Machine learning-based preliminary screening tool for clinical pregnancy prediction: towards management of IVF/ICSI stages.

backgroundAccurate prediction of pregnancy outcomes in assisted reproductive technology (ART) remains a clinical challenge due to the complexity and heterogeneity of IVF/ICSI cycles. Existing models often focus on isolated treatment stages and rely on linear statistical assumptions, limiting their ability to support personalized care throughout the entire treatment process.

methodsThis retrospective study included 1,062 women who underwent IVF/ICSI between 2016 and 2021, with an additional temporal validation cohort of 250 patients treated in 2022. Two machine learning (ML) models were developed to predict clinical pregnancy outcomes during the pre-treatment and treatment phases. Model performance was evaluated using metrics including precision-recall curves, F1 score, calibration, Brier score, and decision curve analysis. SHapley Additive exPlanations (SHAP) were used to enhance interpretability, and restricted cubic spline (RCS) analysis explored nonlinear relationships. Both models were deployed as interactive web calculators to facilitate clinical use.

resultsBoth models demonstrated favorable performance in internal and external validation. Key predictors identified for the pre-treatment phase included female age, antral follicle count (AFC), and body mass index (BMI). For the treatment phase, important predictors comprised serum progesterone level on HCG day, gonadotropin dosage, and endometrial thickness on HCG day. RCS and subgroup analyses revealed significant nonlinear threshold effects of these variables on pregnancy probability.

conclusionWe developed and validated dual-phase ML models for clinical pregnancy prediction across IVF/ICSI stages. Through improved interpretability and online accessibility, our models offer a practical and individualized decision-support tool to optimize ART strategies in real-world clinical settings.