To most Americans, the phrase "Medicare Fraud" brings up images of individuals who cheat the system to collect healthcare when they don't need it. While those cases do exist, that's not what really concerns fraud analysts, insurance companies, and government agencies. The strong majority of fraud cases involve healthcare providers. The problem is that there are so many providers that current healthcare solutions aren't advanced enough to identify fraud in the vast amount of information. One example is the problem of prescription billing abuse. That area is a great example of where machine learning (ML) can have a direct input on a real-world problem.
As of 2015, there were more than 67,000 pharmacies in the USA. One type of prescription fraud is when a pharmacy automatically sends out prescription refills before a patient requests said refills. By charging for prescriptions that the patients are not using, the pharmacies inflate their revenue. Medicare, Medicaid, and private insurance companies can lose large amounts of money to the practice.
One example of the problem was a single pharmacy paying more than $340,000 to settle claims made by the Massachusetts Medicaid program (MassHealth). The challenge is that lack of strong analytics can prevent many state organizations and insurers from analyzing data to more quickly identify potential fraud.
In a similar case from the other side of the country, another auto-refill case was discovered by SCAN Health Plan, a non-profit HMO headquartered in Long Beach, CA. The company received tips from customers that they were receiving lidocaine refills they hadn't ordered, with some receiving lidocaine cream every two weeks. Given the specific charges, SCAN was able to use the Alteryx analytics platform to perform an in-depth analysis of their data, identify a number of pharmacies with questionable transactions, and then remove from SCAN's network the ones who practiced fraudulent billing. Since that 2017 identification of a problem, SCAN has saved more than $2 million dollars through identification of that type of fraud.
Reactive is Good but Proactive is Better
While it is good to know that analytics can find fraud once it's known what the analyst should search, there's an obvious problem. Today, most occurrences of this type of fraud are only identified when a patient complains. Given the time that can pass before that happens and the information is sent to the right people, large losses can happen before the insurer finds out there is a problem. It is far better to be proactive, to identify the problem as soon as the data indicates it is there.
That is where machine learning can come in. Given the large data sets in today's medical industry, ML can be trained to analyze refill patterns for individuals, pharmacies, and regions. When ML is then included in the information infrastructure, exceptions can then be immediately flagged for human investigation. Further study can then determine if the flagged transactions are false positives, good prescriptions that fell outside the expected parameters, or real positives found early.
"Businesses can no longer afford to leave machine learning out of their fraud detection arsenal," said Ashley Kramer, SVP Product Management at Alteryx. "Machine learning can rapidly detect anomalies in data that could indicate fraud, giving analysts the freedom to dig deeper into the data, which in turn, has the potential to significantly reduce financial loss for companies." If only a few unusual events can indicate to an ML system that there might be fraud, humans can be notified by the system early in the fraud attempt, preventing the loses now being identified. Analysis of the large volume of existing data is important in stopping loss and starting the recovery process, but machine learning predictive analysis can shut off the loss much earlier, providing significant savings in the healthcare system.