Anatomy-guided prediction of complications after hypospadias repair: development and multicenter validation of stratified machine learning models.

Background: Prediction of adverse outcomes after hypospadias repair remains difficult because the risk is influenced by penile anatomy and by whether the urethral plate is preserved or transected. We aimed to develop and externally validate stratified machine-learning models for postoperative complications requiring surgical intervention after primary hypospadias repair. Methods: This multicenter cohort study used prospectively collected clinical data from 17 tertiary hospitals in China. The development cohort included 584 boys treated at Beijing Children's Hospital, National Center for Children's Health, and the external validation cohort included 511 boys treated at 16 additional centers. Patients were stratified according to urethral plate transection (UPT) status. The outcome was any postoperative complication requiring surgical intervention, including urethrocutaneous fistula, urethral stricture, urethral diverticulum, or recurrent ventral curvature. Candidate predictors included age, penile and glans dimensions, urethral plate width, meatal position, preoperative ventral curvature, degloving/plication variables, and length of deficient urethra. Eight machine-learning algorithms were compared. SHapley Additive exPlanations (SHAP) were used for model interpretation, and Streamlit-based calculators were developed. Results: The final modeling cohort included 1,095 boys. UPT was performed in 333 of 584 patients in the development cohort and in 244 of 511 patients in the external validation cohort. In the development cohort, complications requiring surgical intervention occurred in 78/251 (31.1%) non-UPT patients and 162/333 (48.6%) UPT patients. In the external validation cohort, the corresponding rates were 64/267 (24.0%) and 118/244 (48.4%). For non-UPT patients, the optimized random forest model achieved an area under the curve (AUC) of 0.85; global SHAP importance ranked length of deficient urethra, age, glans length, preoperative curvature, glans width, and urethral plate width. For UPT patients, the optimized decision tree model achieved an AUC of 0.84; global SHAP importance ranked age, post-correction meatal position, length of deficient urethra, penile length, glans length, glans width, and preoperative curvature. Conclusions: UPT-stratified machine-learning models provided interpretable estimates of the risk of postoperative complications requiring surgical intervention after primary hypospadias repair. These tools should be used to support risk counseling and follow-up planning, not to replace intraoperative judgment, surgeon experience, or center-specific quality assessment.