Neural Networks: Looking at the Bigger Picture
Urologist Alan Partin, M.D., Ph.D., loves statistics, facts, and figures: Rearranging them, making sense out of them, and using what he's come up with to help patients. A prime example: The "Partin Tables" he developed, along with Urologist-in-chief Patrick C.Walsh, M.D., which filled a great need by correlating three facts about a man's disease -- PSA level, Gleason score, and clinical stage--and accurately estimating the extent of a man's prostate cancer to help him make an educated decision about treatment. Now, instead of just three pieces of information, he's taking more than a dozen, feeding them into a sophisticated neural network -- a "thinking" computer program he has helped develop -- and asking new questions, such as: What will the results of this man's biopsy be? What will be the pathologic stage of his tumor? Will he have positive lymph nodes?
"Neural networks are not new," says Partin, "but they're fairly new to medicine. The stock market uses them all the time: They watch trends; the network tells them what's going to happen in the next quarter, so they know which stock to buy. Factories use them to measure the temperature of water, steam coming out of the pipes, the noise level in the building about 15 or 20 variables that they continuously monitor -- and they know two days before the machine's going to go down, because they've seen the pattern before. The neural network says, "You're going to be in trouble, you'd better stop the line and fix something."
With a neural network program he and colleagues developed with funding from the National Cancer Institute, Partin says, "I can take a man's PSA, his age, his race, digital rectal examination, free PSA, and I can give him a very good estimate of his probability of having prostate cancer if he were to get a biopsy. Instead of saying "That's a little high, maybe you should get a biopsy." I can say, "You've got a 48 percent chance of having cancer."
Neural networks recognize patterns, "just as you can recognize your child 500 yards away just by glancing." Their answers are educated guesses. The neural networks -- so called because they function like artificial brains, and have the ability to learn from their mistakes -- can see a bigger picture, says Partin: "For the last 15 years, we have been looking at tumor markers, looking at pathologic information, trying to make predictions for prognosis. We look at slides, Gleason scores: we measure PSAs, we have new blood tests. We've been doing image analysis -- looking at the shape and texture and organization of the DNA in the nuclei of prostate cancer cells. Some of these tests are good, some are great, and some are okay. No single one of them can tell us the answer, but maybe all of them together would give us more of an idea what the future holds for men."
But no human, Partin adds, can comprebend so many variables at one time. Enter the neural network, which uses complex mathematical-statistical analysis "to compare variables that aren't inherently coordinated with each other." The network doesn't even try to figure them out. "It simply doesn't care whether the variables make sense together; it's just looking for a pattern."
How, then, does this brainy computer work? Partin gives the example of a kid trying to learn Spanish with flash cards. Hold up a card, the kid looks at the Symbols and takes a guess. "If he's right, we put that card aside, and pick up the next card. If he's wrong, we tell him the correct answer, put the card back in the stack, an ask him again in a few minutes. Keep going through the flash cards, and eventually he'll learn Spanish," Partin says. "Then you can give him words he's never seen before, and because he's learned all the prefixes, suffixes and conjugation he can make a guess, and often he'll get it right." The neural network is simply a matter of training a computer to look at complex series of results and determine a pattern -- the possibility of cure, perhaps, or the likelihood that cancer will be aggressive. "The computer does this thousands of times until, like a brain, it gets pretty good at guessing which horse is going to win the race."
Compared to standard statistical patterns, the neural networks conclusions, based on retrospective data from prostate cancer patients -- 500 so far -- are "far superior," says Partin, who is on a committee with the World Health Organization and the International Union of Cancer Control to investigate neural network technology worldwide. He believes the network has the potential to save millions of men from unnecessary biopsies. "Last year, 25 million men in the U.S. had PSA tests; 20 million of them have had a negative prostate biopsy and don't know what to do next year. We just can't afford to biopsy 20 million men every year. If the neural network can say, "You don't need that biopsy," if all the knowledge that we can grasp is saying that a man is probably okay, then that's where this technology is going to help."
Shaw et.al., Urology, Vol.54:1999.
Kattan et.al., Urology, Vol.47: 14-221, 1996.
Potter et al., Vol.54:1999.
Wei et.al., Urology, Vol.52: 161-172, 1998.