The Site of Care problem

NPIs are on every medical claim submitted from providers to payers for reimbursement. Those NPIs are tied to physical addresses, and are widely used to determine where the visit documented in the claim took place.

The problem is that NPIs were never designed to indicate where care was provided. CMS describes an NPI as an “intelligence-free” number, which means it carries no information about a provider's location by design. It exists so claims can move between providers and payers without needing multiple identification numbers for the same provider.

There are two flavors of NPI. Type 1 applies to an individual, like a specific physician or nurse. Type 2 belongs to an organization.

A single organization can hold many Type 2 NPIs for different divisions, locations, and service lines, and the address registered to any one of them is the address submitted in documentation sent to NPPES (administered by CMS). That can often be a corporate office, a billing department, or a health system's main campus.

In short:

How billing data became location data

When a physician bills through a larger organization (think physicians’ group or health system), the claim may be submitted using the group's registered NPI, even if the procedure occurred 10 miles away. Large systems can often route billing through centralized entities, so the volume from a dozen real provider locations collapses onto a single, incorrect address.

If you run a provider directory or a claims-based intelligence tool on top of that data, it’ll inherit every one of those distortions.

Market sizing in healthcare can also rest on patient/diagnosis/procedure volumes, as well as on whether care occurs in hospitals, ASCs, or offices. That segmentation drives where a company invests, which accounts it prioritizes, and how it forecasts. 

MedTech companies are using claims-based data to guide territory design, quota setting, account targeting, market access strategy, and resource allocation. If procedures are attributed to a single billing address despite occurring across many locations, teams may overestimate opportunity in one territory, underinvest in another, focus contracting efforts on the wrong organizations, or overlook important accounts entirely.

The impact becomes even more obvious in the field. Healthcare remains a relationship-driven industry where commercial teams rely on in-person engagement. A field rep may be directed to the listed address of a high-value physician, only to arrive at an office the physician has never set foot in.

Once it's discovered that these locations don’t match where the patient was seen, confidence in the underlying data erodes, and every decision built on it becomes harder to make with conviction.

The Site of Care model

The idea is simple, even if the work underneath it isn't. Instead of trusting the billing NPI on a claim, we’re rolling those billing identifiers up into a single real location and attributing providers to the places where they actually practice.

So, yes, we still use the NPI. But building nuance and practical thinking into the model helps contextualize what the NPI is telling us.

A few things happening under the hood:

• Consolidation: Type 2 NPIs for the same location are grouped using standardized address, name, taxonomy, and CMS enrollment. NPIs on the same hospital campus are grouped by a general catchment area, even when the address lines don't match exactly, whereas non-hospital sites must share a standardized “rooftop” address.
  
  
• Deduplication: A single service is usually billed twice, once on the institutional (facility) form and once on the professional (provider) form, with different claim IDs. We count the service once and map it to the most accurate location instead of double-counting.
  
  
• De-obfuscation: Centralized billing is unwound to the true sites of care, but only when the data model has a high level of confidence in the real location. There’s some proprietary data engineering going on here.
  
  
• Re-Classification: Every type of site gets a classification that better reflects how organizations actually view the market: ASC, Hospital, Lab, Urgent Care, Office, etc. Hospitals have inpatient and outpatient segments, so volume in an outpatient surgery department can be viewed separately from that in the inpatient setting.

We’re figuring it out, but there is more to do

Modularity is hot in the streets, and you can't track demand if your data can't distinguish a hospital outpatient department from the ASC across the street. Getting the location right is the prerequisite for every site-of-care question worth asking.

The NPI problem isn’t fully solved, and it probably never will be until we drop the NPI of it all. And we're still deliberately conservative about reassigning a claim's location when the signal isn't strong enough to warrant it.

It’s important to be transparent that no data model will perfectly reconstruct reality. But every improvement in accuracy increases trust in the data and provides a clearer picture of where care is delivered.

‍