RESEARCH DOSSIER
Insurance Claims Process Mining and Event Log Reconstruction
Abstract. Insurance claims process mining reconstructs and analyzes documented claims-handling behavior from evidence distributed across administrative, financial, communication, and supplier information systems. Its central scientific challenge is the distinction between an operational occurrence, its documentary representation, and the event constructed by an analytical procedure. This monograph examines claims event semantics, provenance, cross-system identity resolution, object-centric event logs, process discovery, conformance checking, concurrency, temporal uncertainty, operational case-file validation, and reproducible comparison across insurance lines. It combines a critical synthesis of established process-mining research, official exchange specifications, selected insurance case studies, and international and jurisdiction-specific conduct and data-protection instruments. Explicitly synthetic examples demonstrate how aggregation, delayed recording, uncertain linkage, censoring, and choice of reference process alter findings. The analysis establishes that process reconstruction is an inference problem whose conclusions depend on identifiable relationships and the reliability of observations. A process map provides evidence of documented behavior under a defined extraction and interpretation procedure. Contractual entitlement, causal effects of interventions, and fairness of individual decisions require additional evidence. The contribution is an insurance-specific integration of existing methods and analytical reasoning; no new theory, original empirical effect estimate, or benchmark dataset is claimed.
Keywords: Claims science; insurance claims; process mining; event log reconstruction; event semantics; case correlation; object-centric process mining; OCEL; XES; partial order; conformance checking; provenance; claims operations; data quality; temporal uncertainty; censoring; market conduct; reproducibility.
Introduction
A claim is an assertion of an insurance-related entitlement that enters a documentary and institutional process. The assertion may become a claim record through a telephone call, a digital form, an intermediary, a provider bill, or a notification from another participant. The insurer subsequently makes decisions about coverage, evidence, valuation, liability, benefit calculation, payment, and closure. Each decision may generate traces in several information systems. The scientific object of this research is the relationship between those traces and the operational pathways they support.
The Process Mining Manifesto of the IEEE Task Force on Process Mining situates the discipline within the improvement and control of operational processes. In insurance, its usefulness extends from explaining variation in claim development to auditing selected administrative obligations. Process discovery extracts a representation of observed sequences or relationships. Conformance checking compares observations with a defined reference model. Performance analysis attaches temporal and other measurements to these relationships. Each operation addresses a different analytical question.
Claims workflows are significant because their activities affect both the claimant's experience and the financial representation of obligations. A request for documents may extend the time until a decision. A reserve adjustment may reflect developing knowledge without changing contractual liability. A payment instruction can precede banking completion. An administrative closure can precede a supplement, appeal, litigation, or recovery. These distinctions change the meaning of process sequences and performance indicators.
The first research question concerns identity and semantics: which operations count as events, and how should they be assigned to claims, exposures, contracts, documents, payments, providers, and other objects? The second concerns reconstruction under uncertainty: which temporal and behavioral conclusions remain defensible when the record is incomplete, inconsistent, or retrospective? The third concerns validation: how should an analyst distinguish a data defect, a permitted pathway, a procedural deviation, an operational inefficiency, and potentially unfair treatment?
The study uses a targeted critical literature synthesis rather than a systematic review or meta-analysis. It examines accessible author manuscripts, official standards descriptions and regulatory texts, institutional author abstracts where publisher access is unavailable, and selected case summaries. It analyzes insurance evidence through hypothetical examples and elementary formal reasoning. Research was conducted on November 19, 2026. No confidential insurer event logs or claim files were obtained. Numerical examples expose the consequences of modeling decisions rather than estimate actual portfolio behavior.
The boundaries of the dossier follow the established Claims Science program. The focus is observation and reconstruction. Quantitative leakage estimation, reserve adequacy, predictive model development, litigation strategy, and automated adjudication require distinct research. Their interfaces with process mining are examined through the structure and interpretation of event evidence.
Claims Event Semantics and Provenance
Operational facts, records, and analytical events
An operational event is an occurrence defined at a selected business granularity. A source record is a digital representation produced by a particular system. An analytical event is a representation constructed for an investigation. A user pressing a button, a server committing a transaction, and a document reaching its recipient can correspond to one business activity, several activities, or evidence of different occurrences.
This distinction precedes algorithm selection. A claim notification may be received at 08:00, registered at 10:00, and synchronized at 15:00. Each time can be valid for its own purpose. Using registration to measure the response from receipt removes the initial two hours. Using synchronization as registration adds five hours. Both distortions arise even when every clock is correct.
De Weerdt and Wynn (2022), in their foundations of process event data, describe case identifiers, activity labels, timestamps, and data preparation as foundations of event-log construction. Insurance applications must explicitly define the entity and action represented by these fields. A label such as "approved" requires an object and a proposition: a coverage determination, a repair estimate, a payment authorization, or an exception approval. A shared label obscures these differences if its context is discarded.
Insurance also demands separation of knowledge time from effective time. A correction entered on Friday may assert that coverage became effective on Monday. A reserve adjustment can be attributed to a different accounting period from its posting date. A file note may describe an earlier conversation. An analyst should distinguish the time an event is reported to have occurred, the time it became known to an institution, and the time its account was changed. Reconstructing a process as of an earlier information cutoff requires avoiding hindsight.
The state of a claim is a collection of attributes at a particular time. An event is a supported occurrence or transition. The state "under review" can persist for a month without indicating daily reviews. Reconstructing transitions from successive snapshots requires explicit treatment of observation gaps. If no changes are visible between two snapshots, either the relevant state remained constant or intermediate changes were unrecorded by the snapshot frequency.
Granularity and lifecycle
Claims systems represent actions at multiple granularities. The following event definitions are illustrative.
| Analytical event | Preferred evidence | Distinction |
|---|---|---|
| Notification received | Acknowledged intake or delivery | Claim registration may occur later |
| Coverage decision recorded | Versioned decision and scope | Claimant notification is separate |
| Evidence requested | Outbound request and recipient | Drafting is a different event |
| Evidence received | Inbound document receipt | Review may occur later |
| Settlement authorized | Approval record and amount | Banking clearance is separate |
| Payment completed | Reconciled transaction | Instruction is not completion |
| File reopened | State-transition history | A supplement may be a new object |
A lifecycle-sensitive log can represent start, suspension, resumption, and completion as transitions associated with a stable work-item identifier. The analyst may subsequently derive service intervals or a completion-only view. This choice changes what can be established about waiting, rework, and asynchronous actions. Where start events are unavailable, elapsed time between adjacent events is an inter-event duration, rather than measured service effort.
