A systematic review of AI-based prediction models for outcomes after robot-assisted radical prostatectomy: Identifying the gap in surgical technique integration

Abstract

Robot-assisted radical prostatectomy (RARP) is widely used for localized and locally advanced prostate cancer. However, conventional clinicopathologic models provide limited accuracy in predicting postoperative oncologic outcomes. Artificial intelligence (AI) and machine learning (ML) models integrating multiparametric magnetic resonance imaging (MRI) radiomics, genomic classifiers, and clinicopathologic variables may improve individualized risk stratification. The main objective of this study was to systematically review AI-based models incorporating MRI radiomics, genomic signatures, clinicopathologic factors, or surgical variables to predict oncologic outcomes after RARP, and to determine whether operative technique variables have been included. PubMed, Scopus, and Web of Science were systematically searched. Although the review was not prospectively registered, methodological safeguards were described and applied. Two reviewers independently screened records using predefined eligibility criteria, with disagreements resolved by consensus or third-reviewer adjudication. Six studies were included in the qualitative synthesis. Risk of bias was assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST). Meta-analysis was not performed because of substantial heterogeneity in outcomes, predictors, model architectures, and performance reporting. A descriptive forest-style visualization of the best-reported model discrimination was generated. Of the 296 records identified, 14 duplicates were removed and 282 titles and abstracts were screened. Thirty-one full-text articles were assessed, resulting in 6 studies for the core synthesis and 7 contextual references. Studies were conducted across Europe, Asia, North America, and South America. Predicted outcomes included extraprostatic extension, positive surgical margins, pathological upgrading, early biochemical recurrence, and lymph-node involvement. Best-reported area under the curve (AUC) values ranged from 0.711 for pathological upgrading to 0.970 for positive surgical margins. Three studies used external or independent validation cohorts, while one incorporated a validated genomic classifier. None included intraoperative surgical technique variables as model predictors. Calibration was reported in two studies and decision-curve analysis in three. PROBAST assessment indicated low overall risk of bias in two studies and moderate risk in four. AI-based models show promise for predicting oncologic outcomes after RARP, but current evidence remains heterogeneous and incompletely validated. The absence of operative technique variables represents an important and underrecognized evidence gap. Multimodal data integration, standardized reporting, external validation, and prospective evaluation are required before routine clinical implementation.