Prediction of conservative treatment failure in ectopic pregnancy using paired temporal ultrasound radiomics

Ectopic pregnancy (EP) is a potentially life-threatening condition, and conservative treatment is commonly used in selected clinically stable patients. However, early identification of patients at risk of conservative treatment failure remains challenging. This study aimed to develop, internally test, and temporally validate an ultrasound-based paired temporal radiomics model using paired pre- and post-treatment ultrasound examinations to predict conservative treatment failure in tubal EP, and to evaluate the performance of a combined nomogram integrating beta-human chorionic gonadotropin (β-hCG)-related variables and radiomics information. This retrospective study included 201 patients with tubal EP who underwent initial non-surgical conservative management, including 149 treatment successes and 52 treatment failures. Patients were randomly divided into a training cohort and a test cohort at an 8:2 ratio, stratified by treatment outcome. An independent temporal validation cohort of 40 patients, including 28 treatment successes and 12 treatment failures, was additionally included. Radiomic features were extracted separately from pre-treatment and post-treatment ultrasound images. Paired temporal radiomics features were constructed by concatenating the pre-treatment and post-treatment feature vectors. Random Forest classifiers were used to construct the β-hCG-based, pre-treatment (Pre), post-treatment (Post), and paired temporal radiomics models. A combined nomogram model was constructed using logistic regression by integrating the selected β-hCG-related variable(s) and radiomics signatures. Model performance was evaluated using receiver operating characteristic (ROC) analysis, calibration assessment, decision curve analysis, bootstrap optimism correction, and independent temporal validation. For each ultrasound examination, 1561 radiomic features were extracted. After feature selection, 24 final paired temporal radiomic features were retained for the paired temporal radiomics model. In the test cohort, the paired temporal radiomics model achieved an area under the receiver operating characteristic curve (AUC) of 0.857, which was higher than those of the Pre and Post radiomics models, with AUCs of 0.565 and 0.577, respectively. The combined nomogram model achieved a test-cohort AUC of 0.813 and did not outperform the paired temporal radiomics model. Bootstrap optimism correction showed that the apparent training AUC of the paired temporal radiomics model decreased from 0.921 to an optimism-corrected AUC of 0.799. In the independent temporal validation cohort, the paired temporal radiomics model achieved an AUC of 0.811; however, at the prespecified training-derived cutoff, its sensitivity was 33.3% and specificity was 92.9%. The combined nomogram model achieved an AUC of 0.786. Ultrasound-based paired temporal radiomics showed moderate discriminatory performance for predicting conservative treatment failure in tubal EP, but the low sensitivity observed at the prespecified cutoff in temporal validation limits its use as a stand-alone clinical tool. Further prospective multicenter validation is needed before clinical implementation.