In Post 1, we looked at what the Medicaid spending file contains and what it leaves out. The next question follows naturally: to build a classifier that screens providers for fraud, we need labeled training data. Which providers committed fraud? Which ones stayed clean? Any supervised machine learning (ML) approach starts here.
A common candidate for “known fraudulent” labels is the Office of Inspector General’s (OIG) List of Excluded Individuals and Entities (LEIE). The federal government updates this registry monthly, and published fraud-detection studies often use it as a positive label.
The problem: the LEIE lists excluded providers, and exclusion covers far more ground than fraud.
What Gets a Provider Excluded
The LEIE operates under two sets of authorities from Section 1128 of the Social Security Act. Mandatory exclusions under Section 1128(a) carry a minimum five-year ban:
| Code | Description |
|---|---|
| 1128(a)(1) | Conviction of a program-related crime |
| 1128(a)(2) | Conviction related to patient abuse or neglect |
| 1128(a)(3) | Felony conviction related to healthcare fraud |
| 1128(a)(4) | Felony conviction related to controlled substances |
These are the exclusions most people think of when they hear “LEIE.” Providers who stole from Medicare. Providers who abused patients. Providers convicted of fraud felonies.
But there’s a second set. Permissive exclusions under Section 1128(b) cover a wider range of conduct:
| Code | Description |
|---|---|
| 1128(b)(4) | License revocation, suspension, or surrender |
| 1128(b)(7) | Fraud, kickbacks, and other prohibited activities |
| 1128(b)(14) | Default on a health education loan |
The article reports that the single largest LEIE category is 1128(b)(4): license revocation or suspension, with more than 32,000 entries and roughly 40% of exclusions. Those counts are not publicly reproduced here. The public taxonomy confirms that exclusion reasons extend beyond fraud convictions.
Labeling every listed provider as fraudulent therefore changes the target from fraud to exclusion. A classifier trained on that target estimates a mixture of statutory reasons.
What This Does to a Classifier
To see why this matters, consider what relationship a fraud classifier is estimating. It takes billing features as input (how much a provider bills, how many patients they see, how concentrated their procedure codes are) and predicts whether the provider is fraudulent.
If the training labels include thousands of providers who lost licenses for non-fraud reasons, the model no longer estimates “what does fraud look like in billing data?” Instead, it estimates “what does any kind of professional trouble look like in billing data?” Those two questions have different answers.
Observation windows differ mechanically when exclusion occurs during the 2018 to 2024 panel. The fitted relationship can therefore reflect panel length alongside any billing pattern unless the design uses rates or a common observation window.
The article reports that a model trained on all LEIE exclusion types gives lower total claims, lower total paid, and fewer unique beneficiaries positive SHAP contributions toward exclusion. That feature pattern is not publicly reproduced here. Label mixture and shorter observation windows are hypotheses a frozen rerun should test.
A useful design correction is to compare per-month billing rates rather than panel totals. The article reports roughly 140 claims per month for fraud-excluded providers and 63 for non-excluded providers, but that comparison is under reconciliation. The general point survives: panel totals combine billing rate with observation-window length.
Other researchers have noticed this too. One Medicare fraud study used eight exclusion codes as labels, including 1128(b)(4) [2]. The LEIE captures only “brazen, outlandish billing patterns,” while subtle fraud responsible for the majority of financial losses goes undetected [3]. And 38% of providers with fraud convictions remain in medical practice, with 21% never suspended [4]. Even the fraud-specific codes miss a substantial portion of actual fraud.
The Positive-Unlabeled Problem
Even restricting labels to fraud-specific exclusion codes leaves a deeper problem: what the ML literature calls positive-unlabeled (PU) learning.
In standard binary classification, we have positive examples (fraud) and negative examples (not fraud). In healthcare fraud detection, we have positive examples (providers on the LEIE) and unlabeled examples (everyone else). The difference matters: we don’t know which unlabeled providers are truly non-fraudulent and which simply haven’t been caught yet.
A survey of the PU learning literature identifies a foundational assumption: SCAR, or Selected Completely At Random [1]. The idea is that the selection process chooses labeled positives independently of their features. That assumption is difficult to defend in an enforcement setting. Scheme visibility, billing volume, investigative priorities, and state enforcement capacity can all make labeling probability depend on observed or unobserved provider characteristics. This article does not estimate the size of those selection mechanisms.
The design concern is selection bias: if labeling probability varies by scheme, state, or billing features, a supervised classifier can reflect that enforcement selection. This article does not estimate over- or under-representation by scheme.
This series treats LEIE-based screening as a positive-unlabeled learning problem. Providers who were never reviewed remain unlabeled rather than confirmed negatives [5]. The resulting research question is how enforcement selection shapes the patterns a model estimates.
One way to test whether the federal list tells the whole story is to compare it against state-level exclusion data.
Where State Lists Help (and Where They Don’t)
State Medicaid programs can maintain exclusion or suspension lists that capture state-specific actions. California’s Suspended and Ineligible (S&I) list is one example. Like the LEIE, these lists combine multiple legal and administrative reasons, so adding them increases coverage without producing a pure fraud label.
The article reports about 6% overlap between California’s S&I list and the LEIE. It also reports a compliance-vendor estimate that roughly half of state-level exclusions never appear on the federal list, but the underlying vendor source is not included in this article’s reference list. Neither value is a reproduced result on this site, and the vendor figure is not peer-reviewed. Treat both as unresolved article claims rather than evidence for the design.
