Prostate cancer: diagnosis by computer.

In the battle of man versus machine, chalk one up for the machine. A new study shows that a computer, given the same medical history and screening information as a physician, can determine with greater accuracy whether or not a man has prostate cancer. What's more, it can predict in many cases whether the disease will recur.

The computer relies on a neural network to make its predictions. A form of artificial intelligence, neural networks approximate the way the brain processes information and can be taught to recognize complex patterns in data.

"These results suggest that one day we may be able to reduce the number of unnecessary biopsies of the prostate gland," says William J. Catalona, head of urologic surgery at the Washington University School of Medicine in St. Louis and a coauthor of the study. "On average, for every three patients who undergo prostate biopsy based on abnormal results from prostate screening tests, only one patient is found to have cancer," he explains. The computer did far better.

Catalona and his colleagues present their findings in the November JOURNAL OF UROLOGY.

For now, prostate screening tests remain the primary tools for detecting the cancer. These tests include a rectal exam, an ultrasound exam, and the prostate-specific antigen (PSA) blood test, which measures the concentration of PSA, a protein produced by the prostate. But PSA tests often give false positive results, indicating cancer when none exists, or false negative results, concealing the disease.

For their study, Catalona and his colleagues randomly selected 1,787 men who had participated in an earlier, 4-year study of prostate cancer screening. The men in the test group had each had at least one abnormal PSA test. About 40 percent of them had also had a rectal exam that raised suspicions of cancer and had undergone ultrasound exams and biopsies.

To train the neural network, the researchers entered data on 1,578 of the men, including age, race, and results of PSA tests, rectal and ultrasound exams, and biopsies. They then tested the neural network by giving it the same kinds of data from the remaining 209 men -- except for the biopsy findings. Overall, 87 percent of the network's findings matched the biopsy results.

The researchers used a similar method to predict recurrence of cancer. They selected at random four groups of 240 patients, all of whom had undergone surgery to remove a cancerous prostate. Catalona and his colleagues then trained the neural network byentering data for 95 percent of these men into the computer. Using the remaining 5 percent to test the network, the researchers found the computer's predictions to be 90 percent accurate overall.