Prediction vs. Prevention in health insurance fraud. Why not both?

As introduction, and in the interests of full disclosure, I believe in modeling and analytical tools to identify fraud in health plans. I created and implemented the first use of ‘artificial intelligence’ technology to identify fraud in health insurance over 30 years ago. But practitioners and the public need to know the reasons that health plans use these tools, almost to the exclusion of everything else, and the less-than-optimal results this creates. These models have anointed with prevention capabilities they do not possess, in response to a recognition that prevention of fraud is a necessary initiative. But it goes against the livelihoods of the pay-and-chase infrastructure. -

Medicare and Medicaid and the people who oversee them are enamored with what has been termed ‘predictive modeling.’ For the uninitiated, this involves the use of statistical smoothing techniques to apply probabilities that an event or series of events meet some threshold. These techniques, which originated in the credit card industry, are used to find patterns of activity which may indicate fraud- a outsized percentage of high-reimbursement claim codes by one physician when compared to peers, a physician who refers patients to a DME supplier miles away, and other examples. 

Let’s take a look at some of the pros and cons of these tools, and why they are popular, along with the misconceptions about them. 

First, anti-fraud efforts in public and private healthcare are still ‘pay-and-chase.’ Most health plans have a Special Investigation Unit, or SIU. They are staffed by people with backgrounds in law, criminal investigations and law enforcement. They have no incentive or training in prevention. They look for outliers and anomalies in claims with the help of modeling software, focus on the highest probability targets and ‘build a case.’ So, what’s wrong with that? Nothing really, but it is merely pay and chase in overdrive. This enforcement-first approach is highly dependent on [solely] claims data, while at the same time searching for suspect claims. Claims are not reliable raw data in the same sense as weather, biological interactions, or the location of a shipping container. They represent human attempts to maximize reimbursement and are generated by humans doing such

When Medicare and Medicaid were created, and private plans were tasked with reducing fraud, the natural association was that fraud was a crime, and therefore investigators and prosecutors were best equipped to address it. Of course, these laws and guidelines were created by legislators who were investigators and prosecutors. Each state was required to have a Medicaid Fraud Control Unit, or MFCU, which was an investigative and prosecutorial arm of the Attorney General. The narrow thinking about fraud reduction only counts recoveries and does not address more effective prevention measures. That’s why statistical modeling systems are a law enforcement dream, but in fact a taxpayer nightmare. 

They will help identify more quickly data patterns indicative of fraud, which means that the investigative and prosecutorial complex will now start to ‘build a case.’ This means allowing more fraud to occur [and be paid] while looking at the ‘data’ to see if it passes a threshold of certainty to further step-up criminal investigation efforts. These tools also allow pay & chase practitioners to maintain opacity in selecting targets for investigation and ignoring others.

When news of an arrest hits the government statistical tables and the media, this is when the misguided measures and press releases are let loose. Daily, we read about multimillion dollar fraud cases, many of which should never have been allowed to happen with sensible prevention tools. But prevention of fraud is also prevention of a good headline. Instead of laudatory responses, legislators and their staffs, government finance executives and a responsible media should be asking the questions

1                    How did this fraud get to the multi-million-dollar level before anything was done?

2                    How much of this reported number will the taxpayer actually get back?

Because these entities are measured on recoveries only (see the annual OIG report on MFCU performance), the enforcement-first cadre will take larger cases (where fraud was allowed to continue to ‘build the case’ to prosecute. Aside from good press, there is another reason. The cost of investigation and prosecution, coupled with the percentage of the fraud recovered (overall, less than 20%, I have been told by a former DOJ official) means that these are the only case economically worth pursuing. I have calculated the break-even at about $450,000. As a friend once told me, this sends a message that it’s OK to steal some candy, just don’t loot the store.

In order to justify the expense and return on investment for these projects, they have been endowed by their champions with the ability to predict the future. CMS would not answer a FOIA request for their methodology, but their return-on-investment algorithm counts all the future claims that might be submitted by a provider who gets caught in their pattern recognition net. Preventive tools have no value in the here and now, but according to CMS can predict the claim submission behavior of a DME supplier 3-5 years out. Given that these analyses and probabilities are based on human judgement, technology, (both computer technology and medical technology), legislative initiatives, reimbursement and payment model changes, retirements and many other factors, the accuracy of such predictions is, at best, questionable. As Neils Bohr, one the leading nuclear scientists of the 20th Century said. “Prediction is difficult, particularly about the future.” They are not preventive or predictive in any rigorous analysis

Like most things governmental, money and budgets also play a role in these decision. Preventive tools which can stop inappropriate payments before they are made may alert legislators and spark the greatest fear in any government agency head- a reduction in their budget requests. At least it will engender the question whether the agency needs the money they are asking for.  Statistical software also allows the investigative/enforcement apparatus to operate without oversight in a self-contained ecosphere. This may be a positive development, but also may introduce biases and judgements which are not subject to scrutiny. Investigation and prosecution, aside from keeping the status quo intact, is a multi-year effort. Cases with a verdict often come 3-4 years after an arrest is made. Legislators don’t make the connection that these are monies they allocated years ago that may be superfluous.   

Statistical modeling tools have aided anti-fraud efforts by sifting through mountains of data to identify potential fraud. While they have and can contribute to anti-fraud efforts, they do not address root causes of fraud and by themselves do not make an optimal solution.-