AI presents an entirely new landscape for defense against audits.

The Trump administration has launched a program to employ artificial intelligence (AI) in Medicare audits. Use of AI will make obsolete current legal strategies leveraged when defending providers.

On Oct. 3, at a giant rally in Florida, the U.S. President signed the Executive Order on Protecting and Improving Medicare for Our Nation’s Seniors. Section 9 states that:

“The Secretary (of the U.S. Department of Health and Human Services) shall undertake all appropriate efforts to direct public and private resources toward detecting and preventing fraud, waste, and abuse, including through the use of the latest technologies such as artificial intelligence.”

AI presents an entirely new landscape for defense against audits. Attorneys and their consultants will need to understand technologies such as neural networks, deep learning, and use of artificial neurons.

Even before the application of AI, fraud detection methods proliferated. Anomalous provider behavior was detected, and “outlier detection . . . based on Bayesian inference” became commonplace. Naïve Bayes classifiers tagged providers working outside their specialty norm.

AI is developing quickly. As outlined in a recent paper, these technologies were deployed against a Medicare Part B fraud data set with over 3.3 million records, each with 29 attributes. AI methods were able to successfully “train” the system to detect fraud. Using former fraud examples is necessary to train AI “supervised” methods, such as decision trees, neural networks, genetic algorithms, and Support Vector Machine.

“Unsupervised” approaches develop their own rules for comparing claims using techniques such as clustering, outlier detection, and association rules. A recent paper introduced a fraud miner using Cascaded Propensity Matching. This approach is significantly better than clustering methods or outlier detection and can be used for both transactional and identity fraud applications. In another study, six deep learning methods were compared so as to identify the best approach.

Does Your Attorney Understand AI and the Science of Big Data?

As audits come to rely more heavily on these advanced data mining and AI techniques, providers will learn that attorneys with adequate scientific knowledge to mount a credible defense are a scarce commodity. We also can expect changes in the administrative law governing the methods of discovering fraud, waste, and abuse. The well-trodden chapters of the Program Integrity Manual on Statistical Sampling will fade into the past and become merely a relic.

It was one thing to defend providers against the sloppy and horrible statistical work that has become so commonplace in Medicare audits. The question for the future is: How do you defend against a machine?

Not So Fast

Using AI and other computerized techniques, in general, can be problematic. A recent meta-analysis of 200 published studies concluded that “fraud detection methods are difficult to implement . . . because new fraud patterns are constantly developed to circumvent fraud detection methods.” Some question the trend towards even more reliance on automation.

The only bright side for providers might be that AI will put many of the Recovery Audit Contractors (RACs) out of business. They will no longer be needed if machines can do their work better, faster, and for a fraction of the cost. Perhaps that is today’s good news.

Note: The executive order is not only about the use of AI. It contains a number of new directions for Medicare policy. It encourages Medicare Medical Savings Accounts [§3(i)] and will flow cash or monetary rebates into supplemental benefits [§3(ii)]. It encourages innovation such as telehealth services [§4(b)]. Section 5 has a number of provisions to enable providers to spend more time with patients.

Programming Note:

Listen to Edward Roche’s report this story live during Monitor Monday, Oct. 14, 10-10:30 a.m. EST.

Share This Article

Share on facebook
Share on twitter
Share on linkedin
Share on email
Share on print