Abstract
Background/Objectives: Prostatic artery embolization (PAE) has been increasingly recognized, especially with recent progress in embolization techniques for the management of lower urinary tract symptoms due to benign prostatic hyperplasia. Nevertheless, a proportion of patients undergoing PAE fail to demonstrate clinical improvement. Machine learning models have the potential to provide valuable prognostic insights for patients undergoing PAE. Methods: A retrospective cohort study was performed utilizing a modified prior-data fitted network architecture to predict short-term (7 weeks) favorable outcomes, defined as a reduction greater than 9 points in the International Prostate Symptom Score (IPSS), in patients who underwent PAE with nBCA glue. Patients were stratified into two groups based on the median IPSS reduction value, and a binary classification model was developed to predict the outcome of interest. The model was developed using clinical tabular data, including both pre-procedural and intra-procedural variables. SHapley Additive ExPlanations (SHAP) were used to uncover the relative importance of features. Results: The final cohort included 109 patients. The model achieved an accuracy of 0.676, an MCC of 0.363, a precision of 0.666, a recall of 0.856, an F1-score of 0.731, and a Brier score of 0.203, with an AUPRC of 0.851 and an AUROC of 0.821. SHAP analysis identified pre-PAE IPSS, prior therapy, right embolization volume, preoperative quality of life, and age as the top five most influential features. Conclusions: Our model showed promising discrimination and calibration in predicting early outcomes of PAE with nBCA glue, highlighting the potential of precision medicine to deliver interpretable, individualized risk assessments.