Prediction of sperm extraction in non-obstructive azoospermia patients: a machine-learning perspective

Area of Science:

• Reproductive Medicine and Urology
• Artificial Intelligence in Healthcare
• Male Infertility Research

Background:

• Non-obstructive azoospermia (NOA) affects male fertility, with testicular sperm extraction (TESE) offering a chance for conception via Intracytoplasmic Sperm Injection (ICSI).
• Predicting TESE success in NOA patients is challenging, lacking validated predictive models.
• Machine learning (ML) has not been previously applied to predict sperm retrieval in NOA.

Purpose of the Study:

• To develop and evaluate a machine-learning-based model for predicting the presence or absence of sperm in testicular biopsies of NOA patients.
• To compare the predictive performance of ML models against traditional statistical methods.

Main Methods:

• Retrospective cohort study of 119 NOA patients undergoing TESE between 1995 and 2017.
• Development of gradient-boosted trees (GBTs) and comparison with multivariate logistic regression models (MvLRM).
• Model performance evaluated using receiver operating characteristic (ROC) curve analysis and leave-one-out cross-validation.

Main Results:

• The GBT model achieved an Area Under the Curve (AUC) of 0.807 (±0.032), outperforming MvLRM (AUC 0.75 ±0.052).
• GBT demonstrated a sensitivity of 91% and specificity of 51%, compared to 97% sensitivity and 25% specificity for MvLRM.
• Successful TESE (presence of sperm) was achieved in 65.3% of the patients; clinical and biological variables like FSH, LH, testosterone, and testicular size were included in the models.

Conclusions:

• Machine-learning models, specifically GBTs, show promise in accurately predicting sperm presence in NOA patients undergoing TESE.
• These ML models can form the basis for a clinical decision support system to aid clinicians and patients in TESE planning.
• Further validation through larger, prospective studies is recommended to confirm these findings.