Today, many health plans use unsupervised machine learning to detect claims with the highest likelihood of being overpaid based on a review of the supporting documentation in a patient’s chart. Given the high cost and administrative burden associated with medical record retrieval, unsupervised learning is a valuable tool for minimizing unnecessary work by staff. It also reduces the risk of provider abrasion by focusing only on those claims and/or providers that exhibit a suspicious pattern.
Unsupervised models also can detect instances where a provider’s billing behavior varies sharply from their peers. In one example, a provider billed for more than 20 unique genetic tests per member—well over double the average of the provider’s peers. Without machine learning, these types of fraudulent activity would be more difficult to detect. Humans would be left to write queries to spot this activity—assuming they knew the type of behavior to look for—whereas machine learning can uncover this behavior automatically with a high rate of precision.
The path to transformational value Machine learning is a vital tool for ensuring clean claims are paid quickly and suspicious claims—those exhibiting high potential for fraud, waste and abuse—are detected before payment is made. It increases operational efficiency by accelerating the claim decision-making process. It also reduces the administrative burden for staff—and the likelihood of burnout—by leveraging NLP to review thousands of pages of medical records within seconds, a task that could take a human reviewer hours or days. Moreover, NLP never gets tired of performing the same action over and over. This ultimately makes staff review far more effective and efficient.
But machine learning models in and of themselves do not guarantee payment integrity prowess. Taking a transformational approach to machine learning-driven involves four key elements.
1. Vast amounts of data. Machine learning models must be trained using large quantities of data before they can begin to produce highly effective—and highly accurate—results. The more data that is used—including historical data that reveals trends over time—the greater the statistical likelihood of a correct result. Over time, refinement of the model will depend on continuous access to a data pipeline that feeds the model so it can catch new schemes as they emerge and, just as important, better spot instances where human review isn’t needed.