Texas Comptroller SniffS Out Medicaid Fraud With Neural Nets
- By Charlotte Adams
- Apr 30, 1997
Health care fraud in the United States has reached staggering proportions. According to federal government estimates, one in every 10 dollars distributed in health care benefits may wind up in the hands of an illegal recipient. In Texas, that means that as much as $730 million of the $7.3 billion in state and federal funding for Medicaid acute care and long-term nursing home service in fiscal 1996 may have been inappropriately spent.
That money "could help provide preventive care for several hundred thousand additional children," said Robin Herskowitz, a senior policy analyst with the Texas Office of the Comptroller of Public Accounts. The problem, of course, is how to detect abuse of the system. In seeking a solution, Texas is trying an innovative approach: the use of neural network software, modeled on the structure-and learning prowess-of the human brain.
In November 1995 the Texas Comptroller's Office hired Austin, Texas-based Intelligent Technologies Corp. (ITC) to conduct a feasibility study on neural network technology as part of a Medicaid Fraud Detection (MFD) project, one of Texas comptroller John Sharp's recommendations for practicing better government.
The feasibility study was to determine whether neural net technology was appropriate for MFD and whether data available at the state level was suited to the approach, according to Andy Liebler, a senior analyst with the comptroller's office.
Texas turned to ITC-a 2-year-old start-up-because of its founders' experience developing a neural network-based fraud prevention system at Microelectronics and Computer Technology Corp., an Austin technology research consortium. That system, later licensed to Visa, saved the credit card giant more than $100 million in 1995, according to ITC president and chief executive officer Joseph Brown.
"We were pretty sure, theoretically, that the technology would work for Medicaid fraud," Herskowitz said. But credit card fraud is "an easier nut to crack." Medicaid-in the "complexity of the services being delivered, the number of entities and individuals involved in each case, and the condition of the data"-constituted a greater challenge.
The nine-month study, which covered about 20 percent of the state's population, was completed last year and concluded that neural net technology was appropriate for the task of detecting Medicaid fraud. MFD has now entered a one-year transitional phase, which will expand coverage to 30 to 35 percent of the population. "We'll turn over a fully operational system to the Health and Human Services Commission (HHSC) in September of 1997," Liebler explained. It is expected that if and when HHSC takes over the program, it will issue a request for proposals for the full system.
ITC, using neural networks as a core-along with related technology such as genetic algorithms and fuzzy logic-helped to "identify suspicious providers based on Medicaid claims data," Herskowitz said.
Genetic algorithms helped ensure the "survival of the fittest" approaches to the problem. ITC used the algorithms to select about 100 characteristics that the neural nets search for. For example, the software could look for the number of injections given by a physician per patient. Fuzzy logic, on the other hand, helps the fraud detection system to deal with incomplete and incorrect data.
An iterative process was used to identify suspect providers. First, the government asked a group of agency experts to describe the sort of fraud a certain type of provider-say, an ambulance driver-might commit. From the experts' information, ITC developed code telling the computer to pull related information from its data store. The processor then generated a list of "several hundred" suspects, which the experts reviewed and evaluated. That input was fed back into the computer, which then generated a second list. The experts reviewed the second list, paring it down to more than 100 highly suspicious cases. All told, however, the state is investigating only 59 of these because of limited resources.
Problem-solving using neural nets is different from conventional computing approaches, in which a problem is studied, a solution is worked out on paper and code is written to make it work. Rather, neural nets are adaptive and learn by example. This is important when dealing with fraud, Brown contends, because fraud is dynamic and often a step ahead of the law. Even expert systems-a type of artificial intelligence based on rules written by experts-would have difficulty keeping up. Neural nets "simply outperform other techniques" because "they more exactly model the problem," he said.
The hardware needs for the project so far have been relatively simple. The feasibility phase used a Sun Microsystems Inc. SPARCstation Ultra 2 running Unix with a 30G Redundant Array of Independent Disks drive. Dubbed Sentinel, the C++ neural net program worked off about 20G of formatted information. The transitional phase is using an Ultra 2 multiprocessor with 105G of raw storage and an Oracle Corp. relational database. ITC will also spread its net wider to cover about 224,000 providers-up from about 15,000 in the feasibility phase. Sentinel could scale up to full population coverage with an additional 200G of disk space, Brown said.
In the feasibility phase, the software generated a list of a couple hundred suspicious providers out of the total surveyed. ITC could have produced more, Brown said, but was asked to limit the number because of the "limited number of investigators."
The results so far are encouraging. A random sample of cases last August showed the neural nets getting a 38.5 percent hit rate vs. a 14 percent hit rate for the current system. This means that the neural nets outperformed the base system on the same data in the percentage of cases that then "were validated by the team of experts," Herskowitz said.
Because the data in the transitional phase will be stored in a hierarchically organized relational database rather than in "flat files," as was the case in the first phase of the program, investigators will be able to access the data in ways that were not possible before and run complex queries in a shorter time, Brown says.
The program is also developing improved interfaces to the software, Liebler says, so that investigators can more easily pull out not only the suspect's name but other relevant data, such as income shifts and operational statistics. ITC will be adding filters for an additional 12 to 15 types of fraud, up from about six types in the feasibility phase, Brown says. The company is also designing queries to run against the Oracle database.
Funding for the project-out of the comptroller's budget-amounted to $300,000 in the first phase and is moving up to $900,000 in the transitional phase. During the first two years of the statewide system, costs are anticipated to be about $2.3 million per year, with federal funding of 90 percent and 75 percent, respectively, in years one and two. Federal funding will continue at 75 percent in the out years, although the program's cost will have dropped substantially by then.
Where does the technology go from here? "We look at the dynamics of [fraud] practice and how practice changes over time," Brown said. The technology can also make predictions based on past patterns of behavior, although that's not required in the Texas Medicaid project. "Ideally, you'd take a system and split it in half-proactive and reactive," Brown said. That way, when claims come in, "you'd make some sort of evaluation before the money goes out the door." After payment, the money is much harder to retrieve.
Such changes to the system, however, would be the province of the benefits program managers after the program has gone fully operational, Herskowitz says. The delays in payment to providers that might ensue from such modifications would have to be weighed by policy-makers against the likelihood of fraud detection taking place.
The neural net approach has obvious applications in other Medicaid and medical insurance programs. Now tested in the Texas Medicaid context, the technology would be "relatively inexpensive and straightforward" to apply in other states, Brown said. Moreover, neural nets are adaptable to a wide range of problems, including recognition, optimization and prediction. "If I can couch a problem in these terms, I can probably solve it with neural nets," he said.
"I don't see neural networks as avant-garde," he added, "but as technology that has yielded real systems that have solved real problems."
Charlotte Adams is a free-lance writer based in Arlington, Va. She can be reached at firstname.lastname@example.org.