Workers' comp fraud costs Americans more than $5 billionannually, threatens the jobs of Americans, and hurts employers somuch that in some cases, companies go out of business or are forcedto move. It adds 10 cents to every dollar of premiums. It'spervasive, growing, and often very hard to detect.

In the past several years, rising premium rates for workers'compensation insurance have increasingly pressured both insurersand employers. Despite workers' comp reform legislation that hasbeen enacted recently in states such as California, across thenation medical costs continue to grow significantly despite adecline in the number of claims filed. In a 2004 study by theNational Council on Compensation Insurance (NCCI), the medicalshare of total benefit costs in workers' compensation rose toapproximately 55 percent on a countrywide basis, with someindividual state shares approaching 70 percent.

Because a significant percentage of claims start off aslegitimate, workers' comp fraud and abuse is hard to spot. Much ofthis is abuse, such as a case in which a claimant is malingering,delaying their return to work. Outright fraud is less frequent, butoften costs payers dearly. The longer it takes to discover afraudulent claim, the more money is paid out. That's why earlydetection of fraudulent and abusive claims is critical tocontaining the cost of workers' compensation. And the more insurerscan decrease their losses, the more likely it is that theirinsurance rates will be lower as well.

Claimant fraud comes in many guises. Some workers fake injuriesat the workplace to get paid for staying home. Some exaggerate theextent of injury to prolong time away from work, while others claimtheir injuries occurred at work, when, in fact, they happened offpremises and are unrelated to work. In extreme cases, fraud is theresult of organized crime, or collusion with un-scrupulous doctors,therapists, and attorneys.

How prepared are insurance payers to detect fraud and abuse? Asurprising number of claim departments still rely on manualdetection processes, which simply are not sophisticated enough toidentify the many patterns and types of claimant fraud. Adding tothat is the volume of cases on any adjuster's desk. With anindustry average of 250 claims at any time, adjusters are often toooverwhelmed to perform the detailed analysis needed to find complexpatterns that indicate fraud on top of their growing caseload.

According to the Coalition Against Insurance Fraud, 20 percentof total fraud at most is detected, and much of this fraud isdetected late in the life of the claim. This late discoverydramatically increases the total cost on payers. In addition, ahigh percentage of claims that are referred to SpecialInvestigative Units (SIUs) often are not fraudulent leads. The timeand cost to investigators to pursue suspicious claims that don'tturn out to be fraudulent or abusive contribute to the high cost ofworkers' comp.

Compounding this, a Coalition Against Insurance Fraud studynoted that nationwide, fraud bureau budgets and head-counts havedeclined significantly in the past three years, while fraudreferrals to SIUs are growing. With a growing volume of fraudreferrals per investigator, properly identifying the patternstypical of truly fraudulent or abusive cases (or exceptional butnot fraudulent claims needing aggressive case management) hasbecome a huge challenge.

Greater Predictive Power

That's why insurance payers are turning to highly sophisticatedanalytic software called predictive models. Adjusters can spendsignificantly less time reviewing and processing claims, whilespotting truly suspicious claims faster with greater accuracy.

Often based on a highly advanced software technology known asneural networks, predictive models can enable insurers to identifyfraudulent, abusive and high-risk claims much earlier and with ahigher degree of accuracy than any other known method. It can dothis while swiftly and accurately processing the majority of claimswithout adjuster intervention.

Predictive models are largely responsible for the dramaticturnabout in fraud detection in the credit card industry. Today,predictive models are used to screen 85 percent of U.S. credit cardtransactions for fraud, resulting in a 50 percent reduction inindustry losses. This same technology, used so successfully by 65percent of the world's credit card issuers, is now increasinglybeing deployed to contain the spiraling costs of workers' comp.

In seconds, predictive models can scan thousands of dataelements simultaneously to find subtle, complex, and hiddenpat-terns of suspicious behavior. Their analytical and processingstrength enables high-volume claim departments to perform arigorous, objective review of every claim. While human experts arecapable of identifying some red flags and simple fraud patterns,sophisticated modeling techniques are required to find more complexpatterns of fraud. Based on historical examples already deter-minedto be fraud and abuse claims, neural network predictive models canlearn which subtle patterns are associated with a high likelihoodof fraud and which are not.

As claims evolve, they leave a trail of data. This data oftenshows up as first reports, followed by payment transactions thatindicate workers' compensation payments, medical services rendered,vocational rehabilitation, and other important claim events. Thesepatterns of activity provide the raw material that is used by thepredictive model to score each claim from 1-1,000. Normal claimswill typically have scores of 300 or less. But a high score is asignal that the claim is out of profile when compared to its peers.Each time the predictive model scores a claim, reasons are producedto explain the model score. This can provide useful information toassist in the analysis of a claim being suspected of beingfraudulent, abusive, or exceptional by looking at the reason codesreturned with the score.

Predictive models are accurate because the technology recognizespatterns from the data itself, not from pre-existing assumptions onwhat the data means. As a result, the system provides high-qualityreferrals to investigative units, further reducing losses. Inaddition, neural networks also are the only systems sophisticatedenough to detect fraud types that have not been seen before. Claimsare scored frequently to detect any change in the status of theclaimant.

Finding Claims that Matter

Interestingly, only a small percentage of claims account for themajority of claims costs. These more costly cases includefraudulent, abusive, as well as legitimate claims that for variousreasons require special handling. These exceptional situationscould include a claimant who is not acting fraudulently or showingabusive behavior, but instead is someone in need of aggressive casemanagement. Predictive analytics can help insurers identify thesehigh-risk claims rapidly, allowing these exceptions to be routed toexperienced case managers or investigative units, freeing adjustersto process the remaining claims with no outside re-sources.

The trick is separating these high-risk claims from the rest.Many fraudulent or abusive claims don't look all that remarkable atfirst, even to the eye of a well-trained adjuster. Clues may besubtle and submerged in an ocean of data. In the case of a bodilyin-jury claim, for example, medical factors indicating a need forspecial handling may become evident only after some amount oftreatment. In the case of fraud, opportunists who initially filelegitimate claims may eventually fall prey to the temptation toexaggerate or misrepresent their cases. In fact, the majority ofinsurance fraud starts out this way.

Another factor contributing to high claim costs is that presentmethods are skewed toward apprehending fraud rather thanidentifying other types of high-risk claims. Predictive analyticsare highly effective at reducing claims costs because they areextremely effective at identifying these high-risk exceptionclaims, both legitimate and fraudulent. By accurately culling outhigh-risk claims, predictive analytics can make it practical andsafe for insurers to process and close a vast majority of claimsfaster. Insurers save money by identifying suspicious and high-riskclaims at the earliest possible moment, enabling preemptive actionto be taken to prevent losses from occurring.

Advantages of Predictive Models

Predictive models don't replace the skills and experience of anadjuster. Instead, the power of predictive analytics augments thework of adjusters to work more efficiently and effectively. In afraction of a second, predictive models can consider thousands ofvariables simultaneously, looking at complex relationships betweendata and deciphering subtle clues that might be missed by even themost seasoned adjuster.

These advanced methods are applied not only to the initial claimdocuments, but also to every transaction associated with the claimover its entire lifecycle. Each new piece of data coming in isanalyzed thoroughly against not only the claim history, but a vastdatabase captured from industry claims. In addition, predictiveanalytics are accurate because they are objective. Predictivemodels can recognize patterns from the data itself, not assumptionsabout it. This pattern recognition capability is dynamic; when thedata indicates something new, the software updates its detectioncriteria.

Early detection is just one of the total benefits of deployingpredictive models. Armed with evidence of fraudulent or abusiveactivity, claim adjusters can be proactive in managing claims moreeffectively. Often, fraud or abuse can be stopped quickly by a callor letter to the claimant, helping them understand that theiractivity is being monitored. Once contacted politely, malingeringinjured workers often stage miraculous recoveries from theirinjuries. This preemptive effect also prevents claim costs fromgrowing. In the case of an injured worker whose activities are notfraudulent or abusive, claim adjusters can proactively partner withcase managers to provide the care needed for better outcomes and afaster return to work.

Better Bottom-Line Results

Overall, insurers using predictive analytics software technologyto detect fraudulent and abusive claims have experienced a re-turnon investment of 20-to-1, or up to $300 per claim in savings. Incomparison tests, 55 percent of claims were identified by modelsweeks or months before they were discovered manually, and thesoftware often discovered suspicious claims or claims needing casemanagement that would have been missed by insurance claimsanalysts.

With the advent of predictive software models, insurance payershave a powerful tool in the fight against fraud. Losses andadministrative expenses that were previously accepted as a normalcost of doing business can now be substantially reduced. Predictivemodels allow claim managers to make better decisions earlier, bemore productive, and proactively prevent further fraud and abusebefore it happens. The proper case plan can be established for eachclaim, including referral to SIU or special case handling.

It's an exciting advance in the fight against one of the largestchallenges facing claim managers today. With predictive analytics,insurance adjusters have a powerful tool that can dramaticallyimprove the climate of workers' compensation.

Kevin Lisle is product manager of property and casualtyanalytics at Fair Isaac Corporation. Contact In-formation:949-655-3300, www.fairisaac.com, email [email protected].