A machine-learning approach to predict additional treatment after Bacillus Calmette-Guérin induction in non-muscle-invasive bladder cancer.

Non-muscle invasive bladder cancer (NMIBC) comprises ~ 75% of newly diagnosed bladder cancer, with high-risk NMIBC associated with high rates of recurrence and progression. Nearly 40% of patients experience a lack of efficacy with gold standard Bacillus Calmette-Guérin (BCG) therapy and current methods for predicting BCG response are limited. This multicentre real-world study developed and evaluated machine learning (ML) models using data (April 2008-March 2024) from the Japan Medical Data Vision database to predict whether patients with NMIBC who received BCG induction would require additional treatment (cystectomy). In total, 7962 patients were identified based on NMIBC diagnosis and BCG. After processing, 1524 patients with 56 features were used for ML model development. Each ML model employed distinct feature selection, classification algorithms, and class imbalance strategies. Final ML models used either a nine-feature set plus repeat transurethral resection of bladder tumour (TURBT) obtained using a data-driven approach, or a clinically-informed eight-feature set plus repeat TURBT chosen for clinical relevance. Subsequent model performance suggested that factors other than feature selection, such as data imbalance, were key limitations. These findings demonstrate the feasibility of assembling a real-world dataset and performing exploratory ML modelling, although clinically meaningful prediction remains limited.