Skip to main content

Advertisement

Log in

Multivariate outlier detection in medicare claims payments applying probabilistic programming methods

  • Published:
Health Services and Outcomes Research Methodology Aims and scope Submit manuscript

Abstract

The rising elderly population continues to demand more cost-effective healthcare programs. In particular, Medicare is a vital program serving the needs of the elderly in the United States. The growing number of people enrolled in healthcare programs such as Medicare, along with the enormous volume of money in the healthcare industry, increases the appeal for, and risk of, fraudulent activities. Out of the many possible factors for the rising cost of healthcare, fraud is a major contributor, but its impacts can be lessened through the use of fraud detection methods. In this paper, we assess possible illegitimate activities by looking at the amounts paid to providers for services rendered to patients. We propose a novel method for fraud detection that focuses on discovering outliers in Medicare payment data using multiple predictors as model inputs. Our multivariate outlier detection approach is twofold: (1) create a Multivariate Adaptive Regression Splines model to produce studentized residuals and, (2) use the residuals as input into a general univariate outlier detection model, based on full Bayesian inference, using probabilistic programming. Using this approach, we are able to incorporate multiple variables to detect outliers with a model that provides probability distributions, with credible intervals, rather than just point values, as with most common outlier detection methods. Additionally, these credible intervals further enhance confidence that the detected outliers should in fact be considered outlying values, thus possibly fraudulent activities. Our results show that the successful detection of these possibly fraudulent activities can provide effective and meaningful results for further investigation within various medical specialties.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6
Fig. 7
Fig. 8
Fig. 9
Fig. 10
Fig. 11

Similar content being viewed by others

Notes

  1. http://www.pbs.org/newshour/rundown/how-growth-of-elderly-population-in-us-compares-with-other-countries/.

  2. http://www.aetna.com/health-reform-connection/aetnas-vision/facts-about-costs.html.

  3. https://www.medicare.gov.

  4. http://www.aoa.acl.gov/Aging_Statistics/Profile/2015/.

  5. https://www.cms.gov/Research-Statistics-Data-and-systems/Statistics-Trends-and-reports/NationalHealthExpendData/index.html.

  6. https://questions.cms.gov/.

  7. https://www.cms.gov/research-statistics-data-and-systems/research-statistics-data-and-systems.html.

  8. https://www.cms.gov/research-statistics-data-and-systems/research-statistics-data-and-systems.html.

  9. https://www.medicaid.gov.

  10. http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm.

  11. https://turi.com/learn/userguide/anomaly_detection/local_outlier_factor.html.

  12. http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm.

  13. https://documents.software.dell.com/statistics/textbook/multivariate-adaptive-regression-splines.

  14. http://www.darpa.mil/program/probabilistic-programming-for-advancing-machine-Learning.

  15. https://www.cms.gov/research-statistics-data-and-systems/research-statistics-data-and-systems.html.

  16. http://kff.org/state-category/medicare/.

  17. https://questions.cms.gov/.

  18. https://www.cms.gov/Medicare/Coding/MedHCPCSGenInfo/.

  19. https://nppes.cms.hhs.gov/NPPES/.

  20. https://questions.cms.gov/.

  21. https://questions.cms.gov/, https://www.healthcare.gov/glossary/.

  22. https://secure.wpsic.com/sales-materials/files/28807-medicare-approved-amounts-tip-sheet.pdf, http://www.nolo.com/legal-encyclopedia/part-b-medical-insurance-what-you-pay.html.

  23. www.minitab.com.

  24. https://www.R-project.org/.

  25. https://www.R-project.org/.

  26. https://turi.com/learn/userguide/anomaly_detection/local_outlier_factor.html.

  27. http://www.itl.nist.gov/div898/handbook/eda/section3/eda35e.htm.

  28. http://oig.hhs.gov/exclusions/index.asp.

  29. http://www.ocala.com/article/20160422/ARTICLES/160429933, https://www.justice.gov/opa/pr/government-intervenes-lawsuit-against-florida-cardiologist-alleging-unnecessary-peripheral/.

