Artificial intelligence for predicting BCG response in non-muscle-invasive bladder cancer: a systematic review.

Abstract

Objectives: To systematically identify, appraise and synthesise artificial intelligence (AI) and machine-learning (ML) models that predict treatment response and clinical outcomes after intravesical bacillus Calmette-Guérin (BCG) in non-muscle-invasive bladder cancer (NMIBC), a setting in which current risk calculators underperform, and identifying non-responders has become urgent as alternatives to BCG enter practice.

Methods: PubMed, EMBASE and Web of Science were searched through April 2026 following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines (International Prospective Register of Systematic Reviews [PROSPERO] number CRD420261376808). Studies developing or validating AI/ML models for BCG-associated outcomes with a quantitative performance metric were included. Risk of bias was assessed with the Prediction model Risk Of Bias ASsessment Tool (PROBAST). Evidence was synthesised along a clinically oriented framework: AI as a perceptual tool (extracting signal from histology or imaging) or integrative tool (re-weighting clinicopathological, molecular, or urinary variables).

Results: A total of 15 studies (>24 900 patients) were included: seven perceptual, eight integrative. By input data, six used digital pathology, two radiomics, two genomics/transcriptomics, four clinicopathological markers, and one urinary biomarkers. The digital-pathology Computational Histology Artificial Intelligence (CHAI) platform, validated across 12 international centres, stratified high-grade recurrence (hazard ratio [HR] 2.08), progression (HR 3.87) and BCG-unresponsive disease (HR 2.31), and was the only model providing a first signal of predictive value, demonstrating a significant BCG vs gemcitabine/docetaxel interaction (P = 0.029). Integrative models PROGRxN-BCa (concordance index [C-index] 0.79) and DeepSurv (C-index 0.881) outperformed standard calculators but with modest gains (ΔC-index 0.05-0.10). Only 40% of studies performed external validation, none prospectively; five were at high risk of bias.

Conclusion: Perceptual AI, particularly digital pathology, has the highest external validation and provides the only biomarker with a first signal of predictive value for BCG vs alternatives, increasingly relevant in the context of the global BCG shortage. Integrative models such as PROGRxN-BCa and DeepSurv outperform standard calculators with incremental gains and are freely accessible. Prospective validation, systematic calibration reporting and treatment-by-biomarker interaction analyses, ideally embedded in trials such as the BRIDGE trial (ClinicalTrials.gov identifier: NCT05538663), remain priorities before clinical adoption.