RESEARCH DOSSIER
Algorithmic Claims Adjudication and Fairness in Health Insurance
Abstract. Algorithmic claims adjudication determines, recommends, routes, or records decisions about an insured person's entitlement to payment for health services. Its consequences extend from insurer accounting to provider reimbursement, patient access, clinical continuity, and the distribution of administrative burdens. This monograph develops a comparative, source-grounded account of fairness in such decisions. It separates post-service payment adjudication from prior authorization, prospective benefit determination, fraud screening, risk scoring, and public-benefit eligibility; reconstructs the interaction of coverage rules, evidence quality, prediction, human review, and appeals; and evaluates how error rates, coverage correctness, disparate outcomes, costs, and contestability can be measured without confusing their denominators. Primary evidence includes U.S. Medicare Advantage oversight and policy, the National Association of Insurance Commissioners' insurer survey, federal transparency data interpreted with its documented limitations, U.S. claims-and-appeals rules, European data protection and artificial-intelligence legislation, international insurance conduct standards, technical interoperability specifications, and original scholarship on algorithmic fairness and selective labels. Three principal conclusions follow. First, an aggregate denial rate is not a measure of algorithmic bias or illegality: adjudication quality requires a legally and clinically grounded entitlement reference. Second, audits that observe only appealed or finally denied claims cannot establish overall error rates without accounting for selection, reversals, heterogeneous claims, and unobserved outcomes. Third, procedural protections and clinical reliability are properties of an entire sociotechnical decision process, including model inputs, rule versions, human discretion, reasons, access to records, and timely independent review. The contribution is an integrative methodological and comparative analysis of established evidence, not a new theory or a claim of original causal estimation.
Keywords: Health insurance; claims adjudication; algorithmic decision-making; automated denials; medical necessity; prior authorization; procedural fairness; discrimination; fairness metrics; selective labels; insurance regulation; Medicare Advantage; appeals; independent review; clinical reliability; human oversight; European Union Artificial Intelligence Act; GDPR; claims data; source-critical methodology.
Introduction
Health insurance transforms the uncertain cost of health services into an institutional promise of financial protection. That promise acquires operational meaning when a particular service, claimant, date, contract, clinical record, and reimbursement request are evaluated. The World Health Organization's health-financing policy account locates purchasing alongside revenue raising and risk pooling; its financial-protection guidance makes clear that formal coverage alone does not guarantee affordable access to care. The adjudication of an individual claim is therefore one of the places where financing arrangements become an experienced entitlement.
An algorithmic adjudication system may be as straightforward as a deterministic policy-edit rule or as complex as a predictive model trained on historical claims. The word algorithmic refers here to computable rules or models materially involved in screening, routing, recommending, approving, denying, adjusting, or reviewing claims. Whether a particular system qualifies as artificial intelligence under a named legal instrument is a separate classification question. Such distinctions matter: evidence that an insurer uses software is not evidence that it deploys machine learning; evidence that it uses machine learning somewhere in its operations is not evidence that a specific denial was algorithmically produced.
The central research object is the sequence connecting a submitted claim to a consequential coverage or payment determination. A claim may involve an insured member, a healthcare provider, a third-party administrator, an insurer, a public financing institution, and a clinical reviewer. Their interests can diverge: payment integrity, timely care, contractual enforcement, sustainable pooled financing, and administrative cost reduction may be pursued through the same workflow. The relevant standard for deciding any particular claim is derived from the applicable law, benefit contract, program rules, medical evidence, and procedural safeguards. Disagreement over these standards cannot be settled by a generic prediction score.
This investigation relies on documented public evidence and transparent inferential reasoning. It does not purport to audit a private insurer's source code, obtain confidential claim-level records, estimate causal treatment effects, or establish misconduct in a particular case. All quantitative observations are identified by their source population, year, and denominator. Regulatory statements are attached to jurisdiction and instrument. Analytical recommendations are derived from the reviewed literature and institutional requirements; they are not presented as a newly created doctrine.
Foundations and Source-Critical History
The adjudication object and the insurance promise
A health claim is a request for financial performance under a particular healthcare coverage arrangement. An adjudication determines whether and how that request is recognized under the applicable rules. It may result in payment, partial payment, a request for additional information, an adjustment, or an adverse determination. The claimant and the economically affected person need not coincide: a provider may submit a claim, the insurer may reimburse the provider, and the patient may face a residual liability or interrupted care. Claim accuracy, procedural fairness, distributive equity, and financial protection are therefore related but distinct outcomes.
Four operational stages must be identified. Eligibility verifies whether coverage and relevant membership existed when services occurred. Benefit and clinical review interprets the covered service, exceptions, necessity standard, and supporting medical record. Financial adjudication applies provider contracts, allowed amounts, deductibles, copayments, coordination of benefits, and payment rules. Contestation examines an unfavorable decision through correction, reconsideration, appeal, or independent review. A workflow can be accurate in one stage and defective in another. A correct arithmetic calculation applied to the wrong benefit rule remains an erroneous determination; a clinically justified authorization that arrives after treatment becomes impossible may also frustrate access.
