patients at this address, right now? This can include stale addresses, or addresses that were never a place patients could go at all, such as mailing hubs and billing offices.
  • Bookability (bookability): can a patient book an appointment with this practitioner at this address? A practitioner can be seeing patients at an address that is still not a place to book them: an emergency physician or hospitalist working in a hospital, a surgeon who operates at a hospital but sees consults at a clinic across town, a radiologist reading studies. Bookability does not tell you whether the practitioner is accepting new patients, takes a given plan, or how soon they can be seen.
Both answers come from a single model, so they never contradict each other: a location is never rated more bookable than it is active. We generate predictions for physicians (MD and DO) at every address asserted for them by at least one current public source: Medicare billing files, the aggregated national provider directory, NPPES, and public health plan network directories. Other practitioner types (such as nurse practitioners, physician assistants, and behavioral health clinicians) and physicians in training are not yet covered; their locations are returned without either label. Among physician listings in these sources, only about 30% are places where the physician is currently seeing patients, and fewer than one in five are places a patient could book them.

​
How to read a label

Every covered practitioner–address pair receives one label for each question. Each label stands for a fixed range of probability, and we measure how often each label is right on randomly sampled, independently verified listings. A label is only assigned when the measured rate, including its uncertainty, sits inside the label’s range. uncertain is not a negative signal; it means the evidence supports neither call. These rates are measured across our national set of physician listings. In another directory, such as a health plan’s own directory, the mix of listings is different: the order of the labels holds, but the rates shift with that directory’s baseline. In one plan directory we tested, listings were active more often than the national rates would suggest. To put exact numbers on your own directory, verify a small random sample of it against the labels. Reach out if you’d like help.

​
Where the signal comes from

All model inputs are public data.
  • Medicare’s public clinician files, released monthly. Each monthly release asserts where clinicians are practicing based on recent billing. More than three years of releases provide a month-by-month presence history for every NPI-address pair, including millions of observed departures. These files have several limitations: they only cover clinicians who bill Medicare under their own identity, they lag real events by several months, and roughly one pair in ten that disappears from a release later reappears.
  • Medicare utilization summaries. What a clinician bills (office visits versus hospital visits, procedures performed in facilities, telehealth-only services) says a great deal about whether any of their addresses is a place patients book.
  • The national provider registry (NPPES) and related enrollment and licensure data. Deactivations, license states, practice and mailing addresses, specialty self-descriptions and the shape of the registry record itself: how a building is used across all clinicians registered there, how old the record is, how the clinician attests to it. Registry agreement is weak evidence, but contradictions — such as a deactivated identifier, an address that thousands of clinicians use only for mail, a location on the other side of the country — can provide strong evidence.
  • Public health plan network directories. Marketplace plan directories and state Medicaid network filings: which plans list a practitioner at an address, for how long, and whether they flag the listing as telehealth. These are what payers assert, so we treat them as signals, never as confirmation.
  • Additional public disclosures and geography. Public payment disclosures with dated addresses, how far an address is from where the clinician bills, and how a building is shared among clinicians. Though these data sources are individually weak, together they reveal how many independent sources currently place a practitioner at a location.
We are careful about leakage between time stamps and between datasets. Many directories buy from the same upstream vendors, so we make sure we’re not double-counting corroboration.

​
Model development

1. Billing-history models. From the monthly billing histories we train two models: one estimating whether a practitioner has already left an address, one estimating whether a currently-present pair is about to go absent. Their training labels are observed departures from three years of monthly snapshots, so they learn from millions of examples. Both see only public, time-stamped features, and their scores become inputs to the next step. 2. Location outcome model. A single model, trained on independently verified practitioner–location pairs (see below), estimates what each listing actually is:
  • a place where patients book appointments with the practitioner;
  • a procedure site, such as an operating room or surgery center, where the practitioner performs procedures that are booked elsewhere;
  • a place where the practitioner works but patients cannot book them, such as an emergency department or inpatient unit;
  • a place the practitioner has left, or never practiced;
  • an address that is not a care site at all, such as a billing office or mailing address.
Stale affiliation risk is the probability of the first three outcomes combined; bookability is the probability of the first. 3. Labels and measurement. Each probability is converted to one of the five labels using the fixed ranges above. We then measure every label’s actual rate on randomly sampled verified listings that were not used to make that prediction, and confirm the rates on a fresh random sample before each release. Two kinds of listing are labeled by rule rather than by the model:
  • Inpatient specialties. Physicians whose specialty is practiced inside hospitals — hospitalists, emergency physicians, anesthesiologists, critical care physicians, and radiologists and pathologists without an office practice — are labeled very_likely_not_bookable. When we checked, about 1% of these listings were bookable.
  • Deactivated identifiers. Practitioners whose NPI has been deactivated (confirmed against the live registry) are labeled very_likely_inactive and very_likely_not_bookable.
On verified random listings, the model separates active from inactive listings with an AUC of 0.84, and bookable from not bookable with an AUC of 0.89. On the same listings, our previous stale-location model scored 0.70.

