Causality enhanced machine learning for cumulative live birth rate prediction with threshold band optimization in in vitro fertilization

Abstract

Accurate and individualized prediction of cumulative live birth rates (CLBR) remains an unmet need, as existing models largely overlook individual risk preferences and the translational gap between predictive outputs and actionable clinical decisions. We aimed to develop and validate a machine learning algorithm integrating causal inference and threshold optimization for individualized CLBR prediction and clinical decision support in conventional IVF patients. This retrospective study enrolled 1,058 couples undergoing 1,419 complete IVF cycles (including associated frozen-thawed embryo transfers) from Yantai, China. The cohort was divided into a development cohort (2019–2022) and a temporal validation cohort (2023). Variables were selected from patient characteristics, clinical parameters, and embryo data. Eight machine learning algorithms were employed to construct CLBR prediction models. Multiple counterfactual explanation and causal inference methods, including serial mediation analysis, diverse counterfactual explanations (DiCE), double machine learning (DML), and T-learner, were employed to analyze causal effects of modifiable variables and determine threshold ranges. The final eight-predictor gradient boosting machine model (CLBR-GBM) achieved optimal overall performance and was robustly validated in the temporal validation cohort (AUROC 0.851, F1 score 0.801, Brier score 0.155, ABscore 0.848). Interactive decision curve analysis (iDCA) demonstrated favorable clinical utility, with a cohort-specific optimal threshold of 0.429 yielding a net benefit (NB) of 0.331. The model maintained robust NB (0.312–0.379) across a clinically applicable threshold band of 0.3–0.5. Synergistic adjustment involving decreased BMI (average treatment effect [ATE], -0.010) and increased embryo number (ATE, 0.068) represented the primary pattern shifting predicted probabilities across the 0.429 threshold toward live birth. Among these, BMI intervention demonstrated the most robust causal efficacy in the young high-responder obese subgroup (conditional average treatment effect [CATE], -0.105). The model was deployed as a publicly accessible platform ( https://fertility-yts.shinyapps.io/CLBR_GBM ) for individualized CLBR prediction, serving both patients and clinicians. Causally augmented CLBR-GBM identifies temporally validated optimal decision threshold bands with balanced discrimination and robust transportability for cumulative live birth strategies. Not applicable (retrospective non-interventional study).