Introduction: Intralesional collagenase clostridium histolyticum (CCH) improved curvature by =20% for around half of subjects with Peyronie’s Disease (PD) in IMPRESS I/II RCTs. Improved identification of patients more likely to respond to CCH may optimize the use of a costly and time-intensive treatment. This study aims to determine factors predicting greater reduction of curvature following CCH injections for treatment of PD. Methods: Data extraction focused on demographic and clinical variables for patients with PD who received CCH at our institute from 2009 to 2022. Analyses using various machine learning algorithms including Random Forest, Gradient Boosting, and Extreme Gradient Boosting selected the best-fitting model based on the lowest root mean square deviation (RMSE) derived from 5-fold cross-validation model on a 70% training sample. A hold-out test sample (30%) permitted characterization of the predictive accuracy for each outcome in terms of RMSE, mean absolute error (MAE), and R2. Variable importance scores identified variables with greatest predictive influence on the outcome. Partial dependence plots for each important variable characterized the relationship between that predictor variable and the outcome while accounting for the average effect of all other predictors. Results: 175 patients received CCH for PD and had recorded post-treatment curvature. Of these cases, 62.3% (109) experienced a decrease in curvature. In patients with a decrease in curvature, there was an average of 34.9% (17.0 degrees) improvement. Our best-performing model predicting post-treatment curvature used Gradient Boosting (Training Sample: RMSE = 14.59, MAE = 11.75, R2= 0.39; Test Sample: RMSE = 14.83, MAE = 11.63, R2= 0.22). The most important variables were initial curvature, age at presentation, distance of the plaque from the coronal sulcus, volume of plaque, comorbid ED, and calcification in order of descending importance. Conclusions: Using machine learning techniques, we were able to identify important predictor variables that account for 22% of the variation in curvature outcomes following CCH injections. With further optimization of algorithm parameters and transfer learning using larger datasets, we hope to build a predictive model based on several easily attainable patient characteristics that can be used in an outpatient setting to better guide patient decision-making when selecting treatments for PD. SOURCE OF Funding: This work was supported in part by the 2022 Urology Care Foundation Summer Medical Student Fellowship Program and the AUA South Central Section