Prospective prior authorization determines whether a requested service is eligible to proceed under a plan's utilization-management rules before the service. Retrospective adjudication processes a payment or coverage claim after the event. Predetermination estimates benefits without necessarily promising payment. Fraud detection raises a question requiring further verification rather than automatically establishing noncoverage. Risk adjustment estimates population morbidity for financial transfers and should not be confused with an individual's benefit entitlement. This vocabulary tracks the separate uses of HL7 FHIR ClaimResponse, which distinguishes Claim, Preauthorization, and Predetermination and the differing financial force of their results.
From administrative rules to predictive decision support
Electronic benefits administration grew from structured transactions and coding systems, where rule-based processing could check billing elements quickly and consistently. More recent data-intensive systems classify suspicious submissions, predict claim outcomes, retrieve documents, rank cases for clinical review, and generate summaries. These operations can reduce missing-field errors and expedite straightforward payments; they can also multiply a defective rule or misleading proxy across large populations. The empirical distinction between deterministic edits and learned predictions is indispensable because each fails differently: rule errors arise through inappropriate coverage logic, coding, exception handling, or stale regulatory versions; model errors arise additionally through training data, objective selection, distributional shift, calibration, and proxy discrimination.
The 2025 NAIC health-insurer survey documents that responding companies use AI and machine learning across several operations. Its survey population consisted of 93 companies approached through 16 states, and its definitions included statistical models that other surveys might exclude. The report states that 84% of health insurers in the survey currently used AI/ML in some capacity; this is a finding for responding insurers under its definition, not a census of all insurers or proof that 84% automate payment denials. The associated technical survey memorandum distinguishes current use, planned use, and exploration, which must not be aggregated into a claim about operational deployment.
The survey evidence matters because it establishes real industry adoption without granting access to individual decision logic. The correct analytical question becomes which system is used at which stage, for what purpose, with what human authority and what error-monitoring mechanism. Claims adjudication, prior authorization, and utilization management cannot be read as interchangeable adoption categories. Vendor participation further complicates accountability because a health plan may rely on a proprietary tool while retaining the regulatory responsibility for decisions affecting its insured members.
Evidence of errors and the danger of causal overreach
The U.S. Department of Health and Human Services Office of Inspector General's 2022 Medicare Advantage review found that 13% of sampled denied prior-authorization requests met Medicare coverage rules and 18% of sampled denied payment requests met applicable Medicare coverage and plan billing rules. These percentages condition on previously denied requests, and the two denominators refer to different administrative processes. Reviewers found issues involving restrictive clinical criteria, documentation assessments, human errors, and processing-system errors. The report therefore provides direct evidence that erroneous adverse determinations can occur and sometimes be reversed. It does not establish the fraction caused by predictive AI, the population-wide denial rate, or a particular insurer's algorithmic intent.
An equally important source-critical limitation arises in public denial counts. The KFF analysis of 2024 Marketplace claims reports a 19% in-network final denial rate among reporting HealthCare.gov insurers and 37% out of network. Its underlying administrative series comes from the CMS Transparency in Coverage public files, whose coverage and reporting conventions require attention. The KFF analysis describes post-service claims and excludes pre-service prior authorizations, while claims initially rejected and later corrected and paid may no longer appear as final denials. Aggregated reporting cannot reveal which decisions relied on AI, whether each denial was contractually correct, or whether a patient suffered clinical injury. Treating these rates as observed rates of algorithmic wrongful denial would combine three unsupported inferences.
Comparative standards and institutional responsibilities
The International Association of Insurance Supervisors' Insurance Core Principles, ICP 19, places fair treatment, understandable communication, claims, complaints, and disputes within conduct-of-business supervision. Its 2025 consumer-treatment application paper recognizes that consumer capabilities and needs vary across a product's life cycle. These are supervisory standards and guidance, not a universal private cause of action or a substitute for the applicable national claims law. They help frame the comparative question: whether automated administrative efficiency is accompanied by enforceable mechanisms through which a person can understand and challenge a materially adverse decision.
Decision Models, Rules, and Data Inputs
Reconstructing the operative decision
A properly specified claims decision can be represented as a determination function applied to policy and program rules, a defined service, a member's coverage history, clinical information, financial schedules, and the governing time. Let D = f(P, S, M, C, F, T, V), where P is the applicable policy or public-program rule set; S is service and coding information; M is member eligibility; C is clinical evidence; F is the financial/payment schedule; T is the temporal context; and V is the rule and model version. This is a bookkeeping notation for the sources of a decision, not an asserted predictive model or new theoretical construct.
A reviewable decision record should retain at least the coverage clause or legal criterion, rule version, relevant input provenance, effective date, decision path, output status, reason for denial or adjustment, human intervention, notice, and later modifications. The distinction between reason and score is central. A model's risk score can identify cases for further scrutiny; it does not by itself specify which coverage condition was unmet. A denial letter that states only that an opaque score exceeds a threshold does not reveal the contractual basis of the adverse determination.
Data in this setting are generated by institutions with different incentives and collection regimes. Diagnosis and procedure codes reflect care and billing practices, but may also reflect access, coding intensity, and historically uneven treatment. Payment histories depend on negotiated prices, coverage design, provider networks, and prior denials. Clinical notes contain context but vary in completeness and format. A prediction trained on previous insurer decisions may reproduce earlier misinterpretations of policy or records. Training labels must therefore be classified as observed decisions, observed payments, independent coverage-entitlement judgments, or clinical outcomes; each measures a different object.
