Steven Walczak is not an M.D. He earned a doctorate in artificial intelligence from the University of Florida in 1990 and is now an associate professor at the University of Colorado at Denver’s Business School. But he has spent a good deal of his career searching for how the disciplines of computing, economics and medicine can best work together to solve big health care problems.
Because of advances in computing power, statistical methods and data gathering, it’s now possible to mine huge datasets to discern patterns that can be used to make reasonable predictions about health care costs, outcomes and even treatment effectiveness, Walczak said.
He is a proponent of artificial neural networks (ANNs), which crunch data in a way that mirrors how the human mind operates. Human brains are adept at cutting through noise or incomplete data to see how variables interact to form a pattern. Combine that ability with the computational power of big computers, and it’s a powerful way to spot hitherto unseen patterns in patient populations.
“There’s an underlying algorithm for everything,” Walczak said, adding that what seems like random distribution isn’t.
Some experts believe the technology is a potentially powerful tool for gaining insights into population health, improving the management of health care costs and even enhancing diagnostic techniques.
“The idea behind neural networks is that we try to figure out what are the important variables,” Walczak said. “It’s the networks’ job to figure out what is the algorithm, the pattern here.”
He believes ANNs are a better way to crunch medical data than conventional methods, which mostly rely on tools that assume that populations produce data points with perfect bell-curve distributions.
Predicting blood needs
Take blood-unit consumption. During surgery, the availability of the appropriate type of blood can mean life or death for a patient.
Medical administrators often use a simple statistical analysis of previous surgical procedures to estimate future use. That method, called a maximum surgical blood ordering schedule (MSBOS), has been successful.
But it has limitations. Because it relies on the law of averages, an MSBOS calls for the same amount of blood per procedure for every patient. In other words, it’s a linear tool, one that calculates an average when what’s really needed is a method that can spot how seemingly unrelated data points form a pattern, Walczak said.
In a December 2000 article in the journal Decision Support Systems, Walczak and co-author John Scharf compared blood supply estimates made using an MSBOS to predictions made with an ANN. Using four years’ worth of data collected from surgical operations at two Veterans Affairs Department hospitals, the researchers concluded that heeding ANN results would have saved $3,339 per operation through reduced blood-unit demands.
Having less blood on hand also reduces the problem of excessive transfusions, which are not only costly but dangerous for patients. Studies show that surgeons tend to use more blood when more of it is available.
Is Big ANN watching you?
Hype about ANNs is nothing new. Since psychologist Frank Rosenblatt developed the first network in 1958, ANNs have been touted as a tool for everything from beating the stock market to diagnosing heart attacks.
ANNs “were deployed before they were truly mature in all kinds of industries,” said Kevin Desouza, an assistant professor at the University of Washington’s Information School who has researched ANNs’ health care applications. But as stories about ANNs’ shortcomings accumulated, the medical world “started to be closed about neural networks.”
Moreover, some health care practitioners were cautious about ANNs because they feared that predicting medical resource consumption could lead to health care rationing.
According to their reasoning, if payers can predict the resource consumption of individuals and estimate that for the cost of treating one patient they could cure five patients more cheaply, they’d be tempted to decide that the life of the more costly patient was not worth saving.
“Clearly, there have to be appropriate protections in place,” said Kelly Cronin, director of the Office of Programs and Coordination at the Health and Human Services Department’s Office of the National Coordinator for Health Information Technology. “This country is quite nervous to ration health care. I think our expenditures reflect that basic value.”
But ANN advocates say health care becomes more accessible to everyone when overall costs decrease, and better data analysis is the way forward.
“You try to optimize as much as you can so you can use your limited resource in the most efficient way,” Desouza said.
By themselves, ANNs won’t solve the resource crisis, he added. But they can help by cutting waste — for example, by determining whether a patient really needs that next unit of blood.
The way ahead?
Walczak said ANNs represent a powerful tool for managing population health. But he said physicians must embrace it before it will be used widely for setting health care policies.
“The issue with medicine is, if you’re doing diagnostics, you have to get the physicians to accept it,” he said. “The way to get [technology] in place is to claim you’re not doing anything diagnostic.”
Moreover, doctors are well aware that they’re ultimately responsible for their decisions. “They cannot say, ‘Well, the computer told me to do this, so I did it,’” Walczak said. “It’s unreasonable to assume that they’re going to just rely on the technology.”
Although ANNs have been around since the 1950s, it’s only recently that computers have become powerful enough to support the level of pattern-recognition researchers consider statistically meaningful.
ANNs also represent a different way of thinking, a break with old laws of certainty and a hint that we’re secretly propelled by unknown patterns.
But they are the future, Walczak said.
“Information technology is there to enable” health care professionals, Desouza said. “It’s not there to replace them.”