Computational models for detection of endocrinopathy in subfertile males

The observation that men with sperm density greater than 10 million/ml had low probability of endocrinopathy led to a refinement in the evaluation of subfertility. Using statistical methods, we sought to provide a more accurate prediction of which patients have an endocrinopathy, and to report the outcome as the odds of having disease. In addition, by examining the parameters that influenced the model significantly, the underlying pathophysiology might be better understood. Records of 1035 men containing variables including testis volume, sperm density, motility as well as the presence of endocrinopathy were randomized into 'training' and 'test' data sets. We modeled the data set using linear and quadratic discriminant function analysis, logistic regression (LR) and a neural network. Wilk's regression analysis was performed to determine which variables influenced the model significantly. Of the four models investigated, LR and a neural network performed the best with receiver operating characteristic areas under the curve of 0.93 and 0.95, respectively, correlating to a sensitivity of 28% and a specificity of 99% for the LR model, and a sensitivity and specificity of 56 and 97% for the neural network model. Reverse regression yielded P-values for the testis volume and sperm density of <0.0001. The neural network and LR models accurately predicted the probability of an endocrinopathy from testis volume, sperm density and motility without serum assays. These models may be accessed via the Internet, allowing urologists to select patients for endocrinologic evaluation at http://www.urocomp.org.