Again going back to the example of FWA around opioids, one of the challenges payers and providers face is separating legitimate instances of patients/members seeing multiple providers and filling prescriptions at multiple pharmacies (such as a cancer patient with other comorbid conditions) from patient/member drug-seeking behavior. How many different physicians and/or pharmacies are too many?

It could take a dozen human experts many months to go through the amount of data required to gain enough understanding to develop rules around these questions. Even then, the answer would likely be incomplete. But machine learning can uncover patterns in the data in hours, or a few days at most, helping reduce false positives while showing with a high degree of certainty that a patient/member who goes to five providers and five pharmacies to get five prescriptions filled within a certain amount of time is exhibiting drug-seeking behavior. The payer or provider can then take action to address it with reasonable certainty that the time will be well-spent, ideally resolving the issue to benefit the patient/member’s health, and preventing future FWA.


As it gains more experience within the organization, machine learning can not only point out areas of variability but also begin to score the severity of those variations. Having that information enables investigators to set priorities for both the highest probability of uncovering FWA and the best ROI for containing those costs.

The human factor
Machine learning doesn’t mean that FWA efforts can go on auto-pilot. While machine learning does a great job of teasing out the variables and finding patterns humans wouldn’t ordinarily see, human expertise is still required to determine what to do about them.

Human expertise is also required to improve the basic models. Keep in mind that machine learning bases what’s “normal” on what it sees across the organization. If the proper chronic condition code has never been included in a particular diagnosis, machine learning will view that as normal. It is up to the human experts to recognize this issue and provide a framework that allows the algorithm to “train” itself to understand which medications are normally associated with which conditions and point out any discrepancies.

Another example is determining whether hospital bills have been overpaid. Suppose anesthesia is coded for a procedure in the emergency department that doesn’t ordinarily require it. The difference might go unnoticed by humans, but machine learning will see the anesthesia as unusual and will point it out. It will then be up to a human reviewer to determine whether there was a legitimate exception or the claim should be amended before being submitted. If it should be amended, the algorithm will automatically be improved to “understand” the next time something similar occurs, saving time and the cost of another review down the road.

Health IT Homepage