Construction and evaluation of a machine learning model based on shear wave elastography ultrasound radiomics for predicting frozen embryo transfer outcomes

Despite the increasing use of frozen embryo transfer (FET), success rates remain suboptimal. Enhancing personalized pre-transfer evaluation and improving the accuracy of pregnancy outcome prediction are crucial for optimizing transfer strategies and ultimately increasing the likelihood of a successful pregnancy. We seek to create and evaluate an interpretable machine learning (ML) model that leverages multimodal radiomics (grayscale (GS) and shear wave elastography (SWE)) combined with clinical factors to predict pregnancy outcomes after FET. This prospective cohort study enrolled 601 patients, randomly divided into training ( n = 420) and testing ( n = 181) sets at a 7:3 ratio. Radiomic features of the endometrium and junctional zone (JZ) from both GS and SWE were extracted and screened. Five ML algorithms were employed to construct and select the optimal radiomics model, followed by calculation of a Rad-score. We employed univariate and multivariate logistic regression on the clinical parameters to establish a clinical model. An integrated model combining radiomic features and clinical parameters was subsequently developed. Model performance was compared using characteristic curves (AUCs) to identify the optimal predictive model.We employed SHapley Additive exPlanations (SHAP) to improve model interpretability and to visually represent how each feature attribute to FET results. The combined model demonstrated superior performance in the training and testing sets, with an AUC of 0.822 (95% CI, 0.780–0.859), compared with the clinical model (AUC: 0.612; 95% CI, 0.562–0.666) and the radiomics model (AUC: 0.781; 95% CI, 0.734–0.826)in the training set, with SHAP analysis confirming the Rad-score as the most impactful predictor. The integrated model, incorporating radiomic features of endometrial elasticity and clinical parameters, accurately predicts FET outcomes and, through SHAP analysis, enables personalized factor evaluation to guide pre-transfer protocol adjustments.