I got a chance to catch up with Andrea Allmon of FICO to hear about their new Insurance Fraud Manager release (3.0 in July). This is a timely topic because of the debates around healthcare at the moment. In all the discussions about healthcare costs you never hear about the amount of fraud involved, yet healthcare fraud estimates range from 3-10% or $63B – $210Bn! And lest you think this is a result of lots of crooked people, you need to know that this is increasingly about organized crime. Career criminals find health insurance fraud more appealing than dealing drugs – it’s less dangerous and more profitable. It used to be about detecting bills for unnecessary treatments and finding those pushing the boundaries on what was allowed. Now organized schemes are large and are putting patients at risk to create the fraud ring.
This level of fraud vastly exceeds the personal auto, home, workers comp and credit card fraud added together ($20Bn tops)! Add in wasteful treatments (that have similar patterns to outright fraud) and unnecessary emergency room use and you can get numbers in the range of $400-$600Bn annually! All the money you need to reform the industry. Anyway, enough about the problem, what about the solution.
The new release, like previous versions, uses predictive analytics to find potential fraud and abuse. In particular, the application is driven by unsupervised neural nets that look for outliers. Neural nets are widely used in fraud because their ability to find hidden/unknown patterns is good and their opacity (it’s hard to tell why they gave you a particular answer) is less important. Unlike credit card fraud, there are not as many “tags” in healthcare fraud. A tag in this context is a known fraudulent transaction. In credit, fraud is self revealing over time as every transaction lands in either the good or bad bucket. This generates lots of “bads” to use to train the models. But with healthcare fraud this is harder, as providers can commit fraud without being caught (assuming you are not using some kind of fraud detection engine). To compensate, the system uses variables and profiling that are driven towards fraud patterns as well as the neural nets. The same approach also picks up poor quality of care issues, as these have a similar pattern. The models learn from a company’s own data – the design of the models is pre-defined but they become specific to the company because they are self training.
The application is designed to be used on claims that have already gone through the claims adjudication process – where rules can be used to check if it is complete, is for an insured party, is for service covered for this insured etc. The application supports a detect, review, investigate cycle for these claims.
An adjudicated claim is then scored to detect fraud. The application scores a medical bill or a pharmacy claim after adjudication but prior to payment to create a fraud risk score. Each line of the claim is scored separately. The system then ranks the claims most at risk and outputs the claims, the scores, the reasons for the problem and the contextual information. The system queues up transactions with a score above the min level set by the insurer and orders them based on fraud score and some other criteria such as the value of the claim and by type of claim. Criteria that drive a need for medical expertise can also be used to ensure medical staff reviews the claim. For instance, a J code means a claim is for an injectable drug and these typically need medical review to see if doses are being misstated etc. Besides detecting fraud and abuse in claims, other problems are found by the engine such as duplicate claims with minor date changes for instance. Because 60% of claims still come in on paper(!) there are more of these problems than you would expect.
The application also scores the entities involved – providers, pharmacies, labs, doctors (and soon dentists, long term care organizations and patients who go to many people for prescriptions etc). These are rescored monthly without human intervention. Finally, the application profiles providers who share patients and scores patients and providers together as pairs.
To process the fraud cases, the application includes end users investigation/workflow management, case management and reporting. The system can manage claims identified as fraud and support investigations that have not yet been tied to specific claims – lots of tips come in and must be tracked, prioritized and investigated to see if a case should be opened (there are legal ramifications once a case is opened). The system supports what it calls an evidence locker to manage this. This kind of post-decision management is critical as these fraud decisions are those that generally need rapid human intervention but are in a gray area.
The end result of the solution is that insurers find more fraud – they see a reduction in their loss ratio of up to 3%. And because it runs before claims are paid, fraud is prevented rather than being pursued through the traditional (and inefficient) “pay and chase” model.
The majority of installs are now being sold as hosted solutions – Decisions as a Service if you will – and a number of healthcare software companies (McKesson, EDS and IMS Health ) private label the product.
Despite the huge sums involved only the largest insurers and some parts of Medicare have this or a competitive product to detect fraud. Most mid tier and small insurance companies have nothing. No wonder fraud is so high.
Sometimes the cost of stupid systems is clear…