Abstract
OBJECTIVE: To develop and validate an interpretable delta ultrasound radiomics model for predicting live birth following single vitrified-warmed blastocyst transfer (SVBT). DESIGN: Retrospective cohort study. SETTING: The First Affiliated Hospital of Guangxi Medical University. PATIENTS: 528 infertile women who underwent SVBT between January 2019 and December 2022. INTERVENTION: Delta radiomics analysis of the gestational sac at 6th and 8th weeks' gestation, with extraction of radiomic features and integration with clinical parameters. MAIN OUTCOME MEASURES: Live birth, defined as the delivery of a single infant at ≥28 weeks of gestation. Radiomic feature differences between weeks 6 and 8 (delta features), combined with maternal age, were used to build ten machine learning models. The best-performing model was interpreted using SHapley Additive exPlanations (SHAP). RESULTS: A delta radiomics model based on logistic regression yielded superior predictive performance compared to models using only single-timepoint features, achieving an AUC of 0.883 in the training cohort. Incorporating maternal age resulted in an AUC of 0.747 in the testing cohort. SHAP analysis confirmed that both delta radiomic features (from weeks 6 and 8) and maternal age substantially influenced model predictions. CONCLUSION: The Delta ultrasound radiomics model with SHAP-based interpretability offers a comprehensive and dynamic approach for predicting live birth in SVBT. TRIAL REGISTRATION: Not applicable. CAPSULE: A delta ultrasound radiomics model incorporating 6 and 8 week gestational sac features, combined with clinical factors, improves live birth prediction following SVBT.