A current status snapshot supplies particularly limited evidence about the path to that status. A closed record can have undergone multiple reversals and reopenings. A change-data capture stream may preserve transactions after its activation but cannot recover earlier overwritten versions without additional records. Snapshots remain useful for reconciliation of current states, outstanding claims, and inventory.
Activity abstraction can merge technical operations into business milestones. A letter generation, print job, and queue submission might support a single business event only if the definition and evidence justify that representation. An abstraction intended for operational capacity analysis may be too coarse for a regulatory timing inquiry. Conversely, preserving every low-level click can produce a detailed but substantively uninformative diagram.
A useful event dictionary states the activity definition, business object, triggering condition, lifecycle phase, source evidence, exclusions, and version. Its definitions should be evaluated using actual claim scenarios before large-scale extraction. A changed dictionary is a methodological change, even if the extraction query and source system remain the same.
Provenance and reproducibility
The W3C's PROV data model (Moreau & Missier, 2013) establishes concepts for describing entities, activities, derivation, and responsibility. Applied to claims research, the source record is an entity; an extraction or classification is an activity; the resulting event is a derived entity. The software or authorized reviewer may be associated with responsibility for the generation process.
An event table should retain a stable identifier, activity label and definition version, source system, native record key, time fields and precision, associated objects, transformation version, and provenance reference. Inferred events should be distinguishable from directly observed records. The claim file need not be copied into the analytical dataset. A restricted reviewer can use a provenance reference to examine the source under the applicable access controls.
Provenance is particularly important when automated classification produces an event label from correspondence. A classifier may identify a message as a request for medical evidence. The delivery receipt, message content, and classifier output establish different propositions. The transformation should retain source references, model version, uncertainty information, and correction history.
An immutable snapshot and digest establish which bytes were processed. They do not establish that a source record was truthful when created or that an interpretation is correct. A signature authenticates a designated source within a signature scheme; independent validation connects the authenticated statement with the business occurrence it purports to describe.
Operational observation and legal meaning
A legal obligation can exist before an information system records it. A workflow may include valid manual activity outside monitored applications. A system can also record an unsuccessful communication attempt or a transaction subsequently reversed. The interpretation of an event log consequently depends on the business proposition being tested.
Consider a denial registered on Tuesday, a letter generated on Wednesday, and dispatch on Thursday. The decisive event for a procedural control depends on the applicable requirement. A rule concerning notice requires evidence of the legally relevant communication; an internal approval rule may concern the moment of decision. A timestamp offset cannot repair a mismatch between these concepts.
This distinction has institutional consequences. The completeness of a log must be evaluated in relation to the claims it is used to support. A system sufficient for monitoring routine workloads may be inadequate for establishing individual legal compliance. A scientifically reliable extraction can still produce answers outside the intended legal scope if the questions are formulated imprecisely.
Cross-System Case Identity and Data Preparation
Claim identity as an analytical choice
A conventional trace associates events with one case identifier. Insurance processes have several potentially valid identities. A motor accident may produce vehicle and bodily injury exposures. A person may have claims under multiple coverage sections. A provider bill may be resubmitted under a changed identifier, and a payment may settle multiple invoices. A policy number is usually too broad for an individual claim-process trace; a transaction ID is usually too narrow for its overall pathway.
The unit must correspond to the question. For notification to first decision, a stable claim or coverage-exposure identifier may be appropriate. For the development of repair estimates, an estimate and revision chain are more suitable. For payment reconciliation, payment and allocation objects are central. The same event repository can support different views without assuming that their units are equivalent.
The IEEE XES standard (IEEE, 2023) provides a structured exchange language for event logs and event streams, including schemas and extensible attribute semantics. A case-oriented event representation is suitable when the investigation has one consistent case entity and external relationships can be omitted without altering the relevant meaning. Syntactic schema compliance remains separate from domain validation.
The OCEL 2.0 specification (Berti et al., 2024) allows an event to refer to several objects, with qualified relationships, object changes, and relational, XML, and JSON exchange formats. An insurance event may relate to a claim, an exposure, a document, a payment, and a provider. Preserving such relationships allows later projections without assuming that each event belongs exclusively to one case.
Cross-system linkage
Linkage starts with authoritative keys and their namespaces. The same number in two databases need not identify the same object. A system may reuse identifiers after migration or require an enterprise identifier for uniqueness. Crosswalks should record the originating system, validity period, relationship type, and source of authority.
Explicit foreign keys, audited migration mappings, payment references, and verified document identifiers can establish deterministic links. Where fields are missing, an analyst may construct a limited candidate set from authorized attributes such as policy reference, transaction date, coverage classification, and amount. Probabilistic matches are inferred associations whose validity depends on calibration, treatment of ties, and independent testing. A unique candidate under a restrictive blocking rule is not automatically a verified match.
Bayomie, Di Ciccio, and Mendling (2022) study probabilistic event-case correlation using time-stamped events, process models, and constraints on attributes. Their work shows the methodological relevance of combining different forms of evidence. An insurance application nevertheless requires attention to circularity: if linkage rewards the expected workflow, real deviations may be attributed to the wrong cases. Later conformance analysis using the same model becomes partly determined by the linkage method.
Where practical, linkage used to evaluate a procedural rule should be independent of that rule. A sensitivity analysis can compare exact-key-only reconstructions, attribute-based reconstructions, and process-constrained reconstructions. Deviations that disappear only under the model-based approach warrant additional evidence.
Crosswalks can be one-to-one, one-to-many, many-to-one, or many-to-many. A claim merged during migration may inherit multiple old references. A coverage exposure split for separate adjustment may inherit part of an earlier claim history. Analytical identity must preserve the relation and the period in which it holds. Retrospectively assigning all historical events to a final business object can erase the earlier administrative structure.
Extraction and normalization
Extraction requires an analytical cutoff, a source inventory, and rules for generating events. Transaction journals may provide direct status changes. Document systems may supply deliveries, versions, and access events. Financial systems record instructions, reversals, and reconciliation. Outsourcing systems may supply periodic files with different clocks and granularity.
A pipeline should account for duplicate, late-arriving, and corrected records. An event recovered from a delayed source file need not have occurred when the file was loaded. Idempotent extraction means rerunning the same logic over the same snapshot does not multiply events. Late corrections should produce a new version with an audit trail rather than silently overwrite an earlier published analysis.
Suriadi et al. (2017) identify recurring event log imperfection patterns and propose systematic documentation of quality problems. Insurance applications can organize diagnostics by such defect classes. Examples include duplicate retry records, overwritten statuses, batch times, missing joins, case splits, and inconsistent labels. Specific repairs need domain evidence.
Deduplication should consider durable references, native IDs, payload versions, and transition context. A claimant calling twice within a minute may generate two genuine events. A retransmitted payment instruction may represent one intended transaction, while a reversal followed by replacement is a new pair of financial actions. Deleting all repeated adjacent labels can remove authentic rework.
