Est.

Machine Learning for Benefits Fraud Detection in Self-Funded Plans

Sponsors who fund their own claims need layered, explainable pre-payment fraud scoring.

· 9 min read · Updated
Cover illustration for “Machine Learning for Benefits Fraud Detection in Self-Funded Plans”
AI in HR and Benefits · October 26, 2026 · 9 min read · 2,063 words

Machine learning for benefits fraud detection in self-funded plans is a financial oversight issue. Sponsors pay fraudulent claims from plan funds below stop-loss attachment points, so prudent plan governance includes checking how fraud is detected. To protect plan funds, sponsors need pre-payment scoring, models that work together, findings they can explain and a way to keep track of claims trends, even when a vendor says it uses AI.

No single model is best. The strongest approach combines several: supervised learning (Random Forest, Gradient Boosting, Logistic Regression) to score claims against known fraud, unsupervised anomaly detection to find providers who bill differently from peers treating similar patients, knowledge-guided models that build in expert rules, and graph analytics to expose coordinated billing rings. In the Medicare hospitalization study, a fully unsupervised model gave a nearly five-fold lift over random targeting, and it needs no labeled fraud. Supervised models depend on scarce confirmed fraud data and miss schemes investigators never caught. Whichever mix is used, it should score claims before payment and explain why a claim looks suspicious.

Why self-funded plan sponsors carry fraud risk that fully-insured employers do not

In a self-funded arrangement, the employer pays claims as incurred while a third-party administrator processes and adjudicates them. Unlike a fully-insured employer, the sponsor absorbs fraudulent payments as plan spend rather than leaving that claims exposure with an insurer. Those payments raise effective cost-per-employee and distort renewal benchmarking. Stop-loss coverage has a threshold, and losses below the specific or aggregate attachment point remain entirely the sponsor’s exposure.

That financial structure places the sponsor behind an information gap it cannot close through claims administration alone. Shekhar et al., in Can Machine Learning Target Health Care Fraud? Evidence from Medicare Hospitalizations, published in J Policy Anal Manage, describe the imbalance between providers who know what care was delivered and insurers who pay for it. Providers have the informational advantage. A submitted claim describes the services for which payment is requested, but fraudulent coding can exploit the payer’s limited knowledge of the underlying care. The sponsor funds the consequences when that discrepancy escapes detection.

How the fraud threat facing self-funded plans has changed

Traditional healthcare fraud includes upcoding, duplicate claims and billing for services that were never provided. These schemes often follow recognizable patterns that fixed rules can flag. The threat has since become more difficult to isolate. AI-generated clinical documentation can back up claims for procedures that never happened, and synthetic patient identities and coordinated billing rings can spread fraud across thousands of claims.

Rule-based systems depend on signatures that someone has already identified and translated into an edit. That makes them useful against known patterns, but poorly suited to schemes that do not match those signatures. More rules can also produce false positives that consume investigators’ attention while leaving novel schemes undetected. For a self-funded sponsor, reliance on rule-based pre-payment edits alone therefore leaves a gap between the fraud a TPA is configured to recognize and the fraud the plan may actually fund.

What ML fraud detection actually does that rule-based systems cannot

Machine learning changes the basis for deciding which claims deserve scrutiny. Rather than checking only whether a claim matches a known signature, it models expected billing behavior and identifies departures associated with fraud. Context becomes part of the assessment. Diagnostic codes, procedure codes, billing codes, provider characteristics, claim frequency and amounts can be assessed together, revealing combinations that a human auditor would be unlikely to express as individual rules. A provider’s pattern can look suspicious even when no single claim breaks a fixed rule.

Chy’s 2024 paper in WJARR describes ML models as learning from new data and adapting to emerging fraud schemes, in contrast with the reactive nature of rule-based detection. Operationally, that means detection need not remain tied to the patterns recognized when an edit was written. Timing is equally important for the sponsor. ML scoring can run on every claim before payment is released, rather than only supporting an investigation after funds have left the account. A risk score identifies where review is warranted, however, and does not itself establish that fraud occurred.

