Researchers Teach 'Machines' to Detect Medicare Fraud
About $19 billion to $65 billion is lost every year in the United States because of Medicare fraud, waste or abuse.
Using a highly sophisticated form of pattern matching, researchers from 鈥檚 are teaching 鈥渕achines鈥 to detect Medicare fraud. Medicare, the primary health care coverage for Americans 65 and older, accounts for 20 percent of health care spending in the United States. About $19 billion to $65 billion is lost every year because of Medicare fraud, waste or abuse.
Like the proverbial 鈥渘eedle in a haystack,鈥 human auditors or investigators have the painstaking task of manually checking thousands of Medicare claims for specific patterns that could indicate foul play or fraudulent behaviors. Furthermore, according to the U.S. Department of Justice, right now fraud enforcement efforts rely heavily on health care professionals coming forward with information about Medicare fraud.
A study published in the journal is the first to use big data from Medicare Part B and employ advanced data analytics and machine learning to automate the fraud detection process. Programming computers to predict, classify and flag potential fraudulent events and providers could significantly improve fraud detection and lighten the workload for auditors and investigators.
Researchers from 澳门六合彩历史记录鈥檚 examined Medicare Part B dataset from 2012 to 2015. They focused on detecting fraudulent provider claims within the dataset, which consisted of 37 million cases. Fraudulent activities include patient abuse or neglect as well as billing for services not rendered. Physicians and other providers who commit fraud are excluded from participating in federal health care programs like Medicare, and these cases are labeled as 鈥渇raud.鈥
For the study, the researchers aggregated the 37 million cases down to a smaller dataset of 3.7 million and identified a unique process to map fraud labels with known fraudulent providers.
Medicare Part B data included provider information, average payments and charges, procedure codes, the number of procedures performed as well as the medical specialty, which is referred to as provider type. In order to obtain exact matches, the researchers only used the National Provider Identifier (NPI) to match fraud labels to the Medicare Part B data. The NPI is a single identification number issued by the federal government to health care providers. 聽
Researchers directly matched the NPI across the Medicare Part B data, flagging any provider in the 鈥渆xcluded鈥 database as being 鈥渇raudulent.鈥 The research team classified a physician鈥檚 NPI or specialty and specifically looked at whether the predicted specialty differed from the actual specialty, as indicated in the Medicare Part B data.聽
鈥淚f we can predict a physician鈥檚 specialty accurately based on our statistical analyses, then we could potentially find unusual physician behaviors and flag these as possible fraud for further investigation,鈥 said , Ph.D., co-author and Motorola Professor in 澳门六合彩历史记录鈥檚 Department of Computer and Electrical Engineering and Computer Science. 鈥淔or example, if a dermatologist is accurately classified as a cardiologist, then this could indicate that this particular physician is acting in a fraudulent or wasteful way.鈥
For the study, Khoshgoftaar, along with Richard A. Bauder, senior author, a Ph.D. student at 澳门六合彩历史记录 and a data scientist at FPL, and Matthew Herland, a Ph.D. student in 澳门六合彩历史记录鈥檚 Department of Computer and Electrical Engineering and Computer Science, had to address the fact that the original labeled big dataset was highly imbalanced. This imbalance occurred because fraudulent providers are much less common than non-fraudulent providers. This scenario can be likened to 鈥渨here鈥檚 Waldo,鈥 and is problematic for machine learning approaches because the algorithms are trying to distinguish between the classes 鈥 and one dominates the other thereby fooling the learner.
To combat this imbalance, the researchers used random undersampling to reduce the dataset from the 3.7 million cases down to about 12,000 cases. They created seven class distributions and used six different learners across class distributions from severely imbalanced to balanced.
Results from the study show statistically significant differences between all of the learners as well as differences in class distributions for each learner. RF100 (Random Forest), a learning algorithm, was the best at detecting the positives of potential fraud events.
More interestingly, and contrary to popular belief that balanced datasets perform the best, this study found that was not the case for Medicare fraud detection. Keeping more of the non-fraud cases actually helped the learner/model better distinguish between the fraud and non-fraud cases. Specifically, the researchers found the 鈥渟weet spot鈥 for identifying Medicare fraud to be a 90:10 distribution of normal vs. fraudulent data.
鈥淭here are so many intricacies involved in determining what is fraud and what is not fraud such as clerical error,鈥 said Bauder. 鈥淥ur goal is to enable machine learners to cull through all of this data and flag anything suspicious. Then, we can alert investigators and auditors who will only have to focus on 50 cases instead of 500 cases or more.鈥
This detection method also has applications for other types of fraud including insurance and banking and finance. The researchers are currently adding other Medicare-related data sources such as Medicare Part D, using more data sampling methods for class imbalance, and testing other feature selection and engineering approaches.
鈥淕iven the importance of Medicare, which insures more than 54 million Americans over the age of 65, combating fraud is an essential part in providing them with the quality health care they deserve,鈥 said , Ph.D., dean of 澳门六合彩历史记录鈥檚 College of Engineering and Computer Science. 鈥淭he methodology being developed and tested in our college could be a game changer for how we detect Medicare fraud and other fraud in the United States as well as abroad.鈥
-澳门六合彩历史记录-
Tags: faculty and staff | technology | students | research | engineering