CMS Using Analytics, Machine Learning to Fight Fraud

The Centers for Medicare and Medicaid Services is using data analytics and machine learning to predict and mitigate fraud, waste and abuse of reimbursements and resources, the agency’s Center for Program Integrity Data Analytics and Systems Director Raymond Wedgeworth said at GDIT Emerge last week.

This organization within CMS focuses on program integrity, and reviewing and preventing issues related to fraud, waste and abuse. Wedgeworth and his team focus on systems and investigatory work for this mission across CMS. To sharpen the work he’s doing, Wedgeworth’s team is “really pouring some gasoline” into advanced analytics and machine-learning capabilities.

“We’re applying … predictive analytics and other analytics,” Wedgeworth said. “Some are things that are more basic, more of rule-base, like everybody’s heard of the Impossible Day, where a provider is billing for more services and times than is really even possible, so you have some of those more basic rule-based models. Then you have anomalous models where you’re comparing peer groups to one another and how they provide care to see if there’s anything really anomalous there.”

With predictive analytics, CMS is applying regressions and logistic regressions to models, alongside structured, supervised data, to identify characteristics and behaviors.

“We’re able to get to where we are stopping fraud early,” Wedgeworth said. “We were able to get some pretty good leads with a pretty good high percentage of the leads being true positives and not false positives, so that work has been really fruitful and great overtime.”

CMS has been pushing forward with these predictive analytics pursuits by applying machine learning to its models. This, Wedgeworth added, is enabling his team to use unstructured, unsupervised data in predictive analytics models.

Machine learning is enhancing fraud-detection analytics by allowing Wedgeworth’s team to bring back information about the outcomes of investigation leads to models to further improve them.

Still, Wedgeworth recognized that there are areas the office can grow in sharpening its predictions. One area is in identifying fraud that isn’t even known yet.

“What we’re hoping though is to expand [machine learning] even greater in terms of looking at and identifying fraud that’s not known,” Wedgeworth said. “We have a history of what we’re doing now and what we’ve done in the past. That enters into a bias because what you did in the past is then being used to reinforce the new machine-learning models, so you can end up in an infinite loop, and we just keep going back to the same thing, so we’re not identifying potentially new and emerging fraud.”

Wedgeworth added that he has seen the promise of using machine learning and natural language processing to look through electronic medical records for where there may be improper payments.

“We’d probably still need a medical record reviewer, a nurse medical coder and others to look at the medical record, but maybe after we’ve gone through some automated scrubbing of those medical records to identify, ‘Hey, these are the hundred that you should be looking at.’ — That’ll help us to be more efficient with our limited resources,” Wedgeworth explained.