What are the best machine learning approaches for detecting healthcare billing fraud?

Supervised learning uses historical claims labeled as fraudulent or legitimate to score new claims for similarity to known fraud. Random Forest, Gradient Boosting and Logistic Regression are examples of this approach. Its strength is recognition of established patterns. Buyers should assess whether the labeled examples are reliable and relevant to the plan’s exposure. It depends on scarce confirmed fraud data, which makes it harder to learn when most claims in the training set are legitimate.

Unsupervised learning instead finds claims or providers whose behavior differs sharply from that of peers who treat similar patients. It does not require labeled fraud examples. The Medicare study developed a fully unsupervised approach that provided a nearly five-fold lift over random targeting of hospitals, validated against DOJ anti-fraud lawsuit data. It also identified the diagnosis, procedure and billing codes that made a hospital appear suspicious. That finding supports its use in prioritizing investigations, with each anomaly requiring verification and performance in a self-funded plan requiring separate validation.

Weakly supervised and knowledge-guided models help where these other models fall short. Pure anomaly detection may miss patterns that matter in clinical care, while supervised learning needs labels that are hard to get. Expert knowledge supplies an additional guide. Mozafari et al.’s CleverCatch, published at BigData 2025, embeds expert rules about prescription behavior, including consistent preference for high-cost drugs over clinically equivalent lower-cost alternatives. The study found average improvements of 1.3% in AUC, a model discrimination measure, over four state-of-the-art anomaly detection baselines, improving detection accuracy and interpretability.

Graph analytics looks at the links between providers, billing entities, patients and addresses, going beyond each claim on its own. Its particular strength is coordinated fraud detection. For plans that use a TPA, PBM and specialty pharmacy, those links may span vendors, so buyers should ask what data the models can use. Serious detection programs combine these approaches because recognition of known fraud, discovery of unusual behavior, clinical knowledge and relationship analysis address different weaknesses.

The specific fraud patterns that ML is designed to surface in health plan claims

Upcoding occurs when providers move diagnosis or procedure codes into higher-reimbursement categories without clinical justification. Peer comparison makes the pattern more visible. The Medicare study looks for anomalous hospital coding by comparing how providers code similar patients and finding patterns that repeatedly increase reimbursement. The approach can surface suspected upcoding that has yet to be prosecuted by identifying anomalies independently of prior fraud labels. The relevant signal is a repeated coding pattern.

Prescription drug manipulation poses a related but different risk, including too many prescriptions, PBM billing anomalies and the consistent use of high-cost drugs when clinically equivalent lower-cost alternatives are available. CleverCatch specifically targets that cost-preference signal in prescribing behavior. The finding still needs to be understood in its clinical context. The CAA 2026, signed February 3, 2026, requires PBMs to remit 100% of manufacturer rebates to ERISA plans and make comprehensive compensation disclosures as covered service providers under ERISA. Those disclosures create a data trail that models can analyze, although disclosure itself is not a fraud finding.

Services-not-rendered schemes involve claims for procedures, equipment or visits that did not occur. Identity fraud can use stolen patient identities to back up those submissions, and synthetic identities can make fraud harder to detect. AI-generated clinical records can make false claims appear supported, prompting the use of computer vision to detect document forgeries. For buyers, this is a different capability from detecting unusual billing amounts or code combinations.

Coordinated billing rings spread fraud across providers, billing entities or shell companies. Relationships reveal what isolated claims may conceal. Graph analytics maps shared addresses, ownership, referral patterns and patient overlap to show how those submissions are linked. That makes coordinated billing an important test of whether a vendor examines the network around a claim. It also makes cross-vendor visibility relevant to plans whose medical and prescription spending is administered separately.

