Deep learning approaches for predicting response to neoadjuvant chemotherapy in muscle invasive bladder cancer: A systematic review.

Neoadjuvant chemotherapy (NAC) followed by radical cystectomy is the standard of care for muscle-invasive bladder cancer (MIBC), yet treatment response is highly variable and current clinical predictors are inadequate. Deep learning (DL) offers a data-driven approach for modeling complex medical data and may improve response prediction. The objective of this systematic review was to systematically evaluate the performance, methodological rigor, and clinical applicability of DL models for predicting treatment response to NAC in patients with MIBC. Following PRISMA 2020 guidelines, we systematically searched 14 databases and clinical trial registries for studies applying DL models to predict NAC response in MIBC. Eligible studies included those using imaging or omics data and reporting quantitative performance metrics. Methodological quality was assessed using Prediction model Of Bias Assessment Tool (PROBAST) and APPRAISE-AI framework. Twelve studies comprising at least 1,575 unique patients were included. Most employed convolutional neural networks (CNNs), with some integrating radiomics, transcriptomics, or large language models. Reported area under the receiver operating characteristic curves ranged from 0.69 to 0.89, with hybrid models consistently outperforming standard CNNs. While clinical relevance and data quality were generally strong, as assessed by APPRAISE-AI, reproducibility and external validation were frequently limited. Using PROBAST, risk of bias was low in several domains, but concerns persisted regarding outcome definition and analytical rigor. DL models, particularly those integrating imaging with radiomics or clinical features, demonstrate promising potential for predicting response to NAC in MIBC. However, methodological inconsistencies and limited external validation currently constrain their clinical translation. Future research should prioritize prospective, multi-center validation, and standardized multi-modal integration to enable safe and effective clinical deployment.