Enhancing frozen-thawed embryo transfer outcomes and treatment personalization through machine learning models.

Background: Infertility affects millions globally, with significant social, emotional, and economic consequences. While frozen-thawed embryo transfer (FET) is a cornerstone of assisted reproductive technology, its clinical pregnancy success rates remain inconsistent (29.6-59.2%). Improving predictive accuracy and personalizing treatment strategies for FET outcomes could address critical unmet needs in reproductive medicine.

Objective: To develop and validate machine learning models to accurately predict clinical pregnancy outcomes following FET and to simulate personalized treatment strategies based on individual patient profiles.

Methods: A retrospective analysis of 1013 FET cycles across two medical centers was conducted. Four machine learning (ML) models-XGBoost, random forest, logistic regression, and deep neural networks-were trained using female-specific features, male-specific features, combined female and male features, and combined features supplemented with expert-selected clinical features. Model performance was evaluated via ROC AUC, sensitivity, and specificity. SHAP analysis identified key predictors, while decision curve analysis assessed clinical utility. Personalized FET strategies were simulated to evaluate the potential for tailored interventions.

Results: The XGBoost model trained on combined features supplemented with expert-selected clinical features outperformed all other models, achieving the highest ROC AUC (0.7922) along with balanced sensitivity (0.7309) and specificity (0.7755). SHAP analysis highlighted embryo quality, female age, and anti-Müllerian hormone levels as top predictors. Decision curve analysis confirmed XGBoost's clinical utility, demonstrating optimal net benefit across decision thresholds by balancing true and false positives. Simulated personalized strategies based on model predictions showed potential to refine treatment protocols, enhancing pregnancy success rates through patient-specific adjustments.

Conclusions: XGBoost-based ML models provide a robust, data-driven framework for predicting FET outcomes and personalizing treatment. By integrating key clinical and embryological factors, these models enable precision care strategies that optimize success rates and patient outcomes. This study underscores the transformative role of ML in advancing reproductive medicine, offering a pathway to improve decision-making and reduce the burden of infertility globally.