A machine learning-assisted approach to explain key determinants for achieving live birth in autologous ICSI cycles

Can an explainable machine learning-assisted approach uncover data and prediction trends to select the most transparent model for live birth in autologous ICSI cycles?

A retrospective multicenter cohort study (January 2011-December 2023) included 8,066 patients. Only single autologous ICSI cycles using fresh oocytes and donor or patient sperm (fresh or frozen) were included. A total of 47 pre-treatment and in-cycle variables from electronic medical records were evaluated as potential predictors. Five machine learning (ML) algorithms were applied to select the most predictive variables of live birth in a completed cycle. Using an explainability approach based on SHAP values, the decision-making processes of the algorithms were explored.

A data-driven approach allowed to optimize five ML models for maximum performance, achieving a mean AUC of 88.9 ± 0.5%, balanced accuracy of 80.3 ± 0.5%, precision of 89.1 ± 4.1%, sensitivity of 66.1 ± 3.3% and average F1-score of 0.757 ± 0.009. SHAP values and multicollinearity analysis across all five algorithms selected 16 clinically relevant variables as important predictors of live birth without compromising model performance. An explainable ML approach was able to uncover the decision-making process of the models and reveal that decision tree-based models, particularly the gradient boosting classifier, capture the non-linear distribution of SHAP importance values of the 16 predictive variables, while reflecting current clinical knowledge.

The use of an explainable ML-assisted approach in the prediction of live birth in ART cycles offers the opportunity to uncover underlying data trends that can help select the best model for transparent and more informed patient care by ART professionals.