Differentiating between obstructive and non-obstructive azoospermia: A machine learning-based approach.

BACKGROUND

Infertility is a major global concern, with azoospermia, being the most severe form of male infertility. Distinguishing between obstructive azoospermia (OA) and non-obstructive azoospermia (NOA) is crucial due to their differing treatment approaches. This study aimed to develop a machine learning model to predict azoospermia subtypes using clinical, ultrasonographic, semen and hormonal analysis data.

METHODS

This retrospective study included all subjects diagnosed with azoospermia. All patients were evaluated by at least one urologist, had their semen sample assessed on at least two different occasions for diagnosis and underwent a testicular biopsy to determine the type of azoospermia, categorized into OA and NOA. Clinical factors, hormonal levels, semen parameters and testicular features were compared between the OA and NOA groups. Three machine learning models, including logistic regression, support vector machine and random forest, were evaluated for their accuracy in differentiating the two subtypes.

RESULTS

The study included a total of 427 patients with azoospermia, of which 326 had NOA and 101 had OA. The median age of the patients was 33.0 (IQR: 7.0) years. Our findings revealed that factors such as body mass index, testicular length, volume and longitudinal axis, semen parameters and hormonal levels differed significantly between the two groups. When these variables were input into the machine learning-based models, logistic regression achieved the highest F1-score and area under the curve value among the three models evaluated.

CONCLUSIONS

This study underscores the potential of machine learning to differentiate between azoospermia subtypes using readily available clinical data. However, further research is required to validate and refine the model before it can be applied clinically.