The proxy problem
The Obermeyer et al. (2019) study examined a widely used population-health management algorithm that predicted healthcare costs as a proxy for health need. At the same score, Black patients had greater measured illness burden than White patients; the study explains this discrepancy through unequal historical spending at comparable levels of need. The setting was allocation of extra care-management support, not proof that a health insurer's claims-denial model used the same algorithm. Its direct relevance to adjudication is methodological: historical cost or paid-claim labels can encode differences in realized care access even when a prediction is well calibrated against expenditure.
Suppose an adjudication model predicts the likelihood that a claim will be paid under past practice. High predictive accuracy may simply indicate that the model learns the insurer's inherited decisions. If the objective is correct entitlement, the appropriate target is the payment or coverage determination warranted by contract, law, and clinical evidence after independent review. A model accurate on historically paid claims can be inaccurate on legitimate claims earlier denied, appealed, or never submitted. The audit must ask whose decision is treated as ground truth and why it should be trusted.
Model-assisted, rule-bound, and human decisions
Three deployment arrangements have distinct accountability profiles. In straight-through processing, deterministic rules or classifiers finalize certain routine cases without case-by-case intervention. In decision support, software recommends a disposition for an authorized human. In triage, a model allocates scarce review attention without determining final entitlement. An insurer may use all three simultaneously, but they should not share a single description such as “AI decides claims.” The critical institutional variables are actual power to override, the record available to the reviewer, the time afforded, the burden of justification, and whether the system automatically executes an adverse recommendation when no reviewer intervenes.
Human presence alone does not establish meaningful review. An employee who routinely accepts recommendations without seeing case evidence may function as a nominal approval step. Conversely, structured rules can improve human decision quality when they identify missing information and make the relevant provision easy to verify. The analytical issue is whether human judgment independently assesses the entitlement standard and the claimant's circumstances, not whether a particular interface labels its operator a clinician or supervisor.
Regulatory rule versioning in practice
The U.S. 2024 Medicare Advantage utilization rule requires plans to apply the relevant Traditional Medicare coverage framework and limits when plan-specific internal coverage criteria may be used. A subsequent CMS memorandum on algorithmic assistance states that a software prediction of a patient's expected length of post-acute care cannot, without individual reassessment, support termination of services. It also rejects using algorithmic results alone to deny an inpatient admission or downgrade the patient's status. These requirements concern the specified Medicare Advantage decisions and should not be mechanically extended to every private insurance contract in every jurisdiction.
For a system deployed across years, the audit must reproduce the applicable rule at the time of decision. Clinical guidelines, regulatory effective dates, plan amendments, coding schedules, and provider arrangements change. A decision can be incorrect because the versioned policy engine is stale even when its statistical model is unchanged. Version-control evidence therefore has direct legal and clinical relevance, separate from software engineering convenience.
Fairness Auditing, Disparity, and Human Override
Five questions that fairness analysis must keep distinct
The fairness inquiry begins with coverage correctness: among cases with a verified entitlement to a covered payment or authorization, how often does the process improperly deny or delay that entitlement? It then examines treatment of comparable cases: do people subject to the same applicable coverage rule and clinically relevant evidence receive consistent dispositions? It asks whether disparities track legally protected or socially vulnerable characteristics after careful attention to clinical differences and entitlement conditions. It tests procedural equality: do people have similar practical opportunities to learn the reason, provide additional evidence, and obtain timely review? Finally, it considers consequences, including interruptions to treatment, unpaid provider services, out-of-pocket costs, administrative work, and financial distress.
No single score can answer all five questions. Legal entitlement and clinical judgments differ by service and jurisdiction; observed population-level differences may reflect case mix, benefit design, historical inequity, discriminatory processes, or combinations of these factors. The correct response is to specify what each statistic measures and the assumptions required for interpreting its differences.
Error-rate balance, calibration, and objective selection
The Hardt, Price, and Srebro (2016) analysis formalizes equality of opportunity and related error-rate criteria in supervised learning. Chouldechova (2017) demonstrates that apparently desirable fairness criteria may conflict when outcome prevalence differs across groups. Their demonstrations concern classification and prediction; they do not identify which legal rule defines equal treatment in a particular insurance setting. The Barocas, Hardt, and Narayanan (2023) synthesis emphasizes that statistical fairness must be interpreted within the institution and decision process producing labels and effects.
Let Y = 1 denote independently established entitlement for the claim, and let A = 1 denote an adverse initial disposition. For a group g, the wrongful-denial rate among entitled requests is P(A = 1 | Y = 1, G = g). A disparity in this measure can reveal unequal burden among similarly entitled requests if the determination of Y is comparable and properly observed. It is not the same as the ordinary denial rate P(A = 1 | G = g), which includes legally justified rejections. False approvals among non-entitled requests must also be considered because wrongful payment changes pooled costs and may ultimately affect insured members. The choice of thresholds and loss functions entails substantive tradeoffs; a high-accuracy classifier trained on the wrong target cannot resolve them.
A calibrated risk model forecasts the frequency of a specified event among cases given a score; calibration is an attribute of the selected event and data distribution. If the event is “the insurer historically denied this claim,” calibration tells us about replication of historical behavior. It cannot establish fairness, clinical necessity, or legal validity. A clinically meaningful fairness audit must validate the target and error labels against independent review before drawing conclusions from conventional measures of predictive accuracy.
