DIAGNOSING PROSTATE CANCER is notoriously unreliable. Only about 35% of the men who undergo biopsies based on suspicious blood tests and rectal exams turn out to have the disease. But a neural-network computer program developed by Kaman Sciences Corp. in Colorado Springs, Colo., is correct 73% to 77% of the time when it predicts prostate cancer, according to a paper published in the November Journal of Urology. The program misses a few cases, too--it's correct only 94% of the time when it predicts that a patient doesn't have the disease. But Kaman principal scientist Peter B. Snow says it could still aid doctors in deciding how urgently a man needs a biopsy.
Neural networks, patterned on the human brain, find clues that doctors might overlook--such as correlations between seemingly unrelated variables. Other researchers have tried using neural networks to diagnose prostate cancer but with less success. Snow says he achieved accuracy by carefully preprocessing the data and by getting the program to apply general principles to new cases. Some neural networks chew over a data set for so long that they memorize the cases and can't cope with new ones.
Washington University School of Medicine supplied Kaman's program with data on about 20,000 patients. Together, they're seeking funding to expand the database to several hundred thousand patients nationwide. Snow says Kaman would eventually like to sell the software or perform diagnoses for a fee. Kaman has also found success in using neural networks on testicular cancer and heart disease.