The same concern applies to status histories. Five daily snapshots with status "review" do not establish five review events. A change table containing five entries could include genuine transitions, corrections, or technical re-saves. Examination of audit metadata and source evidence establishes what can be derived.
An audit-friendly preparation process retains the raw source extract under appropriate restrictions, a deterministic transformation specification, a rejected-record register, corrected-event records with reasons, and the versioned final projection. Transformation counts should reconcile at each stage. A difference between the number of source records and the number of analytical events is expected where aggregation or splitting occurs, but unexplained differences should block conclusions dependent on those counts.
Synthetic many-to-many settlement example
Suppose claims A and B each have an approved invoice, IA and IB. Both are allocated through one payment batch P. The event "batch settled" is a single banking event linked to P. Its allocations are 120 monetary units to A and 80 to B.
A claim-oriented projection may show batch settlement in both traces. This gives two claim-level observations of an associated settlement but one banking event. Summing the batch total of 200 on each projected copy produces 400, double the actual settlement. Summing allocations gives 120 + 80 = 200. For a batch reliability study the denominator is one batch; for a claim settlement study it can be two claims with the relevant milestone.
This example establishes an ordinary aggregation constraint. Event incidence, case incidence, and monetary incidence are different measures. Distinct events, allocated amounts, and case exposures should be counted at their appropriate levels. Object-centric representation preserves the necessary relationships; the analyst remains responsible for defining the correct aggregation.
A further difficulty appears when one supplier invoice is partly assigned to several exposures of the same claim. The claim-level expenditure can be correct while exposure-level amounts are duplicated. Reconciliation must check both the allocation total and the allocation target. Identifiers that span multiple object types should be explicitly typed.
Missingness and diagnostic denominators
Missing events affect both what can be inferred and which claims remain in a study. Useful diagnostics include the eligible population, exact and inferred linkage coverage, missingness by source and activity, unresolved duplicates, temporal precision, and the case-length distribution. Payment completeness may be reconciled with an independent ledger, while communication completeness requires an independently justified expected record.
A missing event and an event that never occurred produce the same absence in an incomplete log. The dependence of missingness on channel, claim size, geography, or manual handling requires separate investigation. If only integrated repair suppliers provide detailed logs, the sample may represent particular regions or selected types of claims. A confidence interval calculated within that sample does not correct selection bias.
For a synthetic population of 1,000 eligible claims, suppose 900 have exact linkage, 70 have probable linkage, and 30 remain unlinked. Exact linkage coverage is 90%; the expanded view covers 97% conditional on accepting probable matches. If 45 deviations appear among the first 900, their observed rate is 5%. With unknown binary outcomes for the remaining 100, the full-population logical range is 4.5% to 14.5%, assuming the 45 outcomes are verified. This is a missing-outcome bound, not a statistical confidence interval. Additional independent observations can narrow it.
Process Discovery and Conformance
Directly-follows relations and process models
A directly-follows graph summarizes adjacency in projected traces. The count on an edge from activity a to activity b equals the number of adjacent occurrences of that ordered pair. The count depends on extraction, activity granularity, ordering, and the selected case perspective. Frequent adjacency establishes repeated observation under these conditions; it does not establish dependence or a normative workflow requirement.
Suppose three synthetic claim traces are A, B, D; A, C, D; and A, B, C, D. Directly-follows counts are A–B: 2; A–C: 1; B–D: 1; B–C: 1; C–D: 2. The graph indicates that B and C may be observed between A and D. It cannot determine whether they are required checks with incomplete capture, independent optional activities, or branches of distinct processes. The longest trace establishes recorded joint occurrence in one case.
Frequency thresholds can suppress rare paths for visual readability. The raw support and omissions should remain available. An unusual unreviewed denial or unauthorized payment may have greater governance significance than a common routine sequence. A graph used for exploration should coexist with separate exception analysis.
A Petri net represents transitions and conditions through places and tokens. A process tree represents behavioral structures through operations such as sequence, exclusive choice, parallelism, and iteration. These representations support formal questions about possible states, completion, and allowed sequences. A descriptive graph for discussion and an executable reference for conformance checking serve distinct purposes.
Leemans, Fahland, and van der Aalst (2013) present a constructive block-structured discovery approach that produces sound, fitting models, with recovery of an underlying process language dependent on sufficient conditions. A structurally sound model can be a useful decomposition of claims behavior. Soundness is separate from reliability of the event observations and substantive permissibility of the behavior represented.
An analyst should choose the modeling language according to the question. A broad descriptive investigation may start with a directly-follows view. An investigation into competing and concurrent activities may use a process tree. A question about whether a particular required review precedes a payment may be represented by a specific constraint. A complete flowchart is unnecessary where one clearly defined proposition is the subject.
Evaluating discovered models
Buijs, van Dongen, and van der Aalst (2014) distinguish four discovery quality dimensions: fitness, precision, generalization, and simplicity. A model permitting arbitrary sequences may fit all observations while offering little discrimination. A restrictive model may fit common pathways while excluding valid exceptions. Evaluation must take account of the model's purpose and the representativeness of its data.
A credible experiment separates training and evaluation cases. Temporally later cases are preferable where the aim concerns future operation. In object-centric data, shared objects can create information leakage across nominal case partitions. Grouped splitting or removal of boundary-crossing objects may be necessary. A decrease in evaluation fitness after a policy change can reflect genuine process drift rather than a deficiency of the historical model.
Simplicity matters because investigators must understand the model and trace important diagnoses to source evidence. An abstraction combining different types of requests may serve a capacity study while obscuring distinct legal requirements. A high-dimensional map can capture numerous relationships while overwhelming its intended audience. Different views can be justified when each retains its definition and provenance.
The reported result should include the algorithm and software version, parameter settings, training and evaluation population, activity filters, case representation, and exact metric definition. A fitness score without these details is difficult to interpret or reproduce. Comparisons between algorithms are meaningful only under compatible evaluation conditions.
Alignment and conformance
A reference model originates from a process specification, applicable operational policy, precise procedural rule, or independently validated description of expected behavior. It requires a defined purpose, effective period, and scope. A model discovered from the same observations generally describes their patterns; it is not independent evidence of normative compliance.
Van der Aalst, Adriansyah, and van Dongen (2012) describe how replaying history on process models connects events with model elements to diagnose deviations and analyze performance. In alignment-based analysis, a synchronized move matches an observation with a model activity, a log-only move represents unmatched observed behavior, and a model-only move progresses modeled behavior without the corresponding event observation. Invisible transitions receive treatment specified by the model and cost scheme.