Selective labels, appeals, and the unobserved denominator
The selective-labels analysis by Lakkaraju et al. (2017) establishes a general problem: downstream outcomes are frequently observed only for cases selected by earlier decisions. In insurance, cases that are appealed are a selected subset of denials; a successful appeal may depend on literacy, clinician time, access to records, urgency, and administrative capacity. Claims never submitted after a prior denial or anticipated rejection may disappear from the data entirely. Reviewing the success rate among appeals cannot reveal the wrongful-denial rate among all entitled claims without explicit assumptions or an independent sampling design.
This point also governs interpretation of external review. If few denials are appealed, the proportion reversed among appeals is conditional on individuals having reached that stage. If the insurer changes an earlier decision after informal resubmission, the ultimate administrative record may not classify the case as a final denial. Audit fields must preserve initial decision, correction, internal appeal, independent review, final disposition, and relevant dates as separate stages. Otherwise a system may seem more accurate because its most serious errors were corrected after imposing material costs on claimants.
Appropriate audit sampling and group comparisons
A defensible evaluation draws a stratified sample of initial approvals, initial denials, partial payments, administrative rejections, and cases escalated to clinical review. Stratification should reflect clinically material service type, jurisdiction, product, rule version, provider/network status, urgency, and relevant member characteristics. Independent reviewers should receive adequate medical evidence and the rule applicable at the time. Disagreements between reviewers must be resolved through documented adjudication rather than suppressed. For each stratum, report the sampling probability and inverse-probability weights so estimates represent the intended universe.
If protected-group attributes are unavailable, missing, self-reported, or imputed, the analysis must document their provenance, lawful processing basis, and measurement uncertainty. Proxy labels inferred from postal codes or surnames can misclassify individuals and create their own privacy risks. Small cells produce unstable rates; hierarchical analyses may improve precision but must not obscure substantial clinically important subgroup harm. Fairness testing should include intersectional characteristics where sample size and data governance permit, while separating legally prohibited differential treatment from empirical disparities that require further causal investigation.
Override as an observable institutional practice
Override rates can be informative only when the denominator is meaningful. Report the proportion of recommendations presented to a reviewer, the proportion actually examined, the proportion changed, the elapsed time, and changes in correctness and clinical outcome. Low override frequency may mean accurate recommendations, weak reviewer autonomy, or organizational pressure. High override frequency may mean poor model performance or a thoughtful triage design that intentionally routes difficult cases to experts. Neither metric is self-interpreting.
An override audit should reconstruct whether the reviewer could access original records; whether the model score was explained as a reason or merely displayed; whether divergent decisions triggered penalties or additional review; and whether the patient or provider could supply missing evidence. For medical-necessity determinations, review should identify the relevant clinical specialization. An insurer's performance bonus or vendor fee arrangement can influence incentives without automatically establishing a particular wrongful decision. Empirical evaluation should test these pathways rather than infer motives solely from aggregate cost savings.
Contestability, Clinical Reliability, and Institutional Accountability
The effective opportunity to contest a decision
A health-insurance adverse determination is contestable when the affected person receives usable notice, can understand the operative rule and supporting evidence, can submit pertinent material, and can obtain an appropriately independent review before the loss becomes irreparable. This is an institutional condition that can be measured through elapsed time, information access, reversal pathways, reviewer independence, and observable care consequences. The existence of a nominal appeal form says little about effective access when evidence is unavailable, medical care is urgent, or the required documentation exceeds a person's practical capacity.
In the United States, 45 CFR § 147.136 governs internal claims-and-appeals processes and external review within its defined scope and exceptions. It requires specified reasons and process information in adverse determinations, protections for full and fair review, and external review mechanisms for applicable cases involving medical judgment. Its precise application depends on plan type, grandfathered status, and the relevant category of determination. An effective audit checks the actual claimant's procedural rights rather than asserting that every health plan worldwide provides the same appeal stages.
The 2024 CMS interoperability rule creates additional prior-authorization process and data-exchange obligations for its specified payers. Relevant provisions include more timely decisions and specific denial reasons beginning generally in 2026, while several API obligations begin generally in 2027. These dates and the exclusion of drugs from specified provisions prevent a false impression that every function was legally operational as of November 2026. Prior-authorization timelines must remain analytically separate from the deadlines governing post-service appeals under other instruments.
Clinical reliability as a longitudinal property
Claims decisions can affect clinical outcomes through treatment postponement, interrupted post-acute services, prescription access, referral changes, or unexpected cost exposure. A technically correct denial under a particular coverage rule can still reflect a poor benefit design, but the two questions are analytically separate: the insurer's compliance with an existing entitlement and the welfare adequacy of that entitlement. Clinical reliability studies should distinguish these levels and observe patient health, time to treatment, treatment abandonment, avoidable complications, and clinician administrative burden where data and ethical approvals permit.
The Medicare Advantage OIG investigation describes the access and administrative consequences of denials that complied with applicable coverage rules. Its results support targeted scrutiny of clinical criteria and documentation errors but do not measure a general causal effect of automated adjudication on morbidity or mortality. Such outcomes require a suitable comparison design, sufficiently long follow-up, and careful control for changes in case mix and provider behavior.