The article also reports about 1,240 current LEIE matches by National Provider Identifier (NPI), roughly 580 additional S&I matches, and about 1,817 combined matches. Those counts, the stated 1.8-million-provider denominator, and the resulting 0.1% rate have not been reproduced from one public matching output.
Class Imbalance and What It Means
If a frozen rerun confirms a positive rate near 0.1%, the classification problem will be extremely imbalanced. Published studies provide external comparison points:
| Study | Dataset | Fraud Rate |
|---|---|---|
| Johnson & Khoshgoftaar (2023) | Medicare Part B | 0.046% |
| Johnson & Khoshgoftaar (2023) | Medicare Part D | 0.065% |
| Bauder & Khoshgoftaar (2018) | Medicare Part B (2012-15) | 0.009% |
| Tajrobehkar et al. (2024) | Medicare ophthalmology | 0.038-0.074% |
At these rates, a classifier that predicts “not fraud” for every single provider achieves 99.9% accuracy. That number is meaningless. What matters is precision (of the providers flagged, how many are truly fraudulent?) and recall (of the truly fraudulent providers, how many did we catch?).
Class imbalance also means that even seemingly precise models generate large numbers of false positives. At 50% precision, every two flags produce one true positive and one false positive. A model flag does not itself freeze billing or establish wrongdoing. But if agencies use flags to trigger payment holds, audits, or other interventions, false positives can consume investigative capacity and may burden providers or disrupt patient access. Those downstream consequences depend on the policy attached to the score.
Why such extreme imbalance? Because the labels come from enforcement, and enforcement has its own patterns.
Enforcement Patterns Shape the Labels
The MFCU annual report for fiscal year (FY) 2024 gives us a window into what enforcement actually looks like [7]:
| Statistic | Value |
|---|---|
| Total convictions | 1,151 |
| Fraud convictions | 817 (71%) |
| Patient abuse/neglect convictions | 334 (29%) |
| Total recoveries | $1.4 billion |
| Return on investment | ~$3.50 per $1 spent |
| PCS attendants as % of fraud convictions | 36% |
Personal care service attendants account for 36% of fraud convictions despite being one of roughly 80 provider types. The report does not identify why PCS attendants account for that share. Detectability and prosecutorial burden are hypotheses for future analysis. The defensible observation is that enforcement records reflect which cases investigators pursue and successfully resolve.
The $1.4 billion figure covers recoveries reported by state Medicaid Fraud Control Units, while broad national fraud estimates cover the entire U.S. healthcare system and rely on different definitions and methods [7, 8]. Dividing one by the other would create a false comparison. The defensible point is narrower: enforcement records identify only the cases that enter, survive, and conclude within an enforcement process. Providers without an exclusion therefore form an unlabeled group that can include both legitimate providers and undetected misconduct.
What This Means for the Series
The label problem constrains everything that follows. In Post 3, when we compare billing patterns of excluded versus non-excluded providers, the comparison is between providers labeled as excluded (a heterogeneous group including fraudsters, substance abusers, and loan defaulters) and providers not labeled as excluded (a group that includes both legitimate providers and undetected fraud).
In Post 4, when we train classifiers, label quality caps what the model can achieve. A perfect classifier on imperfect labels still produces imperfect results.
The more careful approach is to treat the LEIE as a noisy, enforcement-biased signal that captures a non-random subset of problematic providers. It is one national public enforcement signal, not ground truth.
References
- Bekker, J. & Davis, J. (2020). Learning from positive and unlabeled data: A survey. Machine Learning, 109(4), 719-760.
- Johnson, J.M. & Khoshgoftaar, T.M. (2023). Data-centric AI for healthcare fraud detection. SN Computer Science, 4(4), 389.
- Tajrobehkar, M. et al. (2024). Utilization analysis and fraud detection in Medicare via machine learning. medRxiv, 2024.12.30.24319784. [Preprint; not peer-reviewed.]
- du Preez, A. et al. (2025). Fraud detection in healthcare claims using machine learning: A systematic review. Artificial Intelligence in Medicine, 160, 103061.
- Kumaraswamy, N. et al. (2022). Healthcare fraud data mining methods: A look back and look ahead. Perspectives in Health Information Management, 19(1), 1i.
- Vinay, M.S., Yuan, S., & Wu, X. (2022). Fraud detection via contrastive positive unlabeled learning. IEEE International Conference on Big Data.
- OIG OEI-09-25-00200. (2025). Medicaid Fraud Control Units fiscal year 2024 annual report.
- NHCAA. (n.d.). The challenge of health care fraud. https://www.nhcaa.org/tools-insights/about-health-care-fraud/the-challenge-of-health-care-fraud/
Last updated: July 21, 2026.
The evidence labels now separate the public LEIE taxonomy from site-specific category shares, provider matches, and billing comparisons that do not yet have a matching public output.
This series references the working paper: Cholette, V. (2026). What Do Medicaid Fraud Classifiers Actually Detect? SSRN Working Paper.
Suggested Citation
Cholette, V. (2026, February 15). The label that isn't: Why "excluded" doesn't mean "fraudulent." Too Early To Say. https://tooearlytosay.com/research/methodology/medicaid-fraud-labels/Copy citation