Development and validation of a LightGBM based machine learning model for predicting postoperative erectile...

Postoperative erectile dysfunction remains a highly debilitating complication following robot assisted radical prostatectomy for localized prostate cancer. Traditional risk stratification heavily relies on anatomical and oncological parameters utilizing simple linear regression algorithms, frequently failing to capture the complex multifaceted physiological and psychological dynamics that govern functional sexual recovery. To engineer and robustly validate an advanced machine learning architecture utilizing a LightGBM framework coupled with SHapley Additive exPlanations to accurately predict postoperative erectile dysfunction by integrating conventional surgical metrics with novel psychosocial and systemic immunological indices. This multicenter retrospective cohort study analyzed 824 patients with strictly localized prostate cancer who underwent robot assisted radical prostatectomy across two independent medical institutions. The primary outcome was functional erectile impairment systematically assessed utilizing validated questionnaires following a minimum six month postoperative observation period. To isolate the optimal prognostic variables, a rigorous two step dimensionality reduction strategy was executed. This approach incorporated Least Absolute Shrinkage and Selection Operator regression alongside a Random Forest feature importance algorithm. The dimensionality reduction pipeline successfully identified seven core independent predictors: the Patient Health Questionnaire 9 score, patient age, the HALP score, baseline serum testosterone, clinical stage, preoperative biopsy Gleason score, and prostate volume. The LightGBM algorithm demonstrated unparalleled predictive superiority in the independent validation cohort, achieving an area under the receiver operating characteristic curve of 0.946, an overall accuracy of 0.927, a sensitivity of 0.775, and a specificity of 0.956. The SHapley Additive exPlanations analysis provided transparent clinical interpretability, unequivocally revealing the Patient Health Questionnaire 9 score as the dominant predictive determinant. Severe baseline depressive symptoms were independently associated with a significantly elevated risk of functional impairment, whereas robust systemic nutritional immunological health and adequate baseline testosterone levels emerged as potential protective factors. This multicenter study developed a LightGBM prognostic framework that incorporates preoperative psychological and systemic physiological metrics to predict erectile dysfunction after robot assisted radical prostatectomy. The integration of these variables provides a more comprehensive approach to risk assessment than traditional physiological models. While further independent external validation remains necessary, this framework may assist urologists in personalized risk stratification and early rehabilitation planning for patients with localized prostate cancer.