Predictive analytics to reduce fraud: Best practices


Posted October 1st, 2015


Workers reviewing programs


Historically, state program integrity (PI) initiatives have focused on retrospective recovery of paid claims, often referred to as “pay and chase.” Payment of improper claims has the potential to grow as Medicaid becomes larger and more complex. It’s not uncommon for providers to make coding mistakes or misunderstand ordering and billing rules, and it’s a safe bet that criminals will continue to grow in sophistication.  To keep budgets intact, states must take their PI game to the next level by integrating cost avoidance activities into their strategy for addressing fraud, waste and abuse.

Predictive analytics: A data-driven approach to cost avoidance

Predictive analytics are increasingly used to compare Medicaid claims across provider peer groups and validated benchmarks. Claims failing to meet expected patterns in type and frequency of visits, diagnoses, prescriptions and other factors are flagged for investigation.

These same technologies enable PI staff to categorically score providers to determine level of screening, such as a background check or site visit. Visual displays of relationships among providers, service organizations and beneficiaries help states identify and investigate fraudulent or ineligible providers, remove them from the system and prevent them from re-enrolling.

Cost avoidance models

As claims make their way through an MMIS, one or more of the following analytics-based approaches may be applied:

  • Predictive modeling is an umbrella term for a variety of methods that analyze relevant historical data to create a statistical model of future behavior. Analysts make predictions based on the model, “train” it to recognize the probability of a behavior and then apply it to incoming claims. In general, prepayment predictive modeling focuses primarily on specific transactions and office visits; postpayment predictive modeling tends to examine patient and provider behavior.
  • Risk scoring, like a consumer credit score, is used to assess providers based on a predictive model that analyzes billing, claims and other relevant public and private data. A risk scoring model might evaluate providers based on 15 variables and develop scores between 0 and 1,000, with the riskiest providers scoring over 900. The overall risk score is a single metric that provides an at-a-glance evaluation and allows providers to be categorized according to risk level, but an analyst can dig deeper and look at individual scores for each variable.
  • Link analysis, a predictive modeling approach that identifies unusual or hidden relationships, helps expose fraudulent providers. Link analysis can examine the various components of false identities, including provider names, aliases, Social Security numbers, locations, addresses and phone numbers to find links between providers, suppliers, employees and beneficiaries. Using algorithms to evaluate these connections, link analysis generates visualizations that map this network, reveal relationships among providers and known criminals, and simplify the process of data interpretation.
  • Trend analysis is an analysis of behavior or activity over time to identify trends and project future direction. For example, if abnormal provider activity raises a red flag, trend analysis can explore historical behavior. Or it might be used to analyze a less specific query for random or geographic audits, and to predict future trends or behavior. Trend analysis can help states evaluate Medicaid policy, which is especially important as the Medicaid population grows and changes under the ACA.
  • Spike analysis recognizes more obvious behavioral changes by illustrating them as “spikes” that stand out from normal behavior. Spikes may not be out of the ordinary depending on location, time of year or other developments, and they can be used to predict future spikes. On the other hand, a provider’s billing behavior might look flat or “normal” when viewed historically. But if his or her identity and credentials are stolen and used illegally, a spike in the number of claims, patients or services will be obvious to data analysts.
  • Cluster analysis, or clustering, is a type of predictive modeling in which common data objects are automatically grouped into discrete clusters based on predetermined parameters. Clusters of providers can be compared to determine normal billing behavior and indicate trends. Multiple algorithms can be used to create clusters depending on the type of model needed. Analysts may experiment with many different clustering algorithms to find the model that best answers a particular question.

Responsible stewardship of billions of federal and state taxpayer dollars — not to mention the health care of Medicaid recipients — compels states to maintain comprehensive PI initiatives to ensure the effective and appropriate allocation of resources.

By integrating the front end of the claims cycle into the PI process, states can create a more comprehensive strategy for addressing fraud, waste and abuse that’s more efficient and effective than standalone pay and chase.

Medicaid PI will always require the knowledge of data analysts and other experts to evaluate, adjudicate and investigate claims and providers. But having the right tools at their fingertips simplifies their job and saves budget dollars in the process.

To learn more about best practices for using predictive analytics to reduce fraud, waste and abuse, read our handbook, “Identify, Predict, Prevent, Recover: A Handbook for Medicaid Program Integrity.”


Program integrity handbook offers best practices

Use our program integrity handbook to help reduce Medicaid fraud, waste and abuse.

Get handbook

Horizontal Rule

meeting with professionals

Medicaid PI units: Best practices for working with MCOs

Measure and monitor MCO program integrity efforts and help them strengthen their procedures.

Read more

Horizontal Rule