Use Neural Networks to Detect and Reduce Fraud in the State Medicaid Program


Neural networks a type of advanced computer software should be piloted in the state s Medicaid program to reduce provider fraud.


Background
The crime of fraud is growing at an ala rming rate in both commercial entities and in state service agencies. VISA International, a major credit card company, has reported that credit card fraud has doubled each year for the last three years, with $550 million lost in 1992 alone. VISA expects to los e over $1 billion in 1993 due to fraud. Other commercial companies report equally alarming statistics. Sprint, a major telephone carrier, reported a $4 billion loss in 1992 in phone card fraud; and Exxon, a major gasoline company, reported that 1 in 20 cre dit card transactions are fraudulent.

Statistics on fraud within federal, state or local governmental agencies are not as easy to compile. For commercial entities, the occurrence of fraud eventually becomes evident as victimized customers refuse charges a nd bottom lines are recalculated. Fraud occurring within a state service agency (for example, the Department of Human Services) often goes undetected and thus unreported. Using an extremely conservative estimate of 2 percent, the Medicaid provider fraud fo r the State of Texas would be around $40 million per year.

The past 15 years has seen a nationwide proliferation of computer intelligence or smart computers ranging from the systems that NASA uses to help dock space vehicles, to diagnosing medical illne ss, to reading text from a printed page. A specific type of computer intelligence called neural networks consists of computer programs that can actually learn and mimic the way a human brain works.

Neural networks are specialized computer software that produce non-linear models of complex problems in a fundamentally different way. From a large database, neural networks can develop a set of rules to recognize and predict certain conditions. This softw are learns from experience how to do the task, instead of waiting for programmers to develop the correct relationship. This software works best at recognizing, predicting and controlling patterns, such as in fraud detection, payment reviews and other areas where large amounts of data are gathered.

Neural networks are being increasingly used as real-world tools to find patterns through pattern recognition and computer algorithms. For example, patterns can be the result of a series of medical tests so that they recognize the pattern of a certain illne ss; patterns can be a peculiar type of transaction behavior that can indicate fraud. Neural networks are somewhat unique in that they learn from example, i.e., they have to be shown examples of what needs to be learned.

In private industry, neural networks have been applied to both industrial manufacturing operations, such as chemical, refining and plastics plants, to optimize non-linear processes that previously were controlled by linear modeling or individual plant mana gers and fraud detection, such as in large c redit card operations. Experience has shown that operations can be significantly improved by use of neural networks. Controlled operations in the chemical and refining area show a 40 percent improvement in efficiency, while fraud detection at a major credi t card operation improved 20 percent. Experts predict that further refinement of the neural networks will allow up to 60 percent improvement in this amount, limited by the ability to predict fraud on the basis of only one transaction. 1~

There is some basis for the successful use of this type of computer analysis by the state. The Comptroller s Tax Audit Division uses sophisticated computer programs to identify audit leads, significantly increasing revenue detections in tax audits.

The 1992 Texas Performance Review report on the Department of Human Services (DHS) inspector general s office identified weaknesses in the detection of provider fraud. 2~ Areas where the use of advanced technology to detect fraud could improve their results included the Medicaid , Food Stamp and Aid to Families With Dependent Children (AFDC) fraud-detection programs. In the area of Medicaid provider fraud, referrals for fraud investigation often come from caseworkers who report a suspected fraud case, not from a systematic review for patterns in Medicaid expenditures. Although fraud is difficult to project where there are only institutional victims, not individuals, an estimated 2 to 5 percent of Medicaid expenditures involve provider fraud. 3~


Recommendations
A. The Legislature should direct the Texas Medicaid agency to be the initial site for the development of neural network software; the software would be specifically used to detect provider fraud in the Medicaid program.

The Medicaid program is currently at the Department of Human Services but is scheduled to be moved September 1, 1993, to the Department of Health, with policy oversight by the Health and Human Services Commission. The Legislature should require that the M edicaid agency prepare an interim report on the development, implementation and integration of fraud detection neural network technology by March 1, 1994. The intent is to have the system operational by September 1, 1994.

It is currently estimated that the development of initial fraud-detecting neural networks would not exceed $300,000 from redirected appropriations made to the designated Medicaid agency. These funds would be about 36 percent state and 64 percent federal, resulting in $108,000 of General Revenue Funds and $192,000 of federal funds. The start-up costs could be offset many times over by fraud detection.

B. The Legislature should mandate that the Department of Information Resources (DIR) assist appropriate agencies to develop, implement and integrate a fraud detection system using neural network software.

The Legislature should direct DIR to evaluate the feasibility of each proposed approach and develop solutions to any procurement barriers. There would be no additional costs since reviews of procurement requests are within DIR s current res ponsibilities.


Implications
Given the current rise in the national incidence of fraud, the State of Texas cannot afford to ignore fraud within its service agencies. Failing to act will only encourage others to take advantage of the system. Word gets out quickly when new fraudulent sc hemes prove to be effective. In private industry, this do nothing and pass on the losses attitude is partly responsible for the amount of money lost doubling for consecutive years in the credit card industry. Private industr y is responding using neural network techniques. Large state governmental programs that are also vulnerable to vendor, provider and recipient fraud are excellent candidates for the use of neural networks.


Fiscal Impact
Because the level of contractor, vendor, provider and recipient fraud cannot be accurately estimated in Texas state government, it is impossible to estimate the precise dollar savings in implementing neural networks. However, no matter what the assumptions made about the percentage of fraud currently conducted in the service agencies, the effects of detecting fraud would have tangible financial benefits to the state.

There are at least two procurement models being used for developing and implementing neural network fraud detection systems. The first is the standard research and development contract in which a system is specified, including performance levels. A contrac tor then develops the system for an agreed-upon price. Under this approach, a first phase of neural network development for the Medicaid provider-fraud project is estimated to be between $200,000 and $300,000. As with other Microelectronics and Computer Technology Corporation (MCC) projects, the development of this application will have to be negotiated with the Medicaid agency and other agencies that are interested in neural networks. The coordination of DIR in these negotiations would allow the state to get the best price and to coordinate all agencies that may have an interest in developing this software.

The second approac h uses a fee based upon the percentage of fraud caught. The more fraud that is caught the more the vendor is paid for the use of the software. DIR should evaluate the feasibility of each approach and develop solutions to any procurement barriers.

The implementation of neural networks would take approximately one year to define the scope, negotiate the contract and develop and test the neural network. Costs of the project would be covered by redirected appropriations to the designated Medicaid agenc y. Savin gs could begin September 1994 after the neural networks are implemented and an effective fraud investigation is made. MCC estimates that a 20 percent fraud-identification rate should be possible for the first year of operation. Based on a 2 percent possibl e fraud figure for $6.1 billion in projected 1992 Medicaid expenditures, this would result in possible savings through identified fraud of $12 million in fiscal 1995 and $24.5 million per biennium thereafter. 4~ Because Medicaid is a federal program, approx imately 64 percent of the savings and cost avoidance would be federal taxpayer dollars, and the remaining 36 percent would be savings of Texas state funds. Final savings would be contingent on how aggressively and successfully Texas pursues recovery of fra udulent and abusive charges by Medicaid providers.



Endnotes
1 Interview with Dr. Joe Brown, Director, Neural Network Project, Microelectronics and Computer Technology Corporation, Austin, Texas, November 2 and December 1, 1992.
2 Texas Comptroller of Public Accounts, Texas Performance Review, Office of the Inspector General, Performance Review, (Austin, Texas, December, 1992).
3 Brown, Interviews.
4 Ibid.