Refining prognostication in non-muscle-invasive bladder cancer: From clinical models to artificial intelligence

Over the past 2 decades, risk stratification in non-muscle-invasive bladder cancer (NMIBC) has evolved considerably, progressing from simple clinico-pathologic scoring systems to sophisticated, molecular, and artificial intelligence (AI)-driven frameworks. This review summarizes evidence from peer-reviewed studies that evaluate prognostic models incorporating clinical, pathological, molecular, radiomic, or AI-derived variables to predict recurrence, progression, or response to BCG. Although traditional models, such as the EORTC and CUETO tables, are widely used, they demonstrate modest discrimination and limited calibration in contemporary cohorts. Molecular systems, including the 12-gene PCR score and the UROMOL21 classifier, offer deeper biological insight and improved prognostic accuracy. However, they require specialized platforms and lack prospective demonstration of clinical utility. More recent AI-based approaches, including machine learning applied to clinico-pathologic data, deep learning from whole-slide images, and MRI radiomics, have been shown to consistently outperform historical risk tables. These approaches offer improved patient stratification. However, challenges remain regarding reproducibility, interpretability, and integration into clinical pathways. Overall, no single prognostic tool currently fulfills all criteria for broad adoption. Future progress will depend on the development of rigorously validated, multimodal models that integrate clinical, molecular, imaging, and digital pathology data to enable more precise surveillance and treatment decisions for patients with NMIBC.