Human review, reason-giving, and vendor accountability
The NAIC Model Bulletin on insurers' AI use articulates state supervisors' expectations for insurer governance, written AI programs, fairness, accuracy, controls, and documentation. The NAIC expressly describes it as a model bulletin rather than a self-executing uniform federal statute. Actual legal effect depends on adoption or implementation in a state, alongside existing insurance laws. Its reasoning nevertheless identifies an institutional requirement central to private-vendor systems: responsibility for a consumer-impacting decision cannot be dissolved merely by outsourcing the model or buying the software.
An audit trail must therefore preserve model/vendor version, policy version, training or validation purpose, input history, decision rationale, reviewer action, and a way to investigate incidents. Oversight should include appropriate separation between model development, benefit-rule interpretation, quality assurance, and appeals. Procurement arrangements should make independent validation and incident investigation possible while respecting privacy and legitimate trade secrets. The insurer's ability to explain the governing decision should survive a vendor change or the retirement of a model.
The NIST AI Risk Management Framework offers a voluntary cross-sector structure for governing, mapping, measuring, and managing AI risks. It cannot replace binding claims law, but it can support documentation, incident investigation, model monitoring, and decision-specific risk assessment. Its voluntary character matters in comparisons: a system described as aligned with a framework is not necessarily legally compliant or independently certified.
Civil rights and the boundaries of current legal claims
Section 1557 of the U.S. Affordable Care Act, as explained by HHS civil-rights guidance, addresses discrimination in covered health programs and activities, including discriminatory limitations of insurance claims. The 2024 Section 1557 final rule expressly addresses patient-care decision-support tools. Its implementation, however, is affected by litigation. The official Medicaid notice of judicial stays documents particular stayed or enjoined provisions, including nationwide limits concerning specified interpretations of sex discrimination. A reliable legal statement must identify the exact provision, jurisdiction, judicial order, and time; it cannot treat either the entire 2024 rule as universally in force or all protections as wholly extinguished.
The international comparator is not a simplified assertion that all automated health decisions are prohibited. The legal status of a system depends on its purpose, user, data category, degree of automation, and effects. These factors must be documented before imposing a legal label on a workflow.
Data, Measurement, and Validation
What the available administrative data can establish
The CMS Exchange public-use documentation supports analysis of reported plan and issuer variables under its coverage and file conventions. The KFF 2024 claims analysis observes about 496 million reported claims in its specified HealthCare.gov insurer population, with approximately 451 million in-network claims and an average 19% final in-network denial rate. The analysis records fewer than 1% of denied claims as internally appealed and reports that 66% of internal appeals were upheld, subject to the data's suppression and denominator definitions. These are descriptive results for the stated observation frame. They do not identify the prevalence of AI-driven wrongful denials, an algorithmic treatment effect, or a census of U.S. group and public insurance.
Reporting at plan or issuer level aggregates together service categories, provider circumstances, contracts, decision routes, and repeated submissions. Claims rejected administratively for incorrect identifiers can differ fundamentally from denials after independent clinical assessment. A percentage aggregated over such cases measures the observed administrative mixture. Comparing this mixture across insurers or years without adjustment may confound product offerings, changing reporting coverage, benefit networks, case mix, resubmissions, and policy revisions. The KFF methodology warns that insurer and plan participation in reporting changes over time and that denial-reason categories are incomplete.
The OIG sample is complementary rather than interchangeable. It conditionally evaluates denied Medicare Advantage requests against coverage and billing criteria, while KFF examines Marketplace reporting on post-service claim disposition. A rigorous synthesis can state that both document significant administrative accountability questions; it cannot pool their percentages or infer a common effect size.
Minimum viable audit data and linkage
For each decision event, retain a privacy-protective claim identifier, de-identified member linkage, plan and jurisdiction, service date, submission and adjudication dates, care setting, diagnosis and procedure codes where lawfully available, clinical necessity criteria, policy version, monetary amounts, and initial/final dispositions. Also retain whether a rule-based edit, learned model, or human reviewer materially contributed; whether the system was used for triage or determination; the model and rule versions; the cited reason; corrections; appeals; and independently assessed entitlement when sampled. Data governance must specify retention, legitimate access, privacy, security, and the handling of sensitive medical information.
The unit of analysis should be explicit. A claim may contain multiple service lines. A member may file repeated claims, and a single course of treatment may generate authorization requests, claims, corrections, appeals, and adjustments. A denial rate over service lines will differ from a rate over claims, persons, or treatment episodes. Clustered standard errors or multilevel uncertainty calculations are required when repeated observations are correlated. Counting every successive resubmission as an independent denial inflates both apparent throughput and dispute incidence.
Sampling and measurement equations
For a defined observation universe U, let N be the number of eligible submitted claims, D the number receiving an initial adverse determination, and F the number ultimately remaining adversely determined. Initial denial rate D/N and final denial rate F/N measure different events; F may be affected by resubmission rules and appeals. Define E as a verified entitlement outcome established through independent policy and clinical review. A weighted estimate of wrongful initial denial among entitled cases is the estimated count of decisions with A = 1 and E = 1 divided by the estimated count with E = 1. If verification is sampled, weights must account for differential selection into review. Confidence intervals should reflect both probability sampling and clustering by member, provider, plan, or reviewer as appropriate.
