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

In 30 seconds

This systematic review assessed artificial intelligence and machine-learning models predicting outcomes after intravesical BCG for non-muscle-invasive bladder cancer. It included 15 studies involving more than 24 900 patients, covering digital pathology, radiomics, molecular, clinicopathological