  30. http://www.miamiherald.com/news/local/community/miami-dade/article1962581.html.

References

Download references

Acknowledgements

The authors would like to thank the anonymous reviewers and the Associate Editor for the constructive evaluation of this paper and also the various members of the Data Mining and Machine Learning Laboratory, Florida Atlantic University, for the assistance with the reviews. Also, we acknowledge partial support by the NSF (CNS-1427536). Opinions, findings, conclusions, or recommendations in this paper are the authors’ and do not reflect the views of the NSF.

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Richard A. Bauder.

Ethics declarations

Conflict of interest

Richard A. Bauder declares that he has no conflict of interest. Taghi M. Khoshgoftaar declares that he has no conflict of interest.

Ethical approval

This article does not contain any studies with human participants or animals performed by any of the authors.

Appendices

Appendix 1: Additional information on Bayesian inference and probabilistic programming

The use probabilistic programming, as a general inference technique, enables us to easily create a model of some event or characteristic, such as the detection of outliers, and perform probabilistic reasoning for prediction, infer causes from past events, and learn from the past to improve predictions. At its core, probabilistic programming is probabilistic reasoning, where the probabilistic model is expressed using a programming language. This is, in part, because the probabilistic program is interpreted as the distribution from which one can use tools to ask questions about the distribution. Additionally, these modeling languages incorporate random events as primitives and, as mentioned, their runtime environment handles inference. There are other representation-type languages available, such as Bayesian Belief Networks and hidden Markov models. However, these methods are simply simulations, which is not machine learning. A probabilistic program is akin to a simulation that you can run and analyze.

Even though Bayesian inference updates the posterior probabilities, or beliefs, based on any prior assumptions and the evidence, the updated beliefs may not necessarily agree with our prior assumptions and, as in the real world, the evidence tends to bolster, or overtake, any prior assumptions we may have had concerning some event. In that case, we can assume non-informative, or vague, prior beliefs with little to no prior knowledge. We also have the ability to incorporate stronger beliefs and assumptions based on prior knowledge, such as expert opinions, thus improving the model fit. Other benefits of Bayesian methods include more interpretable results, the incorporation of subjective inputs such as medical knowledge, and the quantification of uncertainties. Furthermore, Bayesian techniques provide credible intervals for the different parameters in the model. The credible intervals show that a value or parameter has, for instance, an 80 or 95% probability of being within the interval bands (note these are the default intervals exported by Stan). This is much easier to interpret than the traditional confidence intervals, which indicate that if an experiment is repeated many times, the values will be within this interval 80 or 95% of the time. Additionally, as with other means to assess confidence in results, the credible intervals around the results are reliable given that the model is true.

Appendix 2: Probability model code and details

Below is our general outlier detection Stan probability model, with comments, showing inputs, unknown variables, distributions, and generated outputs. Note the Stan language is mostly vectorized, but there are instances where vectorization is not yet implemented. For instance, in the Stan probability model code, the prior distributions are vectorized functions whereas the likelihood distribution is not hence the need for a for loop. Even so, both methods result in their requisite distributions.

figure a

In the case of the prior distributions given by \(mean\_value \sim normal(100, 100);\) and \(stdev\_value \sim normal(100, 100);\) in the Stan model code, both the mean and standard deviation are prior distributions declared as normal distributions with mean 100 and standard deviation 100. The use of 100 is not set in stone, but more a starting point for the distributions used to generate the likelihood distribution for the student’s t distribution. These static parameters are typically chosen by the researcher as assumptions on the prior distributions, or prior knowledge of the data.

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Bauder, R.A., Khoshgoftaar, T.M. Multivariate outlier detection in medicare claims payments applying probabilistic programming methods. Health Serv Outcomes Res Method 17, 256–289 (2017). https://doi.org/10.1007/s10742-017-0172-1

Download citation

  • Received:

  • Revised:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s10742-017-0172-1

Keywords

Navigation