For a simple synthetic reference A, B, C, where A is notification, B a required review, and C payment authorization, the observed trace A, C yields an unmatched B at cost one under unit costs for unmatched visible steps. At least three explanations remain possible: the review was omitted, the review occurred but was not logged, or the case did not fall under the reference's conditions. A model-only move is a diagnostic result rather than a reconstructed fact.
The trace A, B, B, C has one extra recorded review if the model permits only one. A second B may be genuine additional work, a duplicated record, or a permitted reconsideration absent from the reference. If the model allows a loop, the alignment changes. Model choice and activity semantics determine the proposition that the deviation supports.
Equal alignment costs do not imply equal institutional consequences. A missing mandatory authorization can have greater control significance than an extra administrative note. A useful assessment reports a structural alignment cost alongside separately established control relevance and evidence strength. Combining these into one unexplained score conceals the assumptions needed for interpretation.
Data-dependent and declarative controls
Many insurance requirements are more accurately represented through individual constraints than one complete permitted sequence. A payment exceeding a threshold might require approval before execution; a denial may require communication; a material decision change may require a reason. Conditions depend on product, amount, rule version, and applicable exceptions.
An event log used for data-dependent control evaluation must preserve attribute values as they were known at the time. A claim's final amount may exceed a threshold even though its earlier interim payment was below it. Applying the final value retrospectively would misclassify the earlier transaction.
An absence of conformance flags can arise because all cases complied, because the tested model omitted a requirement, or because sources were incomplete. A case with uncertain ordering may be indeterminate with respect to a required sequence. Reports should identify the tested rule, documented rationale, necessary source fields, exceptions, and unresolved cases.
For a payment-control analysis, a preliminary candidate list can be generated automatically and then triaged by evidence strength. An investigator can inspect high-risk mismatches first without treating the ranking as a determination. Cases excluded by the data preparation pipeline should be counted separately, particularly if their exclusion relates to the risk being evaluated.
Concurrency and Temporal Uncertainty
Time fields and ordering
Insurance systems record occurrence, receipt, registration, persistence, transmission, and accounting-effective time. Source timestamp strings, precision, zone information, interpretation rules, and original clock semantics should be retained. UTC normalization facilitates comparisons where offsets are known; local calendars remain necessary for applicable business-day calculations.
A local timestamp in the repeated hour at the end of daylight-saving time can refer to two distinct UTC instants. When the source provides no disambiguation, both alternatives remain possible. Assigning the analyst's machine time zone invents information. Where a record has date-only precision, midnight represents a storage convention rather than evidence of activity occurring at midnight.
Van der Aalst and Santos (2021) examine the interplay between time and order and describe partial-order representations that preserve uncertainty and use domain knowledge. A timestamp sort supplies a total order even if underlying events are independent or their order is unresolved. Parallel investigations and outsourced assessments make this especially relevant to claims workflows.
Consider two assessments X and Y after notification N and before settlement approval Z. Both assessments have day-level precision. Trusted dependencies establish that N precedes both and each precedes Z. The supported order leaves X and Y incomparable. Both N, X, Y, Z and N, Y, X, Z are compatible linear extensions.
An alphabetical tie-breaker creates a reproducible sequence, but its order is a software convention. It should not be used as evidence that X caused Y or that a requirement for X to precede Y was satisfied. Simultaneous timestamps also do not establish simultaneous business events: limited clock resolution or batch registration may explain the equality.
Interval representations
Suppose occurrence time for event e is within a closed interval from l(e) to u(e). The inequality u(e) < l(f) establishes that e strictly precedes event f for all possible realizations. Overlapping intervals generally leave multiple orders possible. Independently justified dependencies may restrict those orders further.
For the duration from e to f, the unconstrained range has lower endpoint l(f) minus u(e), and upper endpoint u(f) minus l(e). Reliable evidence that f occurs no earlier than e allows intersection with nonnegative values. Additional compatible constraints may narrow the result.
In a synthetic example, evidence receipt occurred between 09:00 and 11:00 and assessment completion between 10:30 and 12:00. The unconstrained difference ranges from minus 30 minutes to 180 minutes. If documentary evidence establishes that the assessment used the evidence and therefore could finish only after receipt, the lower bound becomes zero and the upper bound remains 180 minutes. An arbitrary imputation that produces a negative duration does not establish a procedural anomaly.
For a duration threshold T, if the upper bound is at most T, every compatible duration satisfies the threshold. If the lower bound exceeds T, no compatible duration satisfies it. Otherwise the time evidence is inconclusive. The inclusive or exclusive boundary must match the actual rule. This elementary reasoning preserves useful information without inventing exact times.
The underlying source intervals may themselves depend on known constraints. For example, an email attachment could have been received before a file upload but after an external transmission confirmation. The intersection of supported intervals can improve temporal precision. An empty intersection is evidence that the assumptions or sources conflict; it should be investigated rather than silently repaired.
Uncertain events and conformance bounds
Pegoraro, Uysal, and van der Aalst (2020) examine conformance over uncertain event data, including uncertainty about time, activity, and case membership. An insurance log can preserve such alternatives prior to conformance checking. A set of compatible realizations may yield a range of alignment costs. A best-case cost of zero establishes the existence of a conforming realization, not that the realized process necessarily conformed.
For example, suppose the reference requires X before Y and evidence supports either order. One compatible realization conforms and the other does not. A definitive verdict on ordering exceeds the observations. If a separate document establishes X before Y, the supported set shrinks and the result may become conclusive. The two stages of inference should remain auditable.
The number of possible realizations can grow rapidly. A set of k mutually independent activities allows up to k factorial possible total orders, before uncertain identities and labels are considered. A partial-order representation or invariant may establish a proposition without enumerating all realizations. Sampling linear extensions may reveal sensitivity, but sampled extrema are not necessarily mathematical extrema.
An analyst should distinguish the sources of uncertainty. Missing event existence differs from uncertain event time. Ambiguous case assignment differs from ambiguous activity classification. Each has different implications for conformance and different remedial evidence. Collapsing them into one generic confidence score may obscure the reasons for uncertainty.
Causal ordering and causation
Temporal precedence and causal dependence are distinct. Documentary evidence that a decision used a particular report may justify a dependency between report and decision. The observation that a report arrived before the decision does not itself show that it was used. A database foreign key can establish a reference without proving substantive influence.
A partial order should include only justified temporal or dependency relations. If adding a constraint creates a directed cycle, at least one premise or interpretation is inconsistent. A reviewer should isolate the conflict: a date may be wrong, a relation may be retrospective, or a purported activity may actually be a different event. Deleting edges solely to force acyclicity removes the symptom without resolving the evidentiary problem.
