About OpenMedicare

OpenMedicare is an independent data journalism project dedicated to making Medicare physician spending transparent and accessible. We analyze over a decade of Medicare payment data (2014-2023) to help patients, policymakers, and researchers understand how taxpayer dollars flow through the healthcare system.

Our goal is simple: shine light on Medicare spending patterns to promote accountability, identify potential fraud, and help Americans make informed healthcare decisions.

Most Medicare transparency sites rely on pre-aggregated summaries or filtered snapshots. We analyze the raw CMS provider-level data — over 96 million rows spanning 10 years — to build our own indexes, risk scores, and trend analysis from the ground up. This means we catch patterns that summary data obscures: billing anomalies within specialties, geographic outliers, providers whose volumes defy mathematical possibility, and markup patterns that only emerge when you compare individual providers against their peers.

Data Sources

• Medicare Provider Utilization and Payment Data: Annual datasets released by CMS containing detailed payment information for individual healthcare providers
• Medicare Part B National Summary Data: Aggregate spending data across all Medicare Part B services
• Provider enrollment data: Physician credentials, specialties, and practice locations from CMS databases
• Geographic data: Rural-urban classifications and state-level healthcare statistics

Analysis Framework

Our analysis focuses on several key areas:

• Markup Analysis: Comparing submitted charges to actual Medicare payments to identify billing patterns
• Peer Comparison: Benchmarking providers against others in similar specialties and geographic areas
• Trend Analysis: Tracking spending changes over the 10-year dataset to identify patterns
• Risk Scoring: Our composite risk scores combine four signals: statistical deviation from specialty peers, markup ratio analysis, volume anomalies (services per day), and beneficiary concentration patterns. Providers are flagged when multiple indicators exceed peer-adjusted thresholds.
• Geographic Analysis: Examining rural vs urban spending patterns and state-level variations

Limitations & Disclaimers

Important limitations to understand when using our data:

• Data reflects Medicare Part B payments only (physician services), not hospital or prescription drug coverage
• Providers must receive at least $10,000 annually to appear in CMS public datasets
• High payment amounts may reflect legitimate factors like patient complexity or specialty care
• Statistical outliers require further investigation to determine if fraud or abuse occurred
• Data processing involves cleaning and standardization that may introduce minor variations

Beyond statistical outlier detection, we built a supervised machine learning model (Random Forest classifier) trained on real confirmed fraudsters. Our training labels come from two sources:

• HHS-OIG LEIE Exclusions: Providers excluded from federal healthcare programs for fraud, abuse, or other misconduct
• DOJ Criminal Charges: Providers indicted or convicted in federal Medicare fraud cases

Combined, these sources give us 2,198 confirmed positive labels for training. The model learns the billing patterns these fraudsters share — volume anomalies, markup ratios, specialty concentration, geographic signals — and scores all 1,719,625 active Medicare providers on how closely they resemble confirmed criminals.

Results: The model achieves an AUC of 0.83, meaning it correctly ranks a random fraudster above a random legitimate provider 83% of the time. 500 providers scored above the 86% match threshold and are featured on our "Still Out There" page.

As with all our analysis, a high ML score is not an accusation — it identifies providers whose billing patterns statistically resemble those of confirmed criminals, warranting further review.