Where ML fraud detection falls short and why those limits matter for plan sponsors

Overall model accuracy can hide a failure to detect fraud when legitimate claims vastly outnumber fraudulent ones. In labeled datasets, fraud rates can be well below one percent. A model can therefore appear accurate by treating nearly everything as legitimate. The headline metric is not enough for evaluation. Sponsors should ask how the vendor measures its fraud catch rate and false-positive rate, rather than accepting a broad accuracy claim as evidence that plan funds are protected.

Training data also reflects the enforcement process that produced it. Historically confirmed fraud covers only the schemes that investigators caught. The Medicare study warns that supervised methods “are biased by the fact that enforcement is non-random, and will likely fail against the ever-changing nature of fraud”. Novel schemes may be absent from those examples. A sophisticated buyer should therefore ask how the program detects unfamiliar behavior, especially AI-generated documentation and synthetic identities, rather than judging it only by performance against past cases.

A black-box score without a reason does not give investigators a sufficient basis for a coverage denial or investigation referral. The finding must be usable by people. Sponsors should ask whether investigators get the codes or behavioral patterns that make a claim look suspicious, and which decisions need human judgment. The Medicare model’s code-level explanations illustrate the distinction between a system that ranks suspicion and one that helps an investigator understand it.

False positives create a separate operational burden by delaying legitimate claims, generating member complaints and consuming investigation capacity. Probabilistic scoring can improve prioritization compared with binary rules, but it does not eliminate those costs. Threshold setting remains a human judgment call. The point at which a score triggers review determines how much suspicious activity receives attention and how much legitimate activity enters the same queue. Vendor evaluation should therefore cover how false positives are measured and managed alongside how many claims the model flags.

The fiduciary question of what ERISA plan sponsors owe on fraud controls

ERISA’s duty of prudence asks whether fiduciaries act as a prudent person in a like capacity, familiar with such responsibilities, would act in managing plan assets. Delegating administration does not remove oversight responsibility, and employer size does not remove that obligation. Prudence is a process standard. Sponsors should reasonably evaluate and monitor available controls, including pre-payment detection, even after appointing a TPA. Public programs’ shift away from “pay and chase” strengthens the case for examining reliance on post-payment recovery, but it does not establish a blanket legal requirement to buy a particular model. Sponsors should be able to document the TPA’s methodology, testing and updates, catch and false-positive rates, and escalation process.

What should HR and finance leaders ask their TPA and stop-loss carrier about ML fraud controls?

Does the system score every claim before payment, or only claims that first trigger rule-based flags? Which supervised, unsupervised and knowledge-guided models are operating, and how does the program address novel patterns absent from training data? Model names alone do not establish adequate coverage. The same discussion should cover prescription drug fraud, PBM billing anomalies and whether graph analytics looks for coordinated billing across claims or looks at each claim on its own.

Process questions then establish what happens after detection. When a claim is flagged, does the investigation team receive a score, a reason or both, and what requires a human decision? How does the vendor measure and manage its false-positive rate? Stop-loss alignment deserves a separate discussion with the carrier. Does it see the TPA’s detection outputs, how does it define and investigate fraud in a stop-loss claim, and does fraudulent inflation of underlying claims affect the attachment point calculation? These questions reveal whether detection and financial protection operate with shared visibility or in separate systems.

Plan-level analytics add a complementary view beyond the TPA’s investigation workflow. Sponsors that benchmark carrier pricing and plan design annually are better positioned to identify anomalies in their own claims data. An AI-native benefits partner can extend that visibility from enrollment across the full renewal cycle. That is a role limited to providing visibility. Its value is giving HR and finance leaders a basis to recognize unusual trends, question their administrators and monitor whether the plan’s fraud controls remain appropriate to its exposure.

Sources

  1. Proactive Fraud Defense: Machine Learning's Evolving Role in Protecting Against Online Fraud
  2. CleverCatch: A Knowledge-Guided Weak Supervision Model for Fraud Detection
  3. Can Machine Learning Target Health Care Fraud? Evidence from Medicare Hospitalizations - PMC
  4. Unsupervised Machine Learning for Explainable Health Care Fraud Detection

More in AI in HR and Benefits