IE 11 Not Supported

For optimal browsing, we recommend Chrome, Firefox or Safari browsers.

Neural Nets Nab

Neural networks -- used successfully for targeting smart bombs and spotting credit-card crimes -- assist states in the fight against health-care fraud.

According to the U.S. General Accounting Office, health care fraud -- estimated at $95 billion in 1995 -- accounts for 10 percent of the nation's total health care spending. For example, some experts say as much as 30 percent of all Workers' Compensation claims are fraudulent.


John Sharp, the Texas comptroller, recently issued a report entitled Disturbing the Peace, which deals with ways of improving state services and controlling costs. According to the report, while the vast majority of Medicaid providers are honest and ethical, even a small amount of fraud in the $10 billion program -- 25 percent of the entire state's budget -- could cost millions.


Texas State Comptroller John Sharp and
ITC president Joe Brown discuss
Texas' Medicaid Fraud Detection System.




One of the report's key recommendations is adapting neural-network technology -- successful in spotting credit card fraud -- to help state investigators identify suspect claims. These recommendations were based in part on a fraud detection pilot in the Comptroller's Office.

"Most fraud detection systems depend on someone dropping a dime," said Guy Denney, director of systems engineering for Intelligent Technologies Corporation (ITC), primary contractor for the pilot. "Using neural-net technology, for the first time Texas has a good way to go off and identify providers electronically."

ITC's neural-net software was used in some of the target recognition technology in Desert Storm.

"We have been developing these systems since about 1983," said ITC President Joe Brown. "We built smart bombs and land mine detectors. We're using the same pattern recognition technology, but now, instead of finding a tank on a desert, we're trying to find fraud."

Rounding Up Fraud in Texas

According to the comptroller's report, a baseline study of the neural-net technology showed good results. Possible targets for investigation were generated from an existing fraud identification system and the neural-net-pilot program. Both sets of targets were passed to investigators to determine which ones warranted follow-up. Of the leads produced by the standard system, 14 percent were deemed worthy of follow-up; of those produced by the neural-net pilot, 39 percent are now being investigated.

The report identified three elements that contribute to the success of the neural-net-based system. "First, it can look at many Medicaid claim details simultaneously to identify subtleties in the data that are not evident when looking at one piece of information at a time. Second, it can combine and analyze data from different databases and apply that information to the problem. Lastly, and perhaps most importantly, it generalizes from its previous 'learning experience' so that it can identify new schemes as they appear."

The pilot system developed for Texas had access to four different state databases populating an Oracle data warehouse. The neural net then ran against the information in the data warehouse. Because detail data is available online through the data warehouse, investigators can also use the system -- drilling down to details that raise the red flag on a particular claim or set of claims.

A bill making its way through the Texas Legislature would implement the recommendations contained in the comptroller's report, including mandating the use of neural-net technology to improve fraud detection. Once passed, the success encountered during the pilot would move forward into a production system.

Utah to Snare Scofflaws

Utah is engaged in a similar effort to adapt neural net and modeling technology to fraud detection, beginning with the Workers' Compensation program. A claimant fraud component will flag possible medical claim fraud. The employer component will look for employers who have misclassified workers to get a break on Workers' Compensation rates. Another component will look for provider fraud -- when a doctor bills for services not rendered, or over-bills by performing too many services or artificially increasing the frequency of service.

The ability of a neural net to learn and adapt to changing conditions is one of the reasons Risk Data Corp., the primary vendor working with Utah on the project, is excited about the possibilities the technology opens up.

"For example, one characteristic of fraud is that sometimes fraud hot spots develop in certain regions of the country," said Sean Downs, senior vice president of sales and marketing for Risk Data. "Fraud investigators associate that with 'bar talk' -- someone will perpetrate a fraud and they will talk with their friends in a bar or some other social setting, and then you'll have a rash of claims come in. Neural networks are able to help identify when a particular area is becoming a hot spot."

Neural-net technology can also help identify patterns of behavior that wouldn't necessarily be evident at first.

"For example, a strong indicator of credit card fraud is when there is a very low-cost transaction at a gas pump followed by a high-cost one at an electronics store," said Downs. "The thief will pump a dollar or two of gas where he can pay at the pump to see if the card is good, and then run down to an electronic or jewelry store and run up something high."

Neural-net technology has helped spot indicators such as this for credit card companies for years. In fact, the ITC team developed a neural-network-based fraud detection system for VISA International, which helped decrease VISA's fraud losses by 17 percent in 1994 and saved the company over $100 million
David Aden DAden@webworldtech.com is a writer from Washington, D.C.