Datasetpublic

Nason/cms-medicare-database

CMS Medicare Physician & Other Supplier Database A clean, queryable DuckDB database built from the CMS Medicare Physician & Other Practitioners Public Use Files -- provider-level Medicare Part B claims data from CY2012 through CY2023. 132,947,347 rows across 3 tables covering what every physician billed, what Medicare paid, and how many services and beneficiaries per NPI per HCPCS code. Built with cms-medicare-database. Quick Start DuckDB CLI… See the full description on the dataset page: https://huggingface.co/datasets/Nason/cms-medicare-database.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

0likes180downloads

Dataset Card

CMS Medicare Physician & Other Supplier Database

A clean, queryable DuckDB database built from the CMS Medicare Physician & Other Practitioners Public Use Files -- provider-level Medicare Part B claims data from CY2012 through CY2023.

132,947,347 rows across 3 tables covering what every physician billed, what Medicare paid, and how many services and beneficiaries per NPI per HCPCS code.

Built with cms-medicare-database.

Quick Start

DuckDB CLI

sql

INSTALL httpfs;
LOAD httpfs;
ATTACH 'https://huggingface.co/datasets/Nason/cms-medicare-database/resolve/main/cms_medicare.duckdb' AS cms (READ_ONLY);

-- Total Medicare spending by year
SELECT year, ROUND(SUM(line_srvc_cnt * avg_medicare_payment_amt) / 1e9, 2) AS total_spending_billions
FROM cms.physician_services
GROUP BY year ORDER BY year;

Python

python

import duckdb
con = duckdb.connect()
con.sql("INSTALL httpfs; LOAD httpfs;")
con.sql("""
    ATTACH 'https://huggingface.co/datasets/Nason/cms-medicare-database/resolve/main/cms_medicare.duckdb'
    AS cms (READ_ONLY)
""")
con.sql("SELECT * FROM cms.physician_services LIMIT 5").show()

DuckDB uses HTTP range requests, so only the pages needed for your query are downloaded.

Tables

Table	Description	Rows
`physician_services`	Provider-level Medicare Part B claims: one row per NPI per HCPCS code per place	116,190,383
`physician_summary`	Provider-level aggregate summary: one row per NPI per year with total services,	13,528,933
`geography_service`	Geographic aggregate: Medicare utilization and payment by state/national level,	3,228,031

Data Source

CMS Medicare Physician & Other Practitioners PUF -- maintained by CMS. Updated annually. Public domain U.S. government data. HCPCS descriptions include AMA CPT content used under CMS license.

License

Database build code: MIT. Underlying data: public domain (U.S. government work). Note: HCPCS descriptions contain AMA CPT content included as provided by CMS in the PUF.

GitHub

Full source code, build instructions, and data dictionary: github.com/ian-nason/cms-medicare-database

Changelog

2026-07-06 — Full refresh + data-quality audit

Rebuilt from the current CMS PUF releases, adding CY2024, and repaired after an independent SQL-verified audit.

Data changes

—CY2024 added to all three tables: 132,947,347 rows total (physicianservices 116.2M, physiciansummary 13.5M, geography_service 3.2M), CY2013-CY2024.

Fixes

—geography_service.HCPCS_Cd re-ingested as VARCHAR: numeric auto-casting had NULLed all 282,776 alphanumeric codes (G/J/lab codes) and stripped leading zeros from 137,781 anesthesia codes. Geography-level procedure analysis works again, including joins to physician_services.hcpcs_code.
—physician_summary identifiers re-typed as VARCHAR: Rndrng_NPI (joins to services now work), Rndrng_Prvdr_Zip5 (1.35M leading-zero ZIPs restored), State_FIPS, RUCA.

Known caveats (see README for the full list)

—CMS suppression makes physician_services a systematic ~17-21% payment undercount vs physician_summary — use the summary table for provider totals.
—provider_gender is 100% NULL in every year (removed retroactively by CMS); chronic-condition columns are NULL before 2017 and top-coded at 75.
—geography_service mixes National and State rows — filter Rndrng_Prvdr_Geo_Lvl.