Reconstructing documented dependence differs from estimating an intervention's causal effect. Investigation may precede a long settlement interval because complex claims require both investigation and longer handling. The process graph supports the formulation of an explanatory hypothesis and the design of an inquiry; a causal effect requires further identification.
Operational Validation against Case Files
Independent documentary evidence
Case-file validation compares reconstructed propositions with independently accessible documentary evidence. The unit of validation can be an event, object relationship, time interval, variant, allocation, or deviation. Each requires specified reference materials and an appropriate sampling design.
A useful review asks whether the event occurred, which object and activity it represents, what time the file substantiates, whether predecessor and successor relations have support, which applicable rule is relevant, and whether the classification is definitive or indeterminate. These questions can yield different answers. A file may establish that an activity occurred without supplying its exact time.
The PM² project methodology of van Eck et al. (2015) provides an established methodological context for conducting process-mining projects. Insurance applications particularly benefit from iterative discussion with handlers and data owners because screen labels and business activities need not coincide. Independent source evidence should be preserved when an operational explanation differs from recorded facts.
A protocol compares event IDs and audit traces with source correspondence, documents, financial records, or provider confirmations. A dashboard and a replica of the same database are not independent validation sources. Where only one underlying record exists, the conclusion can concern consistency of representations rather than external corroboration.
Validation should produce a reference dataset with review dates, evidence pointers, decisions, and reasons. Review findings can refine extraction rules but should remain separable from the evaluation sample. Correcting an extraction using the same files later cited as independent validation gives an overly optimistic result unless a fresh sample is used.
Sampling and reviewer agreement
A review sample should address common cases, flagged deviations, uncertain linkage, rare activities, severe claims, and cases near decision thresholds. Probability sampling supports design-based estimation; purposive sampling supports discovery of failure modes without directly yielding a population rate.
Suppose a synthetic population contains 100 flagged and 900 unflagged claims. Review of 60 flagged cases confirms 42 deviations; review of 90 unflagged cases identifies 9. Under representative simple random sampling within strata and accurate outcome assessment, the estimated population rate is 0.1 × 0.70 + 0.9 × 0.10 = 0.16, or 16%. The unweighted combined sample rate is 51/150 = 34%. Its distortion arises from oversampling flagged cases.
The uncertainty of the estimate depends on design, sample sizes, and review accuracy. Claims affected by one catastrophe or shared supplier network can be correlated, requiring cluster-aware analysis. Small strata may be better described through counts and uncertainty boundaries than unstable detailed percentages.
Independent reviewers can assess a subset of cases, distinguishing disagreement about activity semantics, time, identity, and normative applicability. Reviewers should record reasons for reconciled labels. Where possible, mask the automated flag during an independent outcome assessment to reduce confirmation bias. Reviewer consensus establishes an agreed expert interpretation; its accuracy still depends on the quality of the file.
Diagnosing deviations
A flag labeled "payment before approval" can represent a genuine unauthorized payment, correctly authorized use of a separate procedure, accounting-effective dates inconsistent with transaction order, omitted approval records in a supplier system, or a reversal misidentified as a payment. Each explanation makes different predictions about the source records.
A proportionate investigation confirms the activity and object, timestamp semantics, operative rule, approval evidence, and surrounding financial allocation. It checks whether an exception was applicable and documented. A claim remains inconclusive when material alternatives cannot be distinguished.
Useful descriptive outcomes include an observed model deviation, corroborated operational deviation, documented exception, data defect, and inconclusive finding. These descriptions are working evidence classifications, not a new legal taxonomy. The language of the report should identify whether the investigator confirmed the actual activity or simply found a difference in the data.
File confidentiality and auditability
Claim files can contain medical information, information about third parties, litigation material, or commercial secrets. Validation requires controlled access, a defined purpose, and appropriately restricted extracts. A dashboard generally needs a coded finding and traceable provenance reference, not unrestricted access to the full file.
A reviewer can record verified propositions under a pseudonymous claim identifier. When restricted or privileged material is necessary, authorized specialists should conduct the review or record the limitation. Failure to verify a procedure from the available file is a different conclusion from proof that it failed.
A complete evidence trail should enable an authorized second reviewer to recover the exact records used to classify an anomaly. This is crucial where a finding could influence disciplinary decisions, supplier relationships, or the handling of individual benefits. Investigation should include correction mechanisms for source records and analytical classifications.
Cross-Line Process Comparison and Reproducibility
Insurance lines and natural analytical units
Claims structures vary with contractual obligations and institutional arrangements. Property claims can involve inspections, repairs, estimate revisions, and supplements. Motor claims may contain property damage and personal injury exposures. Liability claims may involve defense and litigation. Health insurance can include provider submissions, adjudication, and resubmissions. Life claims may require beneficiary verification and evidence of death. Parametric coverage may depend on an external trigger observation.
A valid comparison begins with a common question. Notification to first substantive decision may be comparable if the milestone has equivalent meaning. Notification to closure can reflect substantially different obligations, including litigation and continuing benefits. Ranking mean closure time across lines can principally measure differences in case composition.
| Context | Natural object | Observation risk |
|---|---|---|
| Property damage | Claim, estimate, repair item | Supplements and outsourced inspection |
| Motor and liability | Claim, exposure, defense matter | Multiple claimants and external litigation |
| Health reimbursement | Bill, service line, adjudication | Resubmission and batch settlement |
| Life and disability | Benefit request, eligibility period | Continuing benefits and reassessment |
| Parametric insurance | Trigger, measurement, payment | External observations and corrections |
| Catastrophe claims | Claim, event cohort, repair dependency | Shared interruption and correlated demand |
For long-tail claims, comparison should separate first response, interim decisions, first payment, final settlement, closure, and reopening. Cohort maturity affects each measurement. Harmonized vocabularies remain useful when line-specific meanings are recoverable.
An appropriate cross-line design can define a small set of common milestones, supplement them with line-specific state histories, and report which populations meet common inclusion criteria. Comparisons become less interpretable when a milestone has a different substantive meaning in each line.
Documented insurance applications
The Queensland University of Technology BPM Group's 2014 Suncorp case summary describes an Australian insurance project investigating unstructured claims handling and improvement opportunities for relatively simple claims with lengthy processing. Its public summary establishes the setting and purpose; the underlying data and linked full report were not examined here.
Fluxicon's Curaçao claims case summary (2021) describes analysis of reimbursement claims relating to childbirth care from 2018 through 2020. Its significance for this study is the different analytical object: administrative reimbursement processes rather than physical-loss adjustment. An administrative event should not automatically be interpreted as the time of the clinical service.
