Machine learning-based prediction of post-operative outcomes in robotic-assisted radical prostatectomy: a multi-variable analysis of 758 cases.

Abstract

Robotic-assisted radical prostatectomy (RARP) has become the gold standard treatment for localized prostate cancer. However, predicting post-operative outcomes remains challenging. This study aims to develop and validate predictive models for key outcomes using machine learning approaches and compare them with traditional risk stratification systems. We conducted a retrospective analysis of 758 consecutive patients who underwent RARP between 2014 and 2018. Pre-operative variables included PSA, Gleason score, clinical stage, and IPSS scores. Primary outcomes were biochemical recurrence (BCR), positive surgical margins (PSM) (PSM), and functional outcomes at 12 months. Machine learning algorithms were compared with D'Amico and CAPRA risk stratification systems. The cohort included 758 patients with a mean age of 60.5 years. At 12-month follow-up (n = 634), biochemical recurrence rate was 4.5% (29/634). For pre-operative counseling applications, the machine learning model using only pre-surgical variables achieved AUC 0.783 for predicting 12-month biochemical recurrence, significantly outperforming D'Amico classification (AUC 0.692, p < 0.001). The comprehensive post-operative model incorporating pathological variables achieved optimal performance (AUC 0.847 for 12-month BCR, AUC 0.863 for 24-month BCR). At 12-month follow-up, biochemical recurrence occurred in 4.5% (34/753) of patients. Key pre-operative predictors included PSA (OR 1.23 per ng/mL, 95% CI 1.15-1.31), biopsy Gleason score ≥ 8 (OR 3.45, 95% CI 2.18-5.46), and clinical stage ≥ T2b (OR 2.67, 95% CI 1.89-3.77). Machine learning-based prediction models significantly outperform traditional risk stratification systems for predicting post-operative outcomes in RARP. These models provide personalized risk assessment to guide treatment decisions and patient counseling.