The Netherlands has a healthcare system providing access to care for all with solidarity through compulsory medical insurance to all inhabitants. The services are provided through health insurers that make specific arrangements with healthcare providers. In 2018, annual healthcare costs in the Netherlands totalled €100 billion, according to CBS (2019) sup>1. Almost half of this amount is accounted for by health insurers. Health insurance companies have therefore extended governance in place to manage the financial transactions related to these healthcare costs. To manage and reduce the risk of unlawful payments, insurers perform checks/controls on invoices. This allows insurers to identify mistakes or fraudulent payments, helping to reduce unnecessary costs, as ultimately the cost of fraud would be paid by the population through higher premiums.

In this article, we describe a research study we have developed as part of research for VGZ, one of the largest health insurance companies in the Netherlands. In this study, we analyse potential fraudulent claims in a portfolio of healthcare provider invoices for neighbourhood care. Healthcare fraud is considered a material risk in the Netherlands. NOS (2020)² report that the excessive number of healthcare fraud cases in the Netherlands puts undue pressure on the Dutch prosecution system. The Netherlands Public Prosecution Service and Dutch healthcare insurers approximate an annual loss of €100 million due to healthcare fraud (NOS, 2020) [2].

Given the materiality of health insurance fraud, there is a growing effort by insurers to tackle this issue. With the use of the expertise and supervision of Milliman, we were able to research and produce a Master's thesis on the application of statistical methods to detect fraud. The analyses were conducted on real-world invoice data from VGZ. Based on interviews with fraud experts, key invoice data features were selected that could potentially have an impact on detecting insurance fraud within VGZ. This article briefly describes the approach of the research and the statistical methods applied to detect fraud more effectively at VGZ.

We started our research with interviews of several domain experts within VGZ. These domain experts detect suspicious behaviour related to fraud by both manual inspection of invoices and with the use of a model. We concluded that the application of two specific statistical methods could potentially complement the existing methods for detecting irregular invoices.

The first statistical method is based on Benford's law. Benford’s law is a naturally occurring phenomenon, which makes a statement about the frequency distribution of leading digits in real datasets (e.g., accounting, macro-economic and election data). The method based on Benford's law aims to separate naturally occurring numbers from humanly fabricated numbers. We found that this can therefore also be used to target potential fraud in invoice cost amounts, where these are fabricated in order to invoice treatments that never took place. Applying Benford's law on the actual sets of invoices of a large number of healthcare providers, we were able to identify healthcare providers which, after further research, were considered to have potentially made fraudulent claims.

The second statistical algorithm applied in this work is the Isolation Forest (iForest). The iForest is a machine learning algorithm for anomaly detection based on the principle of isolating anomalies based on Euclidean distances. Compared to the majority of statistical models, the iForest algorithm has a fundamentally different approach. Statistical models for fraud detection often model legitimate instances in order to identify instances different from those considered legitimate. The iForest method directly detects anomalous instances without modelling legitimate instances. The advantage of the iForest method is that it is likely to complement existing statistical methods that model legitimate instances. Furthermore, the iForest method has proven to be a great tool to detect irregular healthcare providers by using a small number of features as input, using only features in which deviations are likely correlated with fraud.

This research is based on the fact that fraudulent behaviour oftentimes manifests itself by different patterns than what would actually be observed in valid behaviour. With the use of different methods that each tackle a particular type of irregular behaviour, more irregularities can be detected earlier in automatized ways. The application of Benford’s law and the iForest provide more insight into the types of potential fraud to detect and categorise each type of potential fraud more effectively. Unlawful payments detected by Benford’s law are likely due to fabricated invoices. The iForest implementation works on detecting deviations in a mix of 17 attribute values. Irregular healthcare providers (i.e., healthcare providers with irregular invoice behaviour) are expected to have a different mix of attribute values from the majority of the data. The attribute values are statistical values of attributes that are often related to fraud (e.g., deviations in age of patients, billing dates and billing amounts). Therefore, healthcare providers flagged as irregular by the iForest are likely a result from deviations in the attribute values.

In this research, the detection of irregular healthcare providers using Benford’s law and iForest models are compared with a model that detects irregular healthcare providers at random. The p-value of the null hypothesis that a model detects irregular healthcare providers at random are approximated. It is found that Benford’s law and the iForest can detect better than randomly selecting irregular healthcare providers, with p-values of 0.123% and 6.84∙10^(-5)%, respectively.

t of the approximately 5000 healthcare providers, each model selected and ranked 50 healthcare providers that were most likely to be fraudulent according to the two methodologies. Fraud inspectors then further investigated these flagged healthcare providers. Based on the occurrence of irregularities within the total population (estimated ~2%), a random model would be expected to have one true positive flag and 49 false positive ones. A true positive is a healthcare provider that is indeed flagged to have irregularities after further investigations by fraud inspectors. Our models were able to detect 11 true positives: two by Benford’s law and nine by the iForest. Interestingly, the two models detected different irregular healthcare providers and hence complement each other. Moreover, it is worth mentioning that the two true positives in the Benford’s law model were ranked in its top three, whereas none of the nine true positive ones in the iForest model occurred in its top three. Therefore, we conclude that Benford’s law is especially powerful where the sample size is limited and the number of false positive needs to be very low, while the iForest model is more powerful if the sample size is larger and false positive ones are less of a problem.

Analyses results Benford’s law and iForest model

• Population size about 5000 healthcare providers
• Sample size 50
• Based on the assumed fraud occurrence in the population a random model expected to have one true positive flag based on an expected 1% irregular healthcare providers
• A true positive is a healthcare provider that is indeed flagged to have irregularities after further investigation by fraud inspectors
• Benford’s law three true positive flags out of 50, of which two occurring in its top three, whereas iForest model nine true positives ones out of 50, but none occurring in its top three
• No overlap in true positives between the models

VGZ is currently implementing the two aforementioned models in a demo, such that fraud investigators can visually observe potentially irregular healthcare providers according to these models. An implementation of these models could result in a more automated and effective way to tackle irregularities oftentimes related to insurance fraud, reducing manual effort. Ultimately, an efficient and effective system that reduces insurance fraud will ultimately lead to lower premiums and the securement of quality healthcare.

¹CBS. (2019, 2 june). Zorguitgaven; kerncijfers. CBS (StatLine). Retrieved on June 9, 2020, from https://opendata.cbs.nl/statline/#/CBS/nl/dataset/84047NED/table?dl=1D372.

²NOS. (2020, 2 january). Controle zorgfraude faalt, OM en verzekeraars luiden noodklok. Retrieved on June 9, 2020, from https://nos.nl/nieuwsuur/artikel/2317099-controle-zorgfraude-faalt-om-en-verzekeraars-luiden-noodklok.html.