Khayatbashi et al. (2025), in an object-centric insurance automation study at If P&C Insurance, describe human and AI-assisted identification of claim parts, object-centric log construction, and validation of selected findings against source systems. Their case illustrates the need to represent multiple actor and object relationships during operational change. An industry-wide causal effect cannot be inferred from this setting alone.
These applications illustrate distinct research questions: unstructured handling, reimbursement pathways, and interaction between human work and automated assistance. Their heterogeneity prevents treating them as estimates of a shared effect. They establish examples of process-mining applications in differing insurance contexts.
Reproducible benchmarks and synthetic tests
The IEEE Task Force's event log directory includes public datasets from domains such as loan processing and hospital operations. Such datasets can support algorithm evaluation and implementation tests. They do not establish the correctness of insurance-specific event semantics or the effectiveness of claims interventions.
A reproducible project should document extraction version, analytical cutoff, vocabulary, object relationships, time rules, uncertainty treatment, filters, and software parameters. Where source data are restricted, a public synthetic fixture can preserve the test property required to assess particular transformations. A synthetic example can test the response to duplicated shared events or uncertain temporal order without revealing or estimating a real portfolio's behavior.
A useful test set contains expected traces and selected perturbations: remove an event, duplicate a retry, replace a precise time with a date, merge case IDs, split a shared payment, or insert a reversal. Observed effects on counts, duration, and conformance reveal sensitivity to the introduced perturbation. The test must identify the known ground truth and expected output before execution.
Reproducing a map includes reproducing its denominators and projection. Two systems may produce visually identical diagrams while counting shared events differently. Two maps may differ because of different activity abstractions rather than substantive operational differences. The analysis should specify which results are stable across relevant transformations and which are contingent.
Performance Measurement and Causal Interpretation
Milestones, elapsed time, and waiting
Claims performance involves acknowledgement, decision time, payment, unresolved inventory, repeated work, communication quality, decision accuracy, and administrative cost. Some measures can be derived from event logs, while others require direct file or outcome evidence.
Elapsed time between recorded milestones is their difference under a specified calendar. Business hours, weekends, local holidays, and suspended periods require a documented calendar convention. Calendar time remains important for claimants. Working time may better describe internal capacity. A period labeled "awaiting evidence" may include responsibilities for follow-up and cannot automatically be excluded from every service measure.
Service effort requires start and completion events plus appropriate information about pauses and shared work. The elapsed time between task assignment and completion usually contains queueing. Two parallel tasks each requiring five hours can consume ten staff-hours during five hours of elapsed time. The sum of individual task durations cannot generally be interpreted as the claimant's waiting time.
A study should report claimant-facing elapsed time and resource effort separately where possible. An intervention shifting waiting from an internal queue to a provider queue can improve internal metrics without improving the claimant's experience.
Censoring and competing outcomes
A claim still open at the study cutoff has a right-censored eventual time to settlement. Excluding such cases changes the population under consideration. A completed-case analysis describes completed cases and needs to state that conditioning.
Consider five synthetic claims notified at time zero and followed for 30 days. Three settle after 4, 8, and 18 days, and two remain unresolved at day 30. The completed-case mean is (4 + 8 + 18)/3 = 10 days. The observed fraction settled by day 30 is 3/5 = 60%. The average time spent unsettled during the first 30 days is (4 + 8 + 18 + 30 + 30)/5 = 18 days, assuming complete follow-up and no competing events. It answers a restricted-duration question without assigning eventual settlement times.
Studies with varying entry dates and censoring can use appropriate time-to-event methods under stated assumptions. Where denials, withdrawals, and settlements preclude one another for the chosen endpoint, treating competing final outcomes as ordinary loss to follow-up can be inappropriate. The outcome must specify administrative resolution, monetary settlement, first decision, or another milestone.
Reopening changes the endpoint. Time to first closure, time until last observed closure, total open-state duration, and time to final irreversible settlement are different quantities. The ultimate closure may be unobservable. State histories allow analysis of the observed transitions without assuming every administrative closure is definitive.
Queue analysis and case mix
A directly-follows edge with a high elapsed time is a candidate bottleneck. The duration is not proof of the responsibility of the team performing the adjacent task. It may reflect staffing, coverage uncertainty, supplier queues, evidence availability, catastrophe load, or priority assignment.
In a stable system with consistent population boundaries and long-run averages, the queueing identity L = λW relates average inventory, throughput, and average time within the system. A synthetic steady process completing 20 claims per day with a mean duration of six days implies average inventory of 120 claims. Nonstationarity, backlog accumulation, and inconsistent system boundaries limit direct use of short measurement windows. The identity alone does not locate a remedial intervention.
Case composition can also produce apparent differences. Suppose one unit handles 80 simple claims taking two days and 20 complex claims taking 12 days. Its mean is four days. Another handles 20 simple claims taking two days and 80 complex claims taking 12 days, giving a mean of ten days. Class-specific times are identical. Standardization to equal class proportions produces seven days for both units. The entire aggregate difference arises from case mix under the example's assumptions.
Case-mix adjustment should be appropriate to the objective. Later variables affected by the process, such as the number of follow-up requests, may be mediators rather than baseline measures. Conditioning on them in a causal evaluation can obscure the process effect. Transparent stratification and sensitivity analysis are helpful alongside more elaborate adjustment.
The distribution of elapsed times is often more informative than one mean. Medians and upper quantiles describe different claimant experiences. Small high-severity subgroups require care because a few cases can dominate the average. Counts, censoring, and confidence intervals appropriate to the design should accompany quantitative comparisons of real portfolios.
Intervention evaluation
Process mining supports hypothesis formation by locating waits, repeat validations, and variations in pathways. A causal intervention inquiry requires a treatment definition, assignment mechanism, outcome, eligible population, and follow-up. Removing a review or automating an activity can change timing, decision quality, and downstream risk simultaneously.
Suppose claims assigned to specialist review take longer than others. Severity, disputed liability, coverage complexity, and incomplete documentation may influence both review assignment and delay. Crude comparison cannot isolate the effect of specialist review. Randomized or plausibly exogenous assignment may support stronger evaluation when operationally and ethically permissible; observational designs require explicit identifying assumptions.
A before-and-after automation comparison is vulnerable to changes in demand, product mix, staffing, logging, and seasonal patterns. Revised logging may increase observed rework even when actual rework falls. The evaluation must assess measurement consistency.
Consequently, observed associations, corroborated operational explanations, and identified causal effects should be reported as separate kinds of findings. Faster payment coupled with increased reversals could represent a transfer of work to downstream stages. A substantive assessment needs quality and financial outcomes, not only speed.
Claims Governance, Privacy, and Operational Deployment
Supervisory and contractual relevance
International insurance governance gives claims handling significance beyond productivity. IAIS ICP 19.10 on claims handling establishes supervisory expectations for timely, fair, transparent handling and includes guidance on documented procedures and claimant communication. ICP 19.11 addresses complaints separately. These are international supervisory standards whose binding implementation depends on jurisdiction and institutional setting.
In the United Kingdom, FCA ICOBS 8.1.1R imposes requirements within its scope concerning prompt and fair handling, guidance and progress information, unreasonable rejection, and settlement after agreement. The analyst must translate the applicable rule into a carefully delimited observable question. A single elapsed-time threshold cannot represent all of these obligations.
A recorded decision and payment time can support analysis of timeliness. The substantive correctness of the decision depends on contract, evidence, and law. A prompt refusal may be unreasonable, and lengthy investigation may be justified. Administrative conformance and substantive fairness have overlapping but distinct evidentiary requirements.
Control evaluation needs a rule register specifying authority, jurisdiction, effective dates, product applicability, event definitions, exceptions, calendar conventions, and escalation arrangements. A rule change should produce a new evaluation version. Historical claims should be examined against the requirements and facts applicable to them at the relevant time.
The interests of claimants may diverge from a purely administrative optimization objective. Minimizing document requests can reduce burdens but may leave essential evidence unavailable. Maximizing early closure can obscure unresolved issues. Process metrics should support a balanced investigation of time, quality, accessibility, and correction.
Privacy and restricted processing
Article 5 of the European Union's General Data Protection Regulation (2016) establishes principles concerning purpose, minimization, accuracy, retention, security, and accountability. Where applicable, these constrain collection, transformation, access, and retention of claims-event information. Lawful basis, special-category conditions, and other obligations require a separate legal assessment.
A study of cycle time may need pseudonymous identifiers, event times, workflow classifications, product line, and broad severity information rather than names or medical histories. Indirect attributes can nevertheless combine to identify individuals. Pseudonymization reduces risk but does not establish that personal data have been anonymized.
Access can be separated by function. Extraction teams maintain controlled link keys; analysts use restricted event datasets; authorized reviewers inspect case files; managers receive aggregated indicators and specific findings when justified. Access history and retention decisions form part of accountability.
Commercial claims may also involve confidential pricing, supplier terms, and litigation material. Cross-organization event sharing must establish rights and purposes for each data field. An interoperable exchange format does not create authorization to disclose or process data.
Automated event construction and oversight
Automated classification, recommendations, and suspicious-path flags can direct investigation. They require evaluation against independently reviewed evidence appropriate to the intended purpose. A classifier trained to imitate historical labels can reproduce coding errors as well as genuine activity patterns.
The If P&C example illustrates a configuration in which human and automated work coexist. An implementation should separately record model version, invocation, output, uncertainty, human review, overrides, and final operational actions. Without that distinction, the process map can assign an event to the wrong participant or omit a decisive human review.
Atypical pathways warrant proportionate investigation rather than automatic adverse decisions. Feedback should distinguish confirmed control failures, corrections, permissible exceptions, and inconclusive flags. Training on unreviewed flags can amplify earlier classification errors.
Decision-support systems should disclose which part of a finding comes from direct observations and which part comes from inference. The ability to trace a recommendation to evidence is important for challenge, correction, and review. An operational process reconstruction should not substitute for contractual or legal assessment where an individual entitlement is at stake.
Monitoring and drift
Process mining operates on changing information systems and workflows. New products, migrations, integrations, reorganizations, and reporting practices alter event frequencies and meaning. A surge of unrecognized activities may indicate a new pathway or a changed code dictionary. Monitoring should therefore include the quality and coverage of data sources.
Deployment should preserve versioned extraction logic, vocabularies, crosswalks, models, control rules, evaluation parameters, and release records. Every report needs an analytical cutoff and the identity of the relevant data and model versions. Later reconstructions should preserve historical comparisons while allowing explicit corrections.
If data coverage deteriorates, monitoring can report an unavailable or indeterminate result. Treating absent flags as good performance under incomplete observation creates a false assurance. Required evidence strength should reflect the consequences of misclassification.
The ability to trace a reported exception to source evidence is also necessary for remediation. The identified cause may be a logging defect, policy mismatch, supplier handoff, capacity limitation, or actual procedural failure. A remediation plan needs an independently supported description of which problem occurred.
Research Limitations and Integrated Analytical Synthesis
Reconstruction as an identification problem
This research integrates event semantics, object identity, partial orders, conformance, case-file validation, and claims governance. Its central analytical interpretation is that a reconstructed trace is an inference conditioned on an observation and transformation procedure. The quality of inference depends on whether materially different operational histories can generate the same observations.
Consider two histories. In the first, a required review occurs but its audit event is lost. In the second, it is omitted. The remaining records are identical. The observed trace alone cannot distinguish the histories. Independent documentation or a validated observation mechanism is required to identify which explanation applies. This is an application of ordinary identifiability reasoning, not a new theory of claims operations.
The same argument applies to times and objects. Two linkage assignments may yield different maps while preserving the event total. Two date-level activities can occur in either order while preserving logged dates. Different allocations can produce identical batch amounts. The scientific response is to seek discriminating evidence, restrict conclusions to supported invariants, or disclose the alternatives.
The practical consequence is a distinction between observed and reconstructed information. Direct records, inferred assignments, constraints, and conclusions should remain traceable as separate elements. A final visual diagram often suppresses these distinctions for accessibility; its accompanying data and report must preserve them.
Alignment of institutional and analytical questions
A claims-process project should start with a concrete proposition. If the objective concerns notification timing, specify which notice was received, by whom, and when. If it concerns payment authorization, define the controlled transaction, amount threshold, operative policy, and relevant evidence. If it concerns repeated evidence requests, distinguish repeated communication from genuinely different requests.
Three relationships govern interpretation: the event-to-object relationship defines the analytical unit; the event-to-time relationship defines supported chronology; the event-to-rule relationship defines the scope of normative comparison. Provenance and validation connect each to the underlying documentary evidence.
A strong finding can therefore state that within a defined cohort and cutoff, under a documented projection and applicable reference, a specified number of observations support a defined deviation, with other observations classified as exceptions or indeterminate. This proposition is independently testable. Broader claims about inefficiency, fairness, or compliance demand evidence that directly addresses those concepts.
Cross-line comparison benefits from definitions that are stable enough for aggregation and specific enough for meaningful interpretation. This balance requires maintaining the original business objects and source semantics even when an analytical projection creates a simplified common vocabulary.
Research limitations
The literature search targeted the architecture-approved questions and selected insurance applications. It was not exhaustive across languages, databases, or jurisdictions. Blocked publisher pages were replaced with institutional author abstracts where available, and their use was restricted to supported propositions. The IEEE standard summary was examined rather than the full licensed text. The Suncorp and Curaçao summaries were examined rather than their linked reports or underlying records. Author manuscripts and official regulatory text were used for the propositions identified in the evidence ledger.
The regulatory comparisons distinguish international supervisory standards, UK conduct rules, and EU data-protection principles; they do not constitute a comprehensive legal analysis of any jurisdiction. The cases and arithmetic examples with invented values demonstrate methodological effects. No real insurer log was collected, no process discovery algorithm was executed on a real claims population, and no original field estimate is claimed.
The authorial contribution consists of domain-specific analytical integration, worked counterexamples, formulation of operational validation tasks, and interpretation of the relationship between process evidence and claims administration. Further empirical validation requires an authorized multi-source dataset, independent file review, tested event mappings, and a prespecified evaluation design. This would constitute a separate empirical research project.
Conclusion
Insurance claims process mining studies the documented pathways through which insurance claims are administered across interacting organizational and technical systems. Its scientific value derives from the relationship between business occurrences, source records, and analytical event representations. Supported definitions, traceable provenance, and appropriate validation determine the quality of the evidence.
Claims events frequently concern multiple objects. The selected case determines the trace projection, while the corresponding financial and organizational measures retain their natural aggregation levels. Object-centric representations can preserve relationships obscured by a single case identifier. Both approaches require explicit definitions and independently checked mappings.
Reliable chronology depends on source clock semantics, precision, and supported dependencies. Partial orders and interval bounds preserve information lost by arbitrary total ordering. They help distinguish demonstrated conformance, possible conformance, and inconclusive observation.
Process discovery, reference-model conformance, and normative assessment answer complementary questions. Discovery describes recorded behavior; a reference defines expected or permitted behavior; case-file validation establishes what a recorded deviation means. The causal effect of an operational change requires a separate identification strategy.
The international significance of this work lies in connecting operational analysis with accountable claims handling. Reliable measurement must preserve the claimant's experience, evidence quality, the applicability of rules, and protection of sensitive information. Insurance process mining becomes a scientific practice through the continuing identification, reconstruction, validation, and responsible interpretation of claims-administration evidence.
References
Bayomie, D., Di Ciccio, C., & Mendling, J. (2022). Event-case correlation for process mining using probabilistic optimization. arXiv:2206.10009. Author manuscript
Berti, A., Koren, I., Adams, J. N., Park, G., Knopp, B., Graves, N., Rafiei, M., Liß, L., Tacke Genannt Unterberg, L., Zhang, Y., Schwanen, C., Pegoraro, M., & van der Aalst, W. M. P. (2024). OCEL (Object-Centric Event Log) 2.0 specification. arXiv:2403.01975. Specification version 2.0, dated October 16, 2023. Specification
Buijs, J. C. A. M., van Dongen, B. F., & van der Aalst, W. M. P. (2014). Quality dimensions in process discovery: The importance of fitness, precision, generalization and simplicity. International Journal of Cooperative Information Systems, 23(1), 1440001. DOI: 10.1142/S0218843014400012. Institutional record
De Weerdt, J., & Wynn, M. T. (2022). Foundations of process event data. In W. M. P. van der Aalst & J. Carmona (Eds.), Process Mining Handbook (pp. 193–211). Springer. DOI: 10.1007/978-3-031-08848-3_6. Chapter
European Parliament & Council of the European Union. (2016). Regulation (EU) 2016/679 of 27 April 2016 (General Data Protection Regulation). Official Journal of the European Union, L 119, 1–88, Article 5. Official text
Financial Conduct Authority. (n.d.). Insurance: Conduct of Business Sourcebook (ICOBS), chapter 8, Claims handling, rule 8.1.1R. Accessed November 19, 2026. Handbook
Fluxicon. (2021). Process mining obstretical care claims data [title as listed]. IEEE Task Force on Process Mining case-study page, updated January 18, 2022. Case summary
IEEE. (2023). IEEE Standard for eXtensible Event Stream (XES) for Achieving Interoperability in Event Logs and Event Streams. IEEE Std 1849-2023. Standard description
IEEE Task Force on Process Mining. (n.d.). Process mining event logs. Official dataset directory. Directory
IEEE Task Force on Process Mining. (n.d.). Process Mining Manifesto. Official project page. Manifesto page
International Association of Insurance Supervisors. (n.d.). Insurance Core Principles and Common Framework for the Supervision of Internationally Active Insurance Groups. Online tool, ICP 19, standards 19.10–19.11 and associated guidance. Accessed November 19, 2026. Official standards
Khayatbashi, S., Sjölind, V., Granåker, A., & Jalali, A. (2025). AI-enhanced business process automation: A case study in the insurance domain using object-centric process mining. arXiv:2504.17295. Author manuscript
Leemans, S. J. J., Fahland, D., & van der Aalst, W. M. P. (2013). Discovering block-structured process models from event logs: A constructive approach. In J.-M. Colom & J. Desel (Eds.), Application and Theory of Petri Nets and Concurrency (pp. 311–329). Springer. DOI: 10.1007/978-3-642-38697-8_17. Institutional record
Moreau, L., & Missier, P. (Eds.). (2013). PROV-DM: The PROV data model. W3C Recommendation, April 30, 2013. Official text
Pegoraro, M., Uysal, M. S., & van der Aalst, W. M. P. (2020). Conformance checking over uncertain event data. arXiv:2009.14452. Author manuscript
Queensland University of Technology BPM Group. (2014). Process mining at Suncorp: Reduction of the processing time of claims in Australia's largest insurance group. IEEE Task Force on Process Mining case-study page, updated January 6, 2020. Case summary
Suriadi, S., Andrews, R., ter Hofstede, A. H. M., & Wynn, M. T. (2017). Event log imperfection patterns for process mining: Towards a systematic approach to cleaning event logs. Information Systems, 64, 132–150. DOI: 10.1016/j.is.2016.07.011. Institutional record
van der Aalst, W. M. P., Adriansyah, A., & van Dongen, B. F. (2012). Replaying history on process models for conformance checking and performance analysis. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 2(2), 182–192. DOI: 10.1002/widm.1045. Institutional record
van der Aalst, W. M. P., & Santos, L. (2021). May I take your order? On the interplay between time and order in process mining. arXiv:2107.03937. Author manuscript
van Eck, M. L., Lu, X., Leemans, S. J. J., & van der Aalst, W. M. P. (2015). PM²: A process mining project methodology. In J. Zdravkovic, M. Kirikova, & P. Johannesson (Eds.), Advanced Information Systems Engineering (pp. 297–313). Springer. DOI: 10.1007/978-3-319-19069-3_19. Institutional record