Editor's Note: The following article is a collectivecontribution of Kevin M. Bingham, Frank Zizzamia, Jim Guszczaand Kirsten Hernan, all of whom work at Deloitte.

Henry Ford is credited with saying, “half of my advertisingbudget is wasted; I just don't know which half.”

Ford might sympathize with insurance claims adjusters, who faceeven more pressing uncertainties. In the workers' compensationdomain, for example, it is well known that 20 percent of the claimsultimately account for 80 percent of the total claims costs. Thisis true of both lost-time and medical-only claims.

So Which 20 Percent?

Traditionally, it has been virtually impossible for the adjusterto know which claims make up the 20 percent early in the claim lifecycle. But this has changed, thanks to advances in claimspredictive modeling methods. Leading claim predictive models canserve as early warning systems, flagging at an early stage thoseclaims most likely to grow large and complex. This enables anadjuster to prioritize his or her workflow and make effective useof special investigative unit (SIU) referrals, medical management,and return-to-work programs. A key advance has been a moreeffective use of injury groups in the claims predictive modeldesign.

Claims predictive modeling in workers' compensation has beenaround for many years, with first-generation modeling effortstypically focused on prospectively identifying the 20percent of claims that drive 80 percent of costs as earlyas possible in the life cycle of the claim. Ideally, theinsurer's goal is to develop such a modeling capability thataccomplishes this goal at first notice of loss(FNOL). However, a major limitation of typicalfirst-generation claim predictive models is that they have tendedto pool the types of injuries together and segment claims alongfairly obvious lines such as injury type. For example, asprain of a worker's lower back is on average more severe than acontusion of a worker's arm. While this is a valid approach, muchmore refined (and economically beneficial) solutions arepossible.

It turns out that the 80/20 rule also applies withineach injury type. For example, 20 percent of the workerswho sprain their lower backs will typically drive 80percent of the total claim costs for all sprains of lowerbacks. Herein lies the opportunity for more effective workers'compensation predictive modeling.

Next, we will describe how International StatisticalClassification of Diseases and Related Health Problems (ICD-9)codes can be classified into workers' compensation-specificgroups. This grouping – used in today's more advancedpredictive model designs – can lay the foundation for identifyingthe most severe claims within each injury grouping[i].

The Claims Lifecycle Before Leveraging AdvancedAnalytics

Over the life of a claim, the information available at aspecific point in time has typically determined how claimprofessionals assess the complexity of workers' compensationclaims. At FNOL this information is quite limited, which iswhy more advanced workers' compensation claim predictive modelsincorporate external data sources and synthetically createdvariables. For loss time claims and more severe medical onlyclaims, the three-point contact with the injured worker, employer,and physician provides additional information that further enhancesthe claim professional's understanding of claimcomplexity.

Even at this stage in the claim's life cycle, it is difficult todistinguish the difference in severity from one back strain toanother, except for some differentiating claimant characteristicssuch as age, tenure, prior claim history, and specific detailsabout the circumstances of the accident—that is, injuryoccurred in the office, injury occurred using heavy machinery,injury occurred from a height, and so on. More often than not, itis still difficult for the claim adjuster to identify the highseverity back sprains without the use of external data andsynthetic variables.

Over time, additional information such as the receipt of medicalbills, pharmaceutical information, and the claimant's medicalhistory begins to paint a much fuller picture of the claim's truecomplexity. However, the opportunity to favorably impact theclaim through early intervention, SIU involvement, and experiencedresource triage may already be lost if it takes nine to 12months for the adjuster to correctly identify the true severity andcomplexity of a claim.

This no longer needs to be the case, however. Employingworkers' compensation claim predictive models that build uponexternal data sources, text mining capabilities, and insightfullyspecified synthetic predictive variables changes thegame.

The Evolution of Claims Analytics

It is helpful to view the evolution of workers' compensationclaims predictive models on analogy with the evolution of workers'compensation underwriting models. Early efforts to segment policieson the basis of expected profitability were heavily dominated bybusiness class code, with other predictive dimensions makingsecond-order contributions. This resulted in models that for themost part told the underwriter what he or she already knew: forexample roofers tend to be less profitable than florists. While itis not a bad thing to refine and further quantify the underwriter'sprior knowledge, the real business value—and the real intent ofpredictive modeling in this domain —has been identifying thepockets of, for example, profitable roofer risks and yes,unprofitable florist risks.

In short, these models enable the underwriter to evolve beyondprimarily class-based underwriting to underwriting based on arefined, multidimensional view of each risk. This approach toinsurance underwriting has been a major success story over the pastdozen years, and today many major US commercial insurers havecultivated their ability to develop, deploy, and continually refinetheir underwriting models. Equations are now routinely used in adomain that had previously been largely judgment-driven as a matterof practical necessity.

Fast forward to today, and we see a similar untapped potentialin the workers' compensation claims modeling domain. Just as earlyunderwriting models had been largely class-driven, today'sfirst-generation workers' compensation claims models tend to belargely injury type-driven.

And analogously with the underwriting application, the realbusiness benefit of claims modeling is not capturing the fact that(for example) contusions are on average less severe than lower backsprains. The adjuster already knows this. Rather, the intent ofclaims predictive modeling is to provide laser-like segmentationwithin each injury type. For example, certain lower backsprains would be good candidates for straight-through processing.On the other hand certain contusions – given the specificcombination of age, comorbidities, and other case-specific riskfactors – might be sufficiently complex to warrant assigning to aclaim adjuster.

In short, next-generation claim predictive models capture thevariation in severity that exists within each injury type. From amodeling point of view, this is a technical challenge. From abusiness point of view, it is an opportunity. Failing to capturewithin-injury type variation is an instance of what the businessstatistician Sam Savage calls “the flaw of averages”. The benefitof large-scale predictive modeling initiatives is to move beyondgroup averages. Just as data-driven retailers now treat customersas an individual rather than members of various consumer segments,claims predictive models should evolve to account for – but alsosee beyond – injury type.

To illustrate, the graph to the right below displays thediffering overall average severities for three injury groups.Claims involving shoulder sprains/strains are clearly more severeon average than those involving neck and backsprains/strains. Claims involving non-back and shouldercontusions have the lowest overall average severity of the threeinjury groups.

In order to move beyond telling claims adjusters what theyalready know, it is crucial to leverage the power of ICD-9 injurygroupings to help identify which contusion is going to be worsethan average, and similarly which sprain of the lower back is goingto be better than average.

Injury Grouping: Evaluating Severity Within LikeInjuries

The first step in developing Deloitte's injury groupingmethodology was to research and analyze thousands of the potentialICD-9 codes that pertain to workers' compensation claims. Thisrequired a detailed analysis of medical severity, clinicalreasonability and the resultant workers' compensation outcome tohelp identify the diagnoses that cluster together with similarprojected claim outcomes. By carefully analyzing a largevolume of historical workers' compensation data, we developedapproximately 70 injury groups. The injury groupings weredetermined by combining similar diagnoses with sufficientstatistical credibility to result in stable statistical patterns,aimed at driving enhanced segmentation in the claim predictivemodeling process.

The second step involved developing a high level body partassignment process. This required assigning an initial bodypart for each injury grouping and ICD-9 combination. Part ofthis process involved reviewing the reasonability of theassignments based on medical science, actuarial peer review and theability to capture variation in lost time duration patterns withina physical part of the human body. We identified 28 differentbody part assignments.

The third step in our injury grouping methodology involved theidentification, development, and assignment of medical specialtytreatment profiles to each injury group. This requiredexamining the reasonability of each treatment profile on the basisof medical science, and comparing the treatment profile assignmentswith the AMA Physician Desk Reference. We developed 16different treatment profiles (that is, chiropractic, physicaltherapy, orthopedics, x-ray, and so on) to help us betterunderstand whether a claimant is likely being over-treated orunder-treated at a specific point in time based on the number ofactual office visits compared with medical protocols for theirspecific injury group.

Last, with our ICD-9 methodology thus defined, we realized thatan approach was needed to help identify the prevalent diagnosis ona claim, since there can often be multiple ICD-9 codes referencedin the claim file. The process of identifying the prevailingdiagnosis, or the condition that is primarily responsible fordriving the claimant's claim outcome, is not normally astraightforward exercise. ICD-9 diagnosis information istypically received and stored in the detailed medical billdata. Often there are multiple ICD-9 codes on a single billand the codes can change over time, reflecting current informationabout the claimant's diagnosis. Our ICD-9 injury groupingmethodology includes a method that leverages the complete set ofdiagnosis information available on a claim to determine theprevalent injury group on an individual claimant level.

Enhancing Claim Predictive Models Leveraging InjuryGroupings

With this injury grouping methodology in hand, it was time toimprove upon first-generation claim predictive models by bettersegmenting claims within each of the defined injurygroupings. We used a normalizationtechnique to remove the variation in claim outcomes due to injurygroup, and analyzed patterns in large databases relating claimcharacteristics observable at FNOL with the ultimate claimoutcomes. We followed a rigorous modeling methodology leavenedwith common sense and insurance claims domain knowledge, resultingin a predictive model that produces significant segmentation veryearly in the life of a workers' compensation claim, literally atFNOL. Such predictive models prospectively segmentclaims within such injury groups as Sprains and Strains of the Neckand Back. The results are dramatic: the highest-scoring claimsare 25 times or more costly than the lowest scoring claims.

The graph below displays actual predictive model results basedon real historical workers' compensation claims. The barsmarked “1-10” represent the average ultimate claims severity forthe best-scoring 10 percent of claims. Similarly, the“91-100” represents the ultimate severity for the worst-scoring 10percent of claims. Note also that this graph is based onblind-test “holdout” data that was not used to build the model.This display demonstrates that (a) the model offers a high degreeof segmentation power and (b) this segmentation is not drivensimply by injury type. Within each injury type, the injurygroup-enhanced models identify as early as FNOL those claimsdestined to be the ones with high severity.

The use of injury groups further enhances the model developmentprocess by assisting the claim adjuster with interpreting in excessof 50 claim characteristics all at once. For example, the useof specific medical procedures can vary greatly based on injurytype. Would the presence of five physical therapy treatmentsbe higher or lower than indicated by medicalprotocols? Depending on the answer to this question, wouldthis information drive the outcome of the claim (Hint: it willvary by injury group). For a sprain orstrain, five treatments may be in line with expectations, or evenlower than indicated by medical protocols, depending on the age ofthe claim. But if the injury was a simple contusion, fivetreatments would likely exceed existing expectations. Usinginjury groups in this way, medical protocols can help to draw moreproper conclusions regarding claim complexity, especially whencombined with other variables such as co-morbidities, externaldemographic and synthetic variables.

The Benefits

Injury grouping, prevalence methodology, and normalization-basedinjury group models constitute the cornerstone of the today's workers' compensation claim predictive modeling. Modelsthat use these components can provide key insights beyond what isalready known by claims professionals.

At claim intake, a newly reported workers' compensation claim isscored by the claim predictive model, resulting in immediate andmore effective routing to claims adjusters and case managers withthe appropriate level of experience. Similarly, these predictivemodels can help identify cases for auto-adjudication or fast trackprocesses, as well as those that may require the involvement of theSIU resources. The business impact of such predictive modelsalso extends to supervision and oversight, helping to trigger whenit is appropriate to involve a more experienced supervisor.

The business impact on a claims organization is breakthroughperformance that results from arming the right resources with theleading information and insights to allow them to take theappropriate actions more quickly, and thereby accelerating theclaims life cycle. The results are real. The drivers of thisimprovement range from more effective resource allocation to morefocused claims management strategies, resulting in better claimsoutcomes and bottom-line loss cost savings of up to 10 percent ofan organization's annual claims spend.

In Conclusion

Using an injury group-based approach to developingstate-of-the-art claim predictive models is a realitytoday. Insurers, self insureds, and third party administrators(TPAs) are using this approach to help them better managetheir claim exposures. These models are enabling organizationsto assign the knowledgeable resource to the appropriate claim, atthe required time for targeted intervention. The benefits arebeing measured and leading claims organizations are helping theirinjured claimants get better faster and return to work earlier.

KEVIN M. BINGHAM is a principal at DeloitteConsulting LLP in Hartford, CT and leader of Deloitte's ClaimPredictive Modeling and Medical Professional Liability practices.FRANK ZIZZAMIA is a director at DeloitteConsulting in Hartford, CT and founder of Deloitte's AdvancedAnalytics & Modeling practice. JIM GUSZCZA isthe national predictive analytics lead in Deloitte's Actuarial,Risk & Analytics practice and is an assistant professor ofActuarial Science, Risk Management, and Insurance at the Universityof Wisconsin-Madison School of Business. KIRSTENHERNAN is a senior manager at Deloitte Consulting LLP inPhiladelphia, PA, and a leader in Deloitte's Claims and RiskManagement practice.

This publication contains general information only and isbased on the experiences and research of Deloitte practitioners.Deloitte is not, by means of this publication, rendering business,financial, investment, or other professional advice or services.This publication is not a substitute for such professional adviceor services, nor should it be used as a basis for any decision oraction that may affect your business. Before making any decision ortaking any action that may affect your business, you should consulta qualified professional advisor. Deloitte, its affiliates, andrelated entities shall not be responsible for any loss sustained byany person who relies on this publication.

About Deloitte

As used in this document, “Deloitte” means DeloitteConsulting LLP, a subsidiary of Deloitte LLP. Please seewww.deloitte.com/us/about for a detailed description of the legalstructure of Deloitte LLP and its subsidiaries. Certain servicesmay not be available to attest clients under the rules andregulations of public accounting.

Copyright © 2012 Deloitte Development LLC, All rightsreserved.

[i] Deloitte Consulting LLP has a patent pending with theUnited States commissioner of patents titled Injury Group BasedClaims Management System and Method, Patent No. 61/199,226 filed onNovember 12, 2008.