​
How we generate labels to verify and calibrate the model

We use agents to research thousands of randomly sampled practitioner–address pairs. These agents search the places where care is actually delivered and announced: the practice’s and health system’s own websites and staff rosters, hospital rosters, state license records, news and announcements. Directory aggregators and reseller sites often copy the same stale registry data, so they never count as strong evidence on their own. Prompts are blinded to the model score and ask agents to report all evidence for and against the listing, then decide what the address is for this practitioner (one of the outcomes above) and whether an appointment can be booked there. A second, blind read of the same listing agrees more than 95% of the time. About three in four randomly sampled listings can be decided with confidence; listings that can’t be decided are left out rather than guessed. This approach is not scalable to millions of practitioner–location pairs, which is why we train a model on it. Examples of agentic outputs: 1. A clear departure: sleep medicine physician, Texas.
2. Right practice, wrong clinician: neurologist, Kentucky.
3. Seeing patients, but not bookable: urologic surgeon, Missouri.
4. Bookable, but not taking new patients: family physician, Wisconsin.

​
Missing addresses

We currently host predictions for ~1-2 unique addresses per NPI, but there are many addresses in other directories (such as a health plan’s own directory) that we do not provide predictions for. We can generate predictions for addresses outside the scope of what is pre-computed. Reach out to learn more.

​
Ongoing updates

Directories decay continuously: practitioners leave, retire, and change how they practice every month. As our input sources publish new releases, we regenerate our predictions and re-measure every label against a fresh random sample of verified listings before publishing. Each response reports the model version that produced its labels in servedModels.

​
Other limitations

  • A group’s address is not always a clinician’s address. Public sources often list a group’s headquarters, billing office, or flagship clinic for every clinician in the group. Telling the satellite office where a clinician actually sees patients apart from the group’s other sites is the largest remaining source of error.
  • Absence of billing is not absence. Clinicians in team-based or facility-billed settings, those who have opted out of Medicare, and direct primary care or membership practices that don’t bill insurance are partly invisible to billing data. We tend to under-call active listings for opted-out and direct primary care physicians.
  • Single-source listings are hard to predict. Listings asserted only by NPPES, or only by a state Medicaid network, have no billing history to learn from. Listings that appear only in a state Medicaid network receive the label measured for that group as a whole rather than an individual prediction.
  • Inpatient specialties are labeled by rule. A hospitalist or emergency physician who also keeps a small outpatient clinic will still be labeled very_likely_not_bookable there.
  • Multi-site practice. A clinician who works one afternoon a week at a satellite office may show no public footprint there. Some of our suspect pairs are real-but-rare practice locations, and even human verification sometimes cannot distinguish “never there” from “rarely there.”
  • Telehealth blurs “location.” Clinicians attached to virtual-first practices are often listed at physical addresses they have never visited. We flag structurally non-clinical addresses, but a telehealth clinician listed at a real clinic they never enter is difficult to confirm.
  • The public record lags. Our primary sources update monthly to annually. A clinician who moved last month will usually not be caught yet; a pair we corroborate today may have quietly broken.
  • Identity confusion. Two clinicians with the same name are a recurring source of both directory errors and verification errors. Same-name confusion is one of the most common mechanisms behind confirmed-wrong location affiliations, and we treat name-only matches as weak evidence everywhere in the pipeline.
  • Address matching is imperfect. We normalize addresses to the building level and pool much of our evidence per building. Campuses, suites, and creative address formatting mean a small share of pairs are compared against the wrong building. We made this choice because we believe a wrong suite is less likely to be problematic, and it makes it possible to aggregate more evidence per address.

​
What these predictions are not

These predictions are not a credentialing or licensure check, not a network-participation check, and not a prediction of who will leave next year. Bookability is not a statement that the practitioner is accepting new patients or about how soon they can be seen. These predictions do not decide whether a listing should be removed, but rather highlight which listings deserve human attention before they are used to support a patient. A very_likely_inactive label is a precisely measured prediction, not a verdict; a very_likely_active or very_likely_bookable label is strong agreement, not a guarantee. A human is needed for the final judgment call.