How Data Scientists Use Health Care Facilities Data

Three facts frame every design decision below. First, the deepest time series ended on purpose - HHS stopped collecting hospital capacity data after May 3, 2024. Second, the widest table is county-level: the AHRF facilities file alone carries 716 columns. Third, only one source here enumerates providers individually - the NPI Registry described below - and it is the join key for everything else.

HealthData.gov - COVID-19 Hospital Capacity by Facility (Historical Time Series) is the only facility-week utilization panel in the slice, and structurally it behaves like a benchmark ML dataset: roughly 6,000 hospitals across about 230 collection weeks from January 1, 2020 through May 3, 2024, several million rows total, each carrying 127 columns. Identifiers pin every row to a facility - hospital_pk matching the CCN where one exists, plus name, address, fips_code, hospital_subtype and an is_metro_micro flag - while measures arrive with three suffixes that tell you how to aggregate them: _avg for 7-day averages, _sum for weekly totals, _coverage for how many times the facility reported that element during the week.

Two caveats belong in your feature-store docs. No statistical adjustment was made for non-reporting or missing facilities - values reflect only reports received - so the _coverage columns double as a per-field missingness indicator you get for free. And the series is static: reporting ended May 3, 2024, ideal for retrospective surge studies, useless as a live signal.

Two sibling file families extend it below the county line. Health Center Service Delivery and Look-Alike Sites ships one row per geocoded site across 56 columns, including FQHC Site NPI Number - your bridge to the provider master described later. HPSA and MUA/P designation files add shortage geometry: primary care HPSA detail runs about 48 MB, MUA/P detail about 11 MB, with polygon boundaries and centroid points in SHP form. The interactive layer of the same warehouse is cataloged separately as HRSA Data Warehouse - Health Center & Facility Finders.

Everything lands over plain HTTP with no registration under public-domain terms whose cards read "Usage limitations: None". One caveat: UDS-derived files carry a FOIA notice stating any alteration or format conversion is the user's sole responsibility.

Yes, and it is the quiet advantage of this industry: standardized annual reporting for the same organizational units going back to 1999. The HRSA Health Center Program Uniform Data System (UDS) Data 2025 release centers on the H80 awardee workbook - about 28.3 MB of XLSX across 38 sheets covering 1,356 health centers, with Tables 3A through 9E holding patients, visits, staffing, quality measures and financials. A look-alike workbook (~2.9 MB) and a service-delivery-sites CSV (~12.8 MB, 56 columns) round out the set, all commercial delivery terms.

Feature-engineering notes that save a day: measure columns arrive coded, T3a_L1_Ca-style, keyed to the UDS Tables manual rather than self-documenting, so build a code dictionary before writing transforms. Blank cells can mean suppression rather than zero - edit checks are logged in the Edits sheet. And the featured year lags: the 2025 Coversheet stamps DateOfLastReportRefresh 05/07/2026.

Respect the method condition CMS attaches: no charge, but per-hour limits apply and bulk queries must use the DDS file instead of the API. The API caps requests at 200 results with skip limited to 1,000 - roughly 1,200 records over six calls - marking where interactive lookups end and batch ingestion begins.

For geometry, the companion portal is HRSA Data Warehouse - Health Center & Facility Finders: an unauthenticated JSON locator takes lat/lon/radius queries, registered web services need a free token tied to a declared calling domain, and facility locator records refresh daily - same-day record dates were observed at research time. A practical spine: land the monthly NPPES file as master, sync HRSA points daily, then left-join AHRF county features and UDS measures on NPI and site name, diffing basic.last_updated for change detection.

A numbered walkthrough from empty cluster to joined analytical table:

Every step runs on free, public-domain inputs - the stack fits a research budget of zero.

Ranked by Datadory's quality score as of August 2026, weighted toward what a modeling workflow can ingest without procurement: