Health insurer detects insurance fraud with IBM SPSS predictive analytics software

Click here for a Dutch version of the case study.

Zorg en Zekerheid is a medium-sized independent regional health insurer in the Netherlands. With more than 460 employees and 380,000 policyholders, the company is committed to providing accessible and affordable healthcare. The customer’s health is key, which is demonstrated by high-quality services, short lines of communication with healthcare providers, and its non-profit basis. By keeping close contact with care providers, such as family practitioners, physiotherapists and other healthcare specialists, Zorg en Zekerheid is able to keep its costs low and its premiums affordable.

The challenge to detect fraud
The majority of policyholders and healthcare providers submits claims for treatments that have actually taken place. However, a small number commit fraud, for example by adapting invoices. There are instances of policyholders who, after returning from vacation, submit invoices for medical costs made abroad. On further examination it is then shown that the invoiced amount has been altered and is many times higher than the original amount. There are also instances of ‘upcoding’, a form of fraud committed by healthcare providers performing simple services, but claiming for more complex alternatives, which results in higher costs. Through active anti-fraud measures, Zorg en Zekerheid aims to reduce costs and ensure premiums of policyholders remain affordable.

These days, most claims are submitted digitally, straight from the care provider to the insurance company. There are millions of records, and the challenge is to quickly identify the records that are fraudulent. In 1999, Zorg en Zekerheid set up the Special Investigations Unit to detect and combat fraud. This department of fraud experts – consisting of four investigators, an analyst and a unit manager – is primarily engaged in detecting anomalies regarding claims. These deviations are then investigated to determine possible fraudulent practices. Once fraudsters have been tracked down, the money can then be recovered.

In order to uncover fraudulent cases, Zorg en Zekerheid had already been using software to analyze data on the basis of pre-defined risk indicators. For this, it was necessary to manually select the data on the basis of these indicators and subsequently determine if fraud was involved. This tended to be a very time-consuming process that did not always produce the desired result. “For the detection of fraud, it is important for us to be able to look into data without knowing in advance what we are going to find and which records will be involved,” said Andor de Vries, fraud analyst with Zorg en Zekerheid. “This is referred to as ‘unsupervised learning’ and requires a solution capable of analyzing larger quantities of data, discovering patterns automatically, and bringing anomalies to light.” After working with the previous software for three years, Zorg en Zekerheid sought a solution that would produce more accurate results. The main criterion was that the ‘chance of being caught’ had to be greater. In other words, the new solution had to be capable of detecting deviations – and ultimately cases of fraud – more accurately and efficiently.

IBM SPSS Modeler increases fraud detection capabilities
Zorg en Zekerheid began examining various data mining solutions, including SAS. “After the pilot with IBM® SPSS® Modeler, we were so enthusiastic about the results that I thought it was redundant to look elsewhere. We have worked with the software for two years now and I’ve never lost my initial enthusiasm. I recommend Modeler to everybody,” said de Vries.

In order to demonstrate the power of Modeler, a pilot project was set up where by a model was created through which deviations in claims could be detected. “We had just solved a fraud case and I gave the data regarding that case to IBM SPSS consultants to incorporate in the pilot project,” continued De Vries. “After a quick analysis, Modeler selected five potential fraudsters from a total of over a hundred healthcare providers. The care provider that had just turned out to be fraudulent was also on the list that was generated by Modeler. From that moment on, I was sure that I was dealing with the right party.”

Modeler improves accuracy and speeds fraud investigations
From day one, Modeler has made a considerable contribution to Zorg en Zekerheid’s fraud detection approach, and the organization has made great progress in this regard ever since. Not only has the accuracy of detected deviations improved significantly, but the process moves much faster as well. “While previously the full investigation process might have taken weeks, we are now able to track down fraud cases within days”, concluded De Vries. “We typically express the added value of our department in terms of financial results. By using Modeler, these results have doubled each year since 2007. We are obviously very satisfied with this score.”