What this module 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. This module answers a deliberately narrow question: 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. We have generated predictions for every practitioner–address pair that appears in at least one current public federal source: Medicare billing files, the aggregated national provider directory, or NPPES. Among these national registries, ~61% of provider address affiliations are predicted to be not-patient-facing.How to read a label
Every practitioner–address pair receives one of five labels. The most suspicious listings are labeledvery_likely_inactive, while the most likely active listings
are labeled very_likely_active. These labels can be used directly to stratify the
most- and least-likely problematic listings in your universe.
To interpret these labels as probabilities, you have to know your universe’s (or
provider directory’s) baseline wrongness — that is, what is the probability that a
random listing is wrong? From there, you can use the table below to estimate the
probability of wrongness associated with a label.
Where the signal comes from
All training data is public.- 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 provides a month-by-month presence history for every NPI-address pair. Crucially, millions of observed departures provide a strong historical baseline of positive 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.
- The national provider registry (NPPES) and related enrollment 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.
- Additional public rosters and disclosures. Federal order-and-referral eligibility, marketplace directories, public payment disclosures with dated addresses, and clinician utilization summaries. Though these data sources are individually weak, together they reveal many independent sources that currently place a practitioner at a location.
Model development
The predictions are built from stacking multiple layers of models and features. 1. Historical departure prediction. 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. During this step, training labels are observed departures from three years of monthly snapshots, so we can train on millions of examples and back-test on any past month. Both models see only public, time-stamped features. 2. Combined training model. The departure model scores and a smaller set of engineered features are combined into a meta-model and trained on a distinct set of labeled affiliations (see How we verify below). 3. Model calibration and prediction. Random samples of location affiliations spanning the entire range of predictions from the model are validated agentically (see How we verify below) to calibrate the model and convert scores to one of 5 labels.How we generate labels to verify and calibrate the model
We use agents to scour the web for hundreds of individual pairs to create labels and calibrate the final model. These agents leverage provider directories, LinkedIn profiles, obituaries, provider profiles and more. Prompts are designed to be blinded from the model score and report all evidence for and against a provider listing being active. Confirmations require at least one source independent of the registries and directory vendors. In our internal tests agentic verification is able to reach a confident determination for 90% of listings. This approach is not scalable to millions of provider-location pairs. Examples of agentic outputs: 1. A clear departure: psychologist in New Hampshire.Missing addresses
We currently host predictions for ~1-2 unique addresses per NPI, but there are many addresses in other directories (such as plan directories) 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
We estimate about 1% of active provider listings go stale every month (that’s 12.6% compounded in a year). Several of the model’s input data sources are updated every month, so we regenerate and publish fresh predictions every month.Other limitations
- Absence of billing is not absence. Hospital-based clinicians, clinicians in team-based or facility-billed settings, those who have opted out of Medicare, and many behavioral-health clinicians in cash-pay practice are invisible to billing data by design. These groups generate most of our false alarms.
- NPPES-only affiliations are hard to predict. Affiliations asserted only by
NPPES have no billing history to learn from; our ranking carries real but weaker
signal there; most of that population is labeled
uncertain. More than half of its confirmed-wrong affiliations were never a place patients could go, which is a different failure than a clinician who left. - 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: two suites at one street address share a label, and 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
It is not a credentialing or licensure check, not a network-participation check, and not a prediction of who will leave next year. It does not decide whether a listing should be removed, but rather highlights which listings deserve human attention before they are used to support a patient. Avery_likely_inactive label is a
precisely measured prediction, not a verdict; a very_likely_active label is strong
agreement, not a guarantee. A human is needed for the final judgment call.