Background <p>Infertility affects millions of people worldwide, imposing substantial psychological, economic, and social burdens on patients. Although frozen-thawed embryo transfer (FET) is increasingly used, the live birth rate remains approximately 40%, highlighting the urgent need for improved predictive tools to enhance clinical outcomes. This study aimed to develop and validate an interpretable machine learning model for predicting live birth outcomes in patients undergoing FET.</p> Methods <p>Patients who underwent FET cycles at the Department of Reproductive Medicine, Affiliated Hospital of Zunyi Medical University, from January 2021 to December 2023 were included as a retrospective cohort, while those treated from January 2024 to December 2024 formed a prospective validation cohort. The retrospective cohort was randomly divided into training and testing sets at a 7:3 ratio for model development and internal validation. Feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO) and multivariate logistic regression. Based on the selected predictors, three predictive models were constructed: Logistic Regression (LR), Random Forest (RF), and XGBoost. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. Calibration curves and Brier scores were used to assess model calibration, while decision curve analysis (DCA) was employed to evaluate clinical utility. The optimal model was selected, and the Shapley additive explanation (SHAP) method was used to interpret the predictions.</p> Results <p>A total of 4,937 subjects were included. The retrospective development cohort comprised 3,470 cases, and the prospective external validation cohort comprised 1,467 cases. The live birth rate was 41.21% in the development cohort and 46.01% in the external validation cohort. LASSO regression and multivariate logistic regression identified ten key features for model development. The AUCs for the LR, RF, and XGBoost models on the training set were 0.826, 0.881, and 0.854, respectively; on the test set, they were 0.818, 0.828, and 0.843, respectively; and on the external validation set, they were 0.798, 0.796, and 0.809, respectively. Considering performance across all datasets, the XGBoost model demonstrated the best predictive performance. ROC curves showed that the predicted probabilities of the XGBoost model were generally consistent with actual outcomes, and the calibration curves were close to the diagonal, indicating good predictive calibration. DCA confirmed the clinical utility of the XGBoost model, showing high net benefit across most clinical decision thresholds. SHAP analysis identified the number of FET cycles, number of blastocysts transferred, number of high-quality embryos transferred, AMH, E2 level on the day of hCG administration, embryo type at transfer, age, endometrial thickness on the day of hCG administration, duration of infertility, and history of intrauterine adhesions as key predictive factors for live birth in patients undergoing FET.</p> Conclusion <p>The XGBoost-based model shows good predictive performance for live birth following FET. SHAP analysis provides interpretability of key features, which can support the development of targeted clinical interventions to improve pregnancy outcomes in patients undergoing FET.</p>