EDITOR’S NOTE: This is the second in a continuing series on the growing use of predictive analytics being deployed by the federal government to identify aberrant claims. Dr. Jost has a Ph.D in Applied Statistics and is an inventor or co-inventor on nine patents dealing with fraud prevention predictive models and scoring systems for both healthcare and financial services.
Healthcare providers today are encountering an increasing number of claim adjudication and reimbursement challenges affecting their business operations and cash flow. Insurance companies continue to push for lower provider reimbursement rates because of their decreasing enrollments and reduced margins.
Government agencies are promulgating more regulations that add costs to an already expensive workflow. Specifically, Recovery Audit Contractors are becoming more aggressive and comprehensive in issuing their costly and disruptive audit requests.
Lastly, payers continue to hold provider claim payments for 40 days or more. These cost shifting, regulatory and audit issues, plus the “payment float” strategy by payers, creates additional and unnecessary burdens for providers.
President Obama last year signed the Patient Protection and Affordable Care Act, increasing transparency and data access across government agencies in an effort to reduce fraud, waste and abuse while extending the period for recovering overpayments due to fraud.
This legislation has generated industry-wide interest in using predictive fraud modeling, similar to what has been used in the financial services industry for the last 20 years. These new predictive models are intended to replace post-payment, rules-based systems that are inefficient and outdated.
Current Healthcare Landscape
The current rules-based approach to fraud risk management in the healthcare industry ensures that most investigative reviews take place long after a provider, or an impostor posing as a provider, has been paid for services billed.
This is because the rules-based systems used to review claims do not rank relative risk. Therefore it is impossible to prioritize risk and accurately distinguish between providers that follow the rules and providers that don’t. This inability to prioritize means all providers are treated with the same degree of scrutiny and all of their claims are reviewed with the same intensity. Also, when there are hundreds or even thousands of rules and many of them are violated simultaneously, it is difficult to determine which combinations of violations are more important than others.
This uncertainty makes claim review more complex and less prone to comparison or prioritization. By failing to eliminate the “good” providers from intense scrutiny, payers create long delays in evaluation and review. This ineffective and time-consuming review environment also makes it easier for unscrupulous providers to exploit the current system and disappear before recovery efforts are deployed.
Predictive model methodology involves identifying the likelihood that a provider will act in bad faith by evaluating current and historical behavior and assigning a score. Generally a high score indicates high risk of fraud, abuse or abnormal behavior. Any provider below a certain score threshold is considered to be at low risk for such behavior. It can be expected that between 90 and 95 percent of all healthcare providers are “good guys” who submit claims that are not fraudulent, wasteful or abusive.
When the providers are ranked by risk it can be expected that the5 percent or so of “bad” providers will have about 10 percent of the flawed claims and perhaps as much as 15-20 percent of the flawed dollars billed.
However, again, not all predictive models are equally effective. The performance of a predictive model depends on the statistical techniques used, the availability of good data and the experience and capabilities of the professionals who build the models. In fact, some of the new predictive models may not be much more effective than existing rules-based systems. Additionally, some models may mix many of the “good” providers in with the “bad” providers, causing their risk ranking effectiveness to deteriorate rapidly. This relative risk ranking is essential in order to prioritize claims for review and investigation. Using risk ranking to identify low-risk providers and claims ultimately will allow payers to streamline their integrity programs, causing less disruption for the vast majority of providers.
Adoption of Analytical Technology in Healthcare
Why hasn’t the healthcare industry embraced the use of analytical technology to manage fraud and abuse risk yet? After all, it has been nearly 15 years since the development of the first healthcare fraud and abuse detection systems. One reason for resistance has to do with the fact that there is not a statistically valid sample of prior fraudulent or abusive transactions. When building statistical predictive models, developers use historical data to create a “picture” of what known fraud and abuse patterns statistically look like. Because there are not many reliable historical examples of fraud and abuse in healthcare, companies that build such models cannot use more sophisticated statistical systems like those used in the financial industry. These more sophisticated models sometimes are referred to as “supervised” models. Because there are a statistically insufficient
number of prior examples of fraud and abuse, supervised models are not currently appropriate for healthcare fraud modeling. Currently, most healthcare analytics are based on an unsupervised modeling technique called a “standardized z-score,” an anomaly detection and outlier identifier. However, both the supervised modeling and standardized z-score techniques have severe limitations for use in building healthcare predictive scoring models (the next article in this series will provide a detailed explanation as to why z-score models and other methods are inappropriate statistical techniques for healthcare predictive fraud models and have a direct negative impact on “good” providers).
The inappropriateness of traditional statistical technology such as supervised models and standardized z-scores combined with the use of rules to detect healthcare fraud and abuse may explain why there are so many law-abiding providers impacted by the abusive practices of a relatively small number of providers acting in bad faith. These statistical methods have a high failure rate in recoupment actions, as reported recently in the Fierce HealthFinance newsletter. The article states that “of those appeals (of recoupment actions) that finished the process, 85 percent are overturned in favor of the provider.”(1) If RACs and other healthcare analytics entities are using inappropriate analytical technology such as rules, traditional standardized z-scores or even supervised models, it is not surprising that a large number of recoupment action appeals are being overturned. In the fraud prevention and detection analytical technology business, these overturned appeals result in what are termed “false positives” because they initially appear to represent fraud or abuse, but upon review are determined to be valid. An overturn rate of 85 percent is extreme and represents an unrealistically high false-positive rate. Using incorrect or inappropriate statistical techniques is one reasonable explanation for such a high “false positive” rate. Aside from the statistical issues related to using inappropriate techniques, high false-positive rates result in disruptive and costly business issues for providers. It takes their focus away from treating patients and directs it toward dealing with administrative audits, which further adds to administrative costs in an already stressed industry.
Legacy, post-payment systems with rules-based payment integrity audits have been notorious for placing significant burdens on providers. Additionally, inappropriate predictive modeling techniques have the capability of imposing the same kinds of draconian obligations, including increased financial and administrative risks, on good and ethical providers. But such providers should be treated better and differently from those suspected of routinely fraudulent, wasteful or abusive behavior. Current legacy systems that cannot rank relative risk effectively are not able to separate good providers from bad providers, while appropriate predictive models can. In order to make a significant impact on fraud and abuse in healthcare, new analytical technology using appropriate predictive models with real-time, pre-payment assessments must be utilized to review all claims in a consistent and unbiased manner. These actions are essential for minimizing the impact on honest and ethical providers and mitigating the administrative and financial impact caused by outdated and manual rules-based claims review procedures.
Again, the future articles in this series will discuss why not all predictive models are created equal and how some “new” predictive models may adversely impact law-abiding providers.
About the Author
Allen Jost is Chief Analytics Officer of Minneapolis based Fortel Analytics (formerly TerraMedica), a healthcare predictive analytics and risk management solutions company. He is a pioneer in fraud prevention for Financial Services, with nearly 30 years of Financial Services and Healthcare experience in advanced analytics and predictive modeling techniques. Dr. Jost has a Ph.D in Applied Statistics and is an inventor or co-inventor on nine patents dealing with fraud prevention predictive models and scoring systems for both healthcare and financial services, including “Falcon”, the most recognized fraud solution in financial services worldwide.
Contact the Author
To comment on this article please go to firstname.lastname@example.org
(1) “Many RAC denials overturned on appeal” March 1, 2011 – 4:13pm ET | By Ron Shinkman, Fierce HealthFinance http://www.fiercehealthfinance.com/story/many-rac-denials-overturned-appeal/2011-03-01#ixzz1XTYOk73A