Pharmacy Fraud in Medicare: How We Detect It

Share:
🤖

Machine learning predictions identify statistical patterns consistent with confirmed fraud cases. A high ML score does NOT mean a provider is committing fraud — it means their prescribing patterns statistically resemble those of providers who were.

Read our methodology →

Medicare fraud costs taxpayers an estimated $60 billion per year. We built a multi-layered detection system that combines statistical analysis, peer comparison, and machine learning to identify prescribing patterns consistent with fraud — all from public CMS data.

1,077,354

Providers Scored by ML

4,183

ML-Flagged Providers

6,706

High-Risk (Rule-Based)

372

OIG-Excluded Active

Our Three-Layer Detection Approach

Layer 1: Rule-Based Risk Scoring

Our 10-component risk scoring model evaluates every provider against their specialty peers. Components include:

  • Opioid prescribing rate vs. specialty median (z-score)
  • Population-level opioid percentile
  • Cost outlier detection (per-beneficiary)
  • Brand preference vs. peers
  • Long-acting opioid rates
  • Antipsychotic prescribing patterns
  • Dangerous drug combinations (opioid+benzo)
  • OIG exclusion list matching
  • Low drug diversity (few unique drugs, many claims)
  • Frequent fill patterns

Most Common Risk Flags

FlagProviders
extreme_opioid_vs_peers3,581
high_la_opioid_vs_peers3,484
95th_pctile_opioid3,188
very_high_opioid_vs_peers2,341
99th_pctile_opioid1,360
90th_pctile_opioid1,073
elevated_la_opioid845
high_antipsych_elderly676
opioid_benzo_coprescriber523
extreme_fills_per_patient469

Layer 2: Machine Learning

We trained a BaggedDecisionTrees model using 281 confirmed fraud labels from OIG enforcement actions. The model uses 20 features to identify prescribing patterns that resemble known fraud cases.

Model Performance

83.0%

Precision

66.6%

Recall (CV)

73.8%

F1 Score

Cross-validated on held-out fraud labels. Production recall on full LEIE match set: 76.5%.

The model flagged 4,183 providers out of 1,077,354 scored — a flag rate of 0.39%. This is intentionally conservative: high precision minimizes false positives.

Layer 3: OIG Exclusion Cross-Reference

We match all prescribers against the OIG List of Excluded Individuals and Entities (LEIE). Result: 372 excluded providers appear in active 2023 prescribing data. These providers are prohibited by federal law from billing Medicare.

Top ML-Flagged Providers

Providers with the highest machine learning fraud probability scores:

ProviderSpecialtyStateML ScoreCostOpioid %
Jerry FlynnFamily PracticeAK1.00$374K6.3%
Milla KarevGeriatric MedicineOH1.00$3.5M1.8%
Marcel HaulardFamily PracticeMO1.00$649K8.1%
Herman ChavisFamily PracticeNC1.00$1.8M10.4%
Jason HamFamily PracticeAL1.00$4.0M8.8%
Joseph ScottFamily PracticeTX1.00$956K10.4%
Vadim BaramPsychiatryMO1.00$9.3M0.0%
Dumitru SandulescuInternal MedicineMI1.00$1.1M9.4%
Jackie MaxeyFamily PracticeKY1.00$3.4M6.7%
Jon BowenFamily PracticeWV1.00$1.4M5.3%

Case Studies: Dual-Flagged Providers

Providers flagged by both our rule-based system (high risk score) and the ML model represent the strongest signals. Here are examples:

Suzanne Morse

Nurse PractitionerCottonwood, AZ

Risk: 64
Claims: 5,441
Cost: $304K
Opioid: 61.7%
Excluded: No
extreme_opioid_vs_peers95th_pctile_opioidhigh_la_opioid_vs_peerselevated_antipsych_elderlyopioid_benzo_coprescriber

How Pharmacy Fraud Works

Common pharmacy fraud schemes that our system can detect signals of:

Phantom Billing

Billing for prescriptions never dispensed. Signal: high claims volume with few unique beneficiaries.

🔄 Upcoding

Substituting cheaper generics but billing for brand. Signal: abnormally high brand percentage vs peers.

🏭 Pill Mill Operations

High-volume controlled substance dispensing. Signal: extreme opioid rates, low drug diversity.

🚫 Excluded Provider Billing

Billing through another provider's credentials. Signal: LEIE match with active claims.

Limitations & Ethics

Our detection system identifies statistical patterns, not confirmed fraud. Important caveats:

  • A high risk score or ML flag is not an accusation — many legitimate reasons can explain outlier patterns
  • Pain management, oncology, and palliative care providers will naturally score higher on opioid metrics
  • Rural providers may appear as cost outliers due to smaller patient panels
  • Only law enforcement with full claims data can determine actual fraud

Our Goal

We aim to make public data accessible for transparency. Every data point comes from CMS public use files. We provide analysis — not accusations. If you believe our data contains errors, contact us.

Related Analysis

View all analysis →