Disparities should be reported alongside accuracy measures, absolute volumes, adverse clinical outcomes, and uncertainty intervals. For example, a one-percentage-point difference over hundreds of thousands of eligible requests may have substantial welfare consequences, whereas large percentage differences in tiny subgroups may be statistically unstable. Because group prevalence and case mix differ, equal overall denial rates can coexist with unequal wrongful-denial rates, and equal false-negative rates can coexist with unequal burdens from appeal cost and delay. Counterfactual claims about discrimination require a justified identification strategy and exposure to the relevant alternative decision process.
Prospective validation and implementation assessment
Retrospective performance on historical claims is a necessary but incomplete diagnostic. External validation should use a temporally and institutionally distinct population, retain the applicable rule and benefit versions, and test calibration and error rates by service, urgency, region, plan, clinical complexity, and relevant groups. Prospective evaluation should measure the complete process after deployment, including human response, appeal rates, time to correction, and changes in care access. Model and policy drift may occur independently: a demographic or coding shift changes data, while an amended contract changes the decision standard. Both should trigger documented investigation.
A credible causal comparison between an algorithm-assisted workflow and a conventional process requires attention to concurrent policy changes, insurer/provider selection, model targeting of difficult cases, staffing and capacity, seasonality, and regression to the mean. Randomized evaluation may be feasible for low-risk administrative triage if protections are maintained; clinical denials and access to urgent care impose heightened ethical and legal constraints. Staggered implementation or interrupted time-series designs can provide information when assumptions are checked, but observational before-and-after differences alone cannot identify algorithmic effects.
Missing evidence and meaningful negative findings
Public disclosure presently leaves substantial gaps around proprietary models, individual coverage criteria, service-level denominators, claims that were never submitted, and the clinical outcomes of disputed decisions. Such gaps must be recorded as limits of inference. No verified private insurer source code, individual claim file, national causal estimate of automated denial harm, or nationally harmonized cross-border fairness dataset was available to this research. This is a finding about this dossier's evidence base, not a claim that such data are impossible to obtain under appropriate legal authority.
International Synthesis and Scientific Conclusions
United States: several distinct regulatory layers
The United States illustrates the need for layered classification. Medicare Advantage has program-specific coverage and utilization-management rules; Marketplace and employer-sponsored plans are governed through different federal and state arrangements; state insurance regulation addresses insurer conduct and may adopt NAIC-derived expectations; and civil-rights provisions operate according to their particular scope and current judicial constraints. The CMS Medicare Advantage memorandum provides a concrete individual-assessment requirement for specified decisions. The federal appeals regulation shows how information, reasons, independence, and review rights can be legally structured. Neither is a global rule for every algorithmic decision in every coverage market.
The NAIC survey results show adoption and governance activity across surveyed insurers, while the OIG findings and KFF descriptive series establish reasons to study erroneous denials and weak contestation. The sources are stronger together when each is used for its own question: adoption, compliance errors, and administrative outcomes respectively. They do not collectively justify assigning all denials to machine learning.
European Union: intended purpose and the actual legal classification
The European Union's Artificial Intelligence Act identifies high-risk uses according to defined purposes. Annex III, point 5(c), expressly covers AI systems used for risk assessment and pricing concerning natural persons in life and health insurance; it does not, merely by that wording, place all health-insurance claims management within that category. The European Commission's AI Act essential-services guidance gives claims management as an example outside point 5(c). However, another use of an AI system may fall under a different high-risk category, such as public-authority decisions on eligibility for essential public healthcare benefits under point 5(a), or satisfy another applicable classification. The legal analysis must be performed on intended purpose and factual use rather than an industry-wide label.
Separately, the EU General Data Protection Regulation, particularly Article 22, regulates decisions based solely on automated processing with legal or similarly significant effects, subject to enumerated grounds and safeguards. The legal consequence depends on whether decision-making is solely automated, whether the effect reaches the threshold, what lawful basis is invoked, and how special-category health information is handled. Human intervention must be assessed in substance, not inferred from a nominal checkpoint. GDPR safeguards and AI Act classifications are distinct legal analyses; satisfying one instrument does not automatically establish compliance with the other.
Cross-border comparability and its limits
The IAIS ICP 19 conduct principles provide a common supervisory vocabulary of fair treatment and claims resolution, while the WHO's health-financing framework directs attention to purchasing arrangements, financial protection, and institutional design. Different systems use public payers, social insurance funds, private carriers, or combinations. An apparently identical denial outcome may concern a different claimant, statutory entitlement, payer-provider contract, or appeal institution. International fairness statistics should therefore be compared only after mapping eligibility, evidence standards, payment mechanics, and available remedies.
Large-scale automated triage may improve timeliness, detect duplicate invoices, and reduce avoidable administrative friction. Those benefits are empirical possibilities whose scale depends on deployment, supervision, and evaluation. The risk of industrialized error is equally real: a single defective rule, mislabeled objective, or obstructed appeal channel can affect many claims before detection. No universal assumption that automation improves or worsens fairness follows from its computational character. The question is the evidenced quality of decisions and remedies in the institution where the system operates.
