AI predicting recurrence in non-muscle-invasive bladder cancer: systematic review with study strengths and weaknesses.

Background: Non-muscle-invasive Bladder Cancer (NMIBC) is notorious for its high recurrence rate of 70-80%, imposing a significant human burden and making it one of the costliest cancers to manage. Current prediction tools for NMIBC recurrence rely on scoring systems that often overestimate risk and lack accuracy. Machine learning (ML) and artificial intelligence (AI) are transforming oncological urology by leveraging molecular and clinical data to enhance predictive precision. Methods: This comprehensive review critically examines ML-based frameworks for predicting NMIBC recurrence. A systematic literature search was conducted, focusing on the statistical robustness and algorithmic efficacy of studies. These were categorised by data modalities (e.g., radiomics, clinical, histopathological, genomic) and types of ML models, such as neural networks, deep learning, and random forests. Each study was analysed for strengths, weaknesses, performance metrics, and limitations, with emphasis on generalisability, interpretability, and cost-effectiveness. Results: ML algorithms demonstrate significant potential, with neural networks achieving accuracies of 65-97.5%, particularly with multi-modal datasets, and support vector machines averaging around 75%. Models combining multiple data types consistently outperformed single-modality approaches. However, challenges include limited generalisability due to small datasets and the "black-box" nature of advanced models. Efforts to enhance explainability, such as SHapley Additive ExPlanations (SHAP), show promise but require refinement for clinical use. Conclusion: This review illuminates the nuances, complexities and contexts that influence the real-world advancement and adoption of these AI-driven techniques in precision oncology. It equips researchers with a deeper understanding of the intricacies of the ML algorithms employed. Actionable insights are provided for refining algorithms, optimising multimodal data utilisation, and bridging the gap between predictive accuracy and clinical utility. This rigorous analysis serves as a roadmap to advance real-world AI applications in oncological care, highlighting the collaborative efforts and robust datasets necessary to translate these advancements into tangible benefits for patient management.