Microdissection Testicular sperm extraction (micro-TESE) represents a crucial intervention in the treatment of male infertility. Despite its significance, micro-TESE remains an invasive procedure, with success rates varying widely, often reaching up to 50%. To date, there have been numerous studies exploring the use of machine learning to predict success rates. However, there is still a lack of a robust predictive model utilizing both clinical and laboratory parameters to achieve higher prediction accuracy. This study aims to evaluate the success rate of sperm retrieval in patients with nonobstructive azoospermia (NOA) using clinical data through a machine learning learning module.

We analyzed 97 NOA patients who underwent micro-TESE at Taichung Veterans General Hospital from October 2017 to October 2020. Preoperative data were collected, and the variables used for training and analysis in the K-Nearest Neighbors (KNN) and Extreme Gradient Boosting (XGB) models included height, weight, FSH, LH, testosterone, prolactin, sperm pH, sperm volume, chromosomal abnormalities, Y chromosome deletion, and testicular size on both the right and left sides. These models were trained and optimized on the collected dataset, and their performance was evaluated on a prospective testing dataset by measuring sensitivity, specificity, AUC-ROC, and accuracy.

The models incorporated clinical and laboratory parameters, including hormone levels, sperm parameters, chromosomal abnormalities, and testicular size. The performance evaluation showed that KNN achieved an accuracy of 81.8%, sensitivity of 81.8%, specificity of 75.0%, and an AUC-ROC of 0.94. Meanwhile, XGB demonstrated an accuracy of 81.8%, sensitivity of 81.8%, specificity of 100%, and a higher precision of 100%.

Machine learning algorithms based on appropriate methodologies, especially XGB, have demonstrated sufficient clinical information to predict successful sperm retrieval in men with non-obstructive azoospermia (NOA) undergoing TESE, achieving excellent predictive accuracy.