Neural networks for the biochemical prediction of bone mass loss

Neural networks are specialized artificial intelligence techniques that have shown high efficiency in dealing with complex problems. Paradigms such as backpropagation have been successfully applied in a number of biomedical applications, but not in attempts to identify women at risk of postmenopausal osteoporotic complications. In this paper, several neural networks were trained using different combinations of biochemical variables as inputs. Bone densitometric measurements in Ward's triangle and in the spinal column were used as separate classification criteria (outputs) between slow and fast bone mass losers. The most parsimonious model with the best performance included plasma concentrations of estrone, estradiol, osteocalcin, parathyrin and urine concentrations of calcium and hydroxyproline (expressed as ratio to creatinine excretion) as input neurons; ten neurons in a single hidden layer; and one neuron in the output layer. Diagnostic efficiency was 76% in Ward's triangle and 74% in the spinal column; sensitivity was 70 and 81%, and specificity was 77 and 65%, respectively. Linear discriminant analysis showed a diagnostic efficiency of 66% in Ward's triangle and 64% in the spinal column, sensitivity was 55 and 86%, and specificity was 75 and 13%, respectively. We conclude that performance of the stepwise discriminant analysis was not superior to the neural networks.