Methodology

MedicaidWatch uses five composable detection algorithms to scan every provider in the dataset. Each asks a different question about billing behavior, and together they produce a composite risk score.

The Data

Our analysis uses the HHS Medicaid Provider Spending Dataset, the largest provider-level Medicaid dataset ever released publicly.

Records: 227,083,361 claims
Period: Jan 2018 – Dec 2024
Total spending: $1.09 trillion
Billing providers: 617,000+
Servicing providers: 1.6 million+
Columns: NPI, HCPCS, MONTH, BENES, CLAIMS, PAID

The Five Algorithms

1. Beneficiary Intensity

Weight: 20% · 4,391 providers flagged

Question: Are they billing too many services per patient?

We calculate the claims-per-beneficiary ratio for every provider for each HCPCS procedure code, then compare it to the national median for that code. Providers are flagged when their ratio exceeds 5x the median and they've billed more than $100K.

The most extreme outlier billed at 79x the median — 235 claims per beneficiary where peers averaged 3.2.

2. Temporal Anomaly Detection

Weight: 15% · 959 providers flagged

Question: Did something suddenly change?

Z-score analysis of each provider's monthly billing against their own historical mean and standard deviation. A spike is flagged when the z-score exceeds 5 and the spike month exceeds $100K.

The largest spike detected was 27.9x a provider's normal billing — $1.26M in a single month versus a $45K average.

3. Code Concentration

Weight: 15% · 22,376 providers flagged

Question: Is this a billing mill?

We measure what percentage of a provider's total revenue comes from a single HCPCS code. Providers are flagged when more than 90% of revenue is from one code and the total exceeds $1M.

Some flagged providers derive 100% of their revenue from a single procedure.

4. Billing Network Analysis

Weight: 15% · 7,357 providers flagged

Question: Is this a shell entity routing claims?

We count distinct servicing providers per billing entity and map network relationships. Entities are flagged when they route claims through more than 50 external servicing providers and bill more than $1M.

Large hospital systems legitimately route through thousands of providers — context is essential. The largest network connected 5,715 providers to a single billing entity.

5. Peer Deviation

Weight: 15% · 22,947 providers flagged

Question: Are they charging far more than peers?

We compare every provider's cost-per-claim against the national median for the same HCPCS code. Providers are flagged when their cost exceeds 3x the median and their total billing exceeds $500K.

The most extreme outlier charged 1,264x the median — $7,409 per claim where peers average $5.86.

Composite Score

Each provider's composite score is the weighted average of all triggered skill scores, normalized to 0–100. Only skills that actually fire for a provider contribute.

COMPOSITE = (
  beneficiary_intensity × 0.20 +
  temporal_anomaly      × 0.15 +
  code_concentration    × 0.15 +
  billing_network       × 0.15 +
  peer_deviation        × 0.15
) / sum(active_weights) × 100

Cross-Reference Enrichment

Every flagged provider is enriched with public data from external sources:

  • NPPES NPI Registry — Provider name, address, specialty, registration date, authorized official.
  • OIG LEIE Exclusion List — Cross-reference against 82,714 excluded individuals and entities. We found 552 matches who billed $395.7M after exclusion.

Important Notes

  1. Anomalies are not accusations. Unusual billing patterns can have legitimate explanations. Large organizations, specialized providers, and certain business models may naturally trigger certain algorithms.
  2. The dataset has limitations. November–December 2024 data is incomplete due to claims processing lag. The dataset covers fee-for-service Medicaid only, not managed care.
  3. AI-generated narratives are analytical, not legal. All narratives should be verified by qualified investigators before any action is taken.
  4. We show our work. Every finding includes exact numbers, peer comparisons, and methodology. Readers can verify every claim against the public dataset.