At present, non-invasive methods are not comprehensive enough to enable urologists to predict sperm retrieval results accurately in patients with non-obstructive azoospermia (NOA). Our aim was to improve the prediction accuracy of sperm retrieval by using leptin and artificial neural networks (ANNs).
Data from May 2004 to July 2010 for 280 patients with NOA were reviewed and assigned into the training and testing set for ANNs. All patients underwent standard diagnostic infertility evaluation and testicular sperm extraction (TESE). Twelve factors were recorded as the input variables for ANNs:
semen volume, seminal pH, seminal alpha-glucosidase and fructose,
serum hormones including FSH, LH, total testosterone (TT), prolactin, estradiol,
serum and seminal leptin. Three ANN models were constructed with the following input variables: ANN1-
. The prediction accuracy for FSH, leptin and ANN models was compared by receiver operating characteristic (ROC) curve analysis.
All ANN models were better than FSH. ANN1 had the largest area under the curve (AUC =0.83) and demonstrated significant improvement compared with FSH (AUC =0.63, P< 0.01) and leptin (AUC =0.59, P< 0.01).
ANNs improve the prediction accuracy of sperm retrieval. Although the leptin AUC is low, combined use of leptin and FSH can significantly improve the prediction accuracy for sperm recovery in NOA patients. Leptin may be a good assistant marker for diagnosing NOA. However, studies with larger numbers of patients are required to confirm the improved predictive performance of ANNs.
Recommendation and review posted by G. Smith