Development and validation of a machine learning model for sperm DNA fragmentation rate in infertile men: a multicenter retrospective study.

Background: The sperm DNA fragmentation index (DFI) provides important reference for evaluating male fertility and assisted reproductive outcomes, and its degree of damage is influenced by multiple factors. Our study aims to develop a machine learning-based predictive model for identifying high sperm DFI in infertile men using clinical and semen parameters. Methods: We retrospectively collected data on infertile male patients from two centers in Shanghai, China from March 2023 to March 2024. We used data from one center as the training cohort to construct the model and data from another center for external validation. The semen data of the included subjects is combined with clinical features as training features for machine learning. We have developed and validated the effectiveness of six models: Decision Tree (DT), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Logistic Regression (LR), and Naive Bayes Classifier (NB). We comprehensively evaluated the performance of machine learning models with different features using ROC curves, accuracy, and other relevant indicators. The SHapley Additive exPlanations (SHAP) diagram was used to illustrate the importance of variables in the model, Lasso regression is used to screen for core features. Finally, a convenient and practical DFI quality early prediction platform was constructed based on core features. Findings: 1037 patients from one center were included in the development cohort, while 290 patients from another center were included in the external validation queue. The RF model performed the best in predicting the quality of DFI in infertile male patients, with a 10-fold cross-validation AUC of 0.979 (0.972-0.986) in the development cohort, and an AUC of 0.945 (95% CI: 0.916-0.975) in the external validation cohort. Finally, the core factors obtained through screening were included in the model, including progressive motility sperm rate, sperm concentration, sperm viability, daily exercise time, smoking status, alcohol consumption, stress level, and insomnia symptoms were incorporated to generate a publicly accessible online platform (https://4sjajo-0-0.shinyapps.io/dfi-prediction-v4/). Interpretation: The RF model using semen parameters and lifestyle factors shows good discrimination for predicting DFI abnormality in infertile men. However, the model exhibits notable miscalibration in external validation (calibration slope 2.196, intercept -0.196), indicating systematic overestimation of risk and insufficient dispersion of predictions. Therefore, in its current form, the model and its associated online calculator should be considered investigational. Prospective validation and recalibration in independent populations are required before any clinical application. Its impact on patient-important reproductive outcomes (e.g., live birth) remains unknown.