Integrated findings
The evidence supports seven linked scientific judgments. Entitlement validity is the principal anchor for studying wrongful denials. Decision provenance is necessary to attribute an error to a model, a rule, human review, or organizational policy. Group fairness needs clinically and legally meaningful comparison sets rather than raw denial parity. Appeal records are selected evidence and cannot by themselves estimate population error. Procedural protections have practical content only when timely review, understandable reasons, and usable records are available. International regulation classifies systems by jurisdiction, purpose, and legal effect. Finally, efficient payment administration and effective health insurance both require measurement of the eventual financial and clinical consequences to the insured population.
These judgments are original analytical syntheses of documented sources and established statistical methods. They do not assert the discovery of a new theoretical mechanism, an original experimental effect, or a jurisdiction-independent legal doctrine.
Conclusion
Algorithmic health-insurance claims adjudication is a consequential institutional practice because it converts coverage rules and clinical information into decisions over payment and access. Its quality depends on the validity of those rules, the evidence used to apply them, the properties of software and predictive models, the independence of human judgment, the availability of understandable explanations, and effective review when a decision is challenged. Fairness is therefore observed through an entire decision sequence: from submitted service and policy entitlement to initial disposition, correction, appeal, final payment, and effects on care.
The reviewed evidence establishes both the spread of algorithm-supported insurer operations and documented administrative errors. It also establishes strict limits: U.S. denial rates are not synonymous with illegal denials; documented erroneous payment decisions are not necessarily AI decisions; and disparities in historical spending can distort apparently accurate models without establishing that a particular insurer's denial system shares the same mechanism. Better evidence requires independent assessment of entitlement, versioned decision records, representative review samples, analysis of legally relevant groups, and linkage to patient and provider consequences.
A scientifically adequate account of fairness combines the contract and public-law standard of each jurisdiction, clinical evidence, statistical error analysis, and practical contestability. The institutional benchmark is a reasoned, traceable, timely, correctable determination of the insured entitlement. Further progress depends on access to suitable case-level data, validated clinical review, stronger disclosure of decision pathways, and prospective comparative studies. This dossier provides the evidence-grounded analytical foundation for such work within the international encyclopedia of insurance and risk science.
Technical Appendix: Definitions and Audit Calculations
Defined observations
Initial decision means the first consequential response after a complete claim enters the adjudication workflow. Final decision means the administrative endpoint under the study's observation rule, which must say whether corrected resubmissions, internal appeals, and independent review have been incorporated. Entitled claim means a claim that an independently specified and reviewable standard of contract, law, and necessary clinical evidence recognizes as payable or authorizable; an audit should document uncertainty in contested cases. Algorithm involvement means a versioned rule, model, or software recommendation materially influenced a stage of processing, distinct from mere storage or transmission.
Example audit table, with no empirical values assumed
| Measure | Numerator | Denominator | Main interpretation risk |
|---|---|---|---|
| Initial adverse-decision rate | Initially adverse claims | Claims received and meeting study inclusion criteria | Submission-stage and claim/line definitions |
| Final denial rate | Claims finally adverse | Comparable eligible submitted claims | Resubmission and follow-up rules |
| Wrongful-denial rate | Initially denied claims independently found entitled | All independently entitled claims in representative audit sample | Verification bias and review selection |
| Appeal uptake | Denials entering a defined appeal stage | Denials eligible to appeal | Barriers and unrecorded informal corrections |
| Appeal reversal rate | Reversed adverse decisions at specified stage | Adverse decisions reaching that stage | Selection into appeal |
| Override rate | Human changes to model recommendations | Recommendations actually presented for human decision | Triage mix and nominal human involvement |
| Time to effective resolution | Elapsed time from defined origin to final actionable disposition | Eligible cases with complete follow-up | Censoring, urgent care, absent claims |
Uncertainty and error accountability
Statistical uncertainty should be expressed with appropriate intervals for defined estimands, and the audit report should disclose missingness, sampling probabilities, reviewer agreement, and sensitivity to alternative entitlement interpretations. Claimant and provider identifiers must be handled under applicable privacy and security law; comparison by protected characteristics requires an appropriate lawful basis. The loss from a wrongful denial may include administrative friction and delayed treatment even if eventual payment is made. Such consequences require direct observation rather than inference from a model score.
