Can AI health-insurance fraud tools be trusted?
A review of 33 studies finds broad technical promise but little comparable evidence on external validation, cost, fairness or real-world deployment. A suspicious-claim score is not proof of fraud and still requires accountable human investigation.
The short version
- What happened
- Seven databases produced 450 records; after 227 duplicates and screening, 33 peer-reviewed studies from 2016–2025 were included.
- Who it affects
- Patients can face delayed care or reimbursement when false alerts are treated as verdicts, so human review and appeal routes are essential.
- What remains unproven
- English-only peer-reviewed literature through March 2025
Keep in mind
English-only peer-reviewed literature through March 2025
Large heterogeneity prevented quantitative meta-analysis
One author performed study selection synthesis and repeated the quality assessment
External validation, interpretability, fairness and governance were rarely evaluated
Cost-effectiveness and real-world operational outcomes were sparse
Source analysis · 1 linked source. Primary analysed source date: 15 October 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
The answer: promising filters, not automated fraud verdicts
Machine-learning, deep-learning and graph-based systems often outperformed the comparators chosen in 33 reviewed studies. That makes AI a credible tool for ranking claims or networks for further investigation. It does not show that insurers can safely reject care, recover money or accuse a person or provider from a model score. Most evidence came from country-specific datasets with inconsistent fraud definitions and evaluation, while external validation, economic evidence and real-world governance were sparse.
Fraud is rare relative to legitimate claims. A model can report excellent overall accuracy by labelling almost everything legitimate, while still missing costly fraud or burdening honest patients and clinicians with false alarms. The operational question is therefore not whether one algorithm posts a high percentage. It is how many genuinely fraudulent cases are found, how many legitimate claims are disrupted, whether investigators can understand the alert, and whether the system produces net savings after review costs and appeals.[1]
How 450 records became 33 studies
The author followed PRISMA 2020 and searched Scopus, Web of Science, OVID Medline, Embase, PubMed, Dimensions and OpenAlex for peer-reviewed English-language work published from 2016 through March 2025. The search identified 450 records, removed 227 duplicates and screened 223 titles and abstracts. Forty reports were sought, three were unavailable and four failed the full-text criteria, leaving 33 studies for qualitative synthesis.
The review extracted dataset, country, fraud type, AI method, performance measures, funding and limitations. Twenty-four studies addressed fraudulent claims, while municipal and international insurance schemes appeared in six studies each. Machine learning was most common, followed by deep learning, graph methods and hybrids. The review did not pool a single effect because inputs, labels, comparators and metrics were too heterogeneous; that was the methodologically safer choice.[1]
Why the reported performance cannot form a league table
Recall was reported in 18 studies, accuracy in 17, precision in 16, F1 score in 15 and AUROC in 13. Only three reported precision-recall AUC, a particularly informative measure when positive cases are rare. Four reported specificity. Different train-test splits, resampling methods, class prevalences and definitions of confirmed fraud mean that a 95% accuracy in one dataset cannot be compared directly with an 85% result in another.
Thirty studies claimed improved performance over an existing method, but only three discussed generalisability and two highlighted limited interpretability. The pattern is consistent with a development literature rewarded for model improvement, not a deployment literature designed to expose failure. Fraud labels may themselves reflect past investigative priorities: if audits historically targeted some providers or regions more heavily, a model can learn enforcement patterns rather than underlying misconduct.[1]
The strongest practical finding is what remains unmeasured
Only nine studies addressed cost-effectiveness and eight runtime efficiency. The review found little consistent evidence on implementation cost, operational workflow, fairness, privacy, appeals or long-term financial sustainability. A model may raise recall while creating so many low-quality alerts that investigators spend more time, providers wait longer for payment and patients face delayed care. Technical performance is only one component of public value.
External validation was also weak. Many models were tied to the coding rules, reimbursement incentives and fraud patterns of China, the United States or another single system. A graph pattern that indicates collusion in one claims network may be routine referral structure in another. Before transfer, insurers need locked testing on local data from a later period and an independent organisation, with thresholds chosen around the real cost of misses and false accusations.[1]
What this means for patients, clinicians and investigators
Patients need due process when an algorithm affects authorisation, reimbursement or investigation. An alert should trigger proportionate review, not automatic denial. People must be told the reason for an adverse action in understandable terms and have a route to correct coding errors, identity misuse or unusual but legitimate treatment. Protected characteristics and proxies such as postcode, provider network and diagnosis mix require fairness testing because unequal scrutiny can become self-reinforcing.
Clinicians and providers need timely notification and a human reviewer who can understand context. Fraud teams should separate anomaly detection from evidence of intent, record which alerts led to confirmed findings and monitor whether investigators over-trust model rankings. Public payers and regulators should demand audit logs, access controls, data minimisation, drift monitoring and published error rates. Procurement language should never equate a suspicious claim with proven fraud.[1]
Review-level limits and what would change the assessment
The review included only English-language peer-reviewed studies through March 2025, so it may miss later, non-English and operational evidence. One author conducted the quality assessment twice rather than using independent reviewers, and titles and abstracts were screened with an AI-assisted platform before full eligibility review. The Drummond economic-evaluation checklist rated 31 studies good, yet the review simultaneously found that comprehensive cost analysis was usually absent; those broad ratings should not be read as proof of low bias or deployment readiness.
Confidence would rise with shared benchmark definitions, later-period and cross-system external validation, precision-recall reporting, documented human review and prospective evaluation of savings, delays, appeals and unequal error. The review received no financial support and declared no commercial conflict. Its defensible conclusion is narrower than a sales claim: AI can help prioritise investigation, but health-insurance organisations still lack comparable evidence that these systems are fair, economical and reliable at scale.[1]