(747) A Novel Deep Learning Algorithm for Predicting Treatment Response to Phosphodiesterase Type 5 Inhibitors in Men with Erectile Dysfunction

Introduction Erectile dysfunction (ED) is a prevalent condition that severely impacts the quality of life and psychological well-being. While phosphodiesterase type 5 inhibitors (PDE5is) are the cornerstone of first-line therapy, a significant subset of patients experiences suboptimal or no response. The inability to reliably predict individual treatment outcomes with conventional clinical assessment represents a major challenge in personalized ED management. Objective This study sought to develop and rigorously validate a deep learning model leveraging comprehensive, multimodal clinical data to accurately predict the probability of a successful treatment outcome following PDE5i administration in men with ED. Methods A retrospective cohort of 2,850 men diagnosed with ED and prescribed PDE5is was analyzed. Input features comprised detailed demographic characteristics, medical history, laboratory values (including testosterone and lipid profile), baseline International Index of Erectile Function (IIEF) scores, and hemodynamic parameters obtained from penile Doppler ultrasound. The primary outcome was treatment success, defined as an IIEF-Erectile Function (IIEF-EF) domain score of ≥22 after a 3-month period of optimized PDE5i therapy. The dataset was randomly partitioned into training (70%), validation (15%), and test (15%) sets. A sophisticated convolutional neural network (CNN) architecture, enhanced with an attention mechanism, was designed to identify complex predictive patterns. Model performance was evaluated based on the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. To ensure interpretability, Shapley Additive exPlanations (SHAP) were employed for global and local feature importance analysis. Results The proposed deep learning model demonstrated superior performance, achieving an AUC of 0.91 (95% CI: 0.88–0.94) on the independent test set. This represented a significant improvement over a traditional logistic regression model (AUC: 0.75, 95% CI: 0.70–0.80; p<0.001). The model's accuracy, sensitivity, and specificity were 85.2%, 86.5%, and 83.7%, respectively. SHAP analysis identified the baseline IIEF-EF score, patient age, serum testosterone level, and specific penile Doppler flow parameters as the most critical predictors of a positive treatment response. Conclusions This study presents a highly accurate deep learning algorithm capable of predicting PDE5i treatment response in men with ED. By integrating readily available clinical data, this tool has significant potential to guide personalized therapeutic decision-making, thereby reducing ineffective treatment trials, optimizing resource allocation, and ultimately improving patient satisfaction and clinical outcomes. External validation in prospective, multi-center cohorts is recommended to confirm its generalizability and clinical utility before widespread adoption. Disclosure No