References
Barocas, Solon, Moritz Hardt, and Arvind Narayanan. 2023. Fairness and Machine Learning: Limitations and Opportunities. Cambridge, MA: MIT Press. https://fairmlbook.org/
Centers for Medicare & Medicaid Services (CMS). 2023. “2024 Medicare Advantage and Part D Final Rule (CMS-4201-F).” https://www.cms.gov/newsroom/fact-sheets/2024-Medicare-advantage-and-part-d-final-rule-cms-4201-f
Centers for Medicare & Medicaid Services (CMS). 2024a. “CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F).” January 17. https://www.cms.gov/newsroom/fact-sheets/cms-interoperability-prior-authorization-final-rule-cms-0057-f
Centers for Medicare & Medicaid Services (CMS). 2024b. “Frequently Asked Questions Related to Coverage Criteria and Utilization Management Requirements in CMS-4201-F.” Health Plan Management System memorandum, February 6. https://www.cms.gov/files/document/hpms-memo-faq-coverage-criteria-and-utilization-management-cms-4201-f-02-6-2024-pdf.pdf
Centers for Medicare & Medicaid Services (CMS). 2026. “Health Insurance Exchange Public Use Files,” including Transparency in Coverage data and data dictionaries, accessed November 16, 2026. https://www.cms.gov/marketplace/resources/data/public-use-files
Chouldechova, Alexandra. 2017. “Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments.” Big Data 5 (2): 153–163. https://doi.org/10.1089/BIG.2016.0047
European Commission. 2026. “AI Act Service Desk: Essential Services,” interpretive examples for Annex III, accessed November 16. https://ai-act-service-desk.ec.europa.eu/en/essential-services
European Parliament and Council. 2016. Regulation (EU) 2016/679 (General Data Protection Regulation). Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2016/679/2016-05-04/eng/pdf
European Parliament and Council. 2024. Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated online text as available November 16, 2026. https://eur-lex.europa.eu/eli/reg/2024/1689
Hardt, Moritz, Eric Price, and Nati Srebro. 2016. “Equality of Opportunity in Supervised Learning.” Advances in Neural Information Processing Systems 29. https://proceedings.neurips.cc/paper_files/paper/2016/hash/6a9659feb1216f14f7384ba499518b38-Abstract.html
Health Level Seven International (HL7). 2023. “FHIR R5: ClaimResponse Resource.” Version 5.0.0, trial-use resource. https://hl7.org/fhir/claimresponse.html
International Association of Insurance Supervisors (IAIS). 2024. Insurance Core Principles and Common Framework (ComFrame), December 2024, published online June 2025. https://www.iais.org/uploads/2025/06/IAIS-ICPs-and-ComFrame-December-2024.pdf
International Association of Insurance Supervisors (IAIS). 2025. “IAIS Publishes Application Paper on Fair Treatment of a Wide Range of Consumers.” July 7. https://www.iais.org/2025/07/iais-publishes-application-paper-on-fair-treatment-of-a-wide-range-of-consumers/
Kaiser Family Foundation (KFF). 2026. “Claims Denials and Appeals in ACA Marketplace Plans in 2024.” Analysis of CMS reporting data, accessed November 16, 2026. https://www.kff.org/patient-consumer-protections/claims-denials-and-appeals-in-aca-marketplace-plans-in-2024/
Lakkaraju, Himabindu, Jon Kleinberg, Jure Leskovec, Jens Ludwig, and Sendhil Mullainathan. 2017. “The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables.” Proceedings of KDD 2017. https://www.kdd.org/kdd2017/papers/view/the-selective-labels-problem-evaluating-algorithmic-predictions-in-the-pres
National Association of Insurance Commissioners (NAIC). 2023. “NAIC Members Approve Model Bulletin on Use of AI by Insurers.” December 4. https://content.naic.org/article/naic-members-approve-model-bulletin-use-ai-insurers
National Association of Insurance Commissioners (NAIC). 2025a. Health Insurance Artificial Intelligence/Machine Learning Survey Results. May. https://content.naic.org/sites/default/files/inline-files/Health%20Survey%20Report%20-%20FINAL%205.9.25.pdf
National Association of Insurance Commissioners (NAIC). 2025b. 2025 Health Artificial Intelligence (AI)/Machine Learning (ML) Survey Analysis. May 9. https://content.naic.org/sites/default/files/inline-files/Health%20Survey%20Memo%20to%20BDAIWG%2005092025%20-%20Final_1.pdf
National Institute of Standards and Technology (NIST). 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
Obermeyer, Ziad, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. 2019. “Dissecting Racial Bias in an Algorithm Used to Manage the Health of Populations.” Science 366 (6464): 447–453. https://pubmed.ncbi.nlm.nih.gov/31649194/
Office of Inspector General, U.S. Department of Health and Human Services (HHS OIG). 2022. Some Medicare Advantage Organization Denials of Prior Authorization Requests Raise Concerns About Beneficiary Access to Medically Necessary Care. Report OEI-09-18-00260. https://oig.hhs.gov/reports/all/2022/some-medicare-advantage-organization-denials-of-prior-authorization-requests-raise-concerns-about-beneficiary-access-to-medically-necessary-care/
U.S. Department of Health and Human Services. 2016–2021, amended. “Internal Claims and Appeals and External Review Processes,” 45 CFR § 147.136, text accessed November 16, 2026. https://www.law.cornell.edu/cfr/text/45/147.136
U.S. Department of Health and Human Services. 2024. “Nondiscrimination in Health Programs and Activities,” 89 Federal Register 37522 (May 6, 2024). https://www.govinfo.gov/content/pkg/FR-2024-05-06/pdf/FR-2024-05-06.pdf
U.S. Department of Health and Human Services, Office for Civil Rights. “Section 1557: Coverage of Health Insurance in Marketplaces and Other Health Plans.” Accessed November 16, 2026. https://www.hhs.gov/civil-rights/for-individuals/section-1557/fs-health-insurance/index.html
U.S. Centers for Medicare & Medicaid Services, Medicaid.gov. 2026. “Nondiscrimination Policy Update,” including notices of judicial stays and injunctions concerning the 2024 Section 1557 rule, accessed November 16. https://www.medicaid.gov/nondiscrimination-policy-update
World Health Organization (WHO). “Financial Protection.” Accessed November 16, 2026. https://www.who.int/health-topics/financial-protection
World Health Organization (WHO). “Health Financing Policy.” Accessed November 16, 2026. https://www.who.int/teams/health-financing-and-economics/health-financing/health-financing-policy