Hospital price transparency data all comes from one place: the machine-readable files hospitals are required to publish under 45 CFR Part 180. The files are free. Parsing them is not hard. What separates a usable dataset from a large one is knowing which of those numbers are real — and a meaningful share of them are not.
When a hospital's billing vendor has no price to publish for a line, it does not leave the field empty. It writes a number. These are the conventions, measured across all 6,113 facilities in our corpus:
| Value published | Rows | Hospitals | What it means |
|---|---|---|---|
| $0.01 | 28,095 | 1,347 | no price available |
| $0.70 | 13,393 | 337 | vendor default |
| $0.00 | 13,141 | 89 | no price available |
| $0.35 | 3,860 | 261 | vendor default |
| $999,999,999 | 1,402 | 23 | max-integer, "not applicable" |
These are syntactically valid numbers. A parser accepts them, a schema validates them, and a database stores them. Nothing about the ingest fails.
Placeholders are not randomly distributed — they cluster at the extreme ends of the price distribution. That makes them maximally damaging to exactly the statistics a rate-benchmarking team depends on:
CPT 27447, total knee replacement. Same source files, two ways of handling them:
| Approach | Reported figure |
|---|---|
| Ingest as published, report the mean | $1,866,246 |
| Detect placeholders, exclude, report the median | $9,069 |
The first number is not a typo and not a rounding difference. It is what the arithmetic produces when a handful of max-integer rows are treated as prices. A knee replacement does not cost $1.8 million, and any benchmark built on that figure is unusable — but nothing in the pipeline errors.
A crude filter — drop anything implausible — destroys real data. Gene and cell therapies genuinely bill in eight figures, and they are in these files:
| Code | Therapy | Published charge | Verdict |
|---|---|---|---|
J3399 | onasemnogene abeparvovec (Zolgensma) | $20,072,246 | real — retained |
J3394 | lovotibeglogene autotemcel (Lyfgenia) | $16,577,889 | real — retained |
J3392 | exagamglogene autotemcel (Casgevy) | $11,764,953 | real — retained |
Q2055 | idecabtagene vicleucel (CAR-T) | $10,000,000 | real — retained |
MS075 | viral meningitis w/ CC-MCC | $999,999,999 | sentinel — excluded |
A $5M cut-off looks sensible and silently deletes every one of the real therapies above. Ours sits at $99M — above every genuine charge line we have observed, below every max-integer convention. Detection is relative to the peer distribution for the same code, with an absolute ceiling as a backstop.
Not a claim — a file. Every flagged row, by hospital, code, published value, and the peer median it was measured against. If you are already licensing this data from someone, run it against what you have and see what comes back.
4,077 facilities scored on whether they meet 45 CFR Part 180 at all: 427 publish no usable file, 3,650 publish a file containing placeholders, 212 are majority unusable. Each with a phone number.
6,113 facilities across 55 states and territories — including Puerto Rico, the US Virgin Islands, Guam and American Samoa, which most corpora stop short of. Plus 59,618 clinic and surgery-center locations and 1,102,404 individual practitioners alongside the hospital data.
Send us a state and a handful of codes you already have rates for. We will return our figures beside the exclusion list for those codes, so you can see exactly which rows differ and why. No commitment, and you keep the comparison.
Figures on this page are computed from our own corpus of hospital machine-readable files and are reproducible from the exclusion list. They describe what is in the source files published by hospitals — they are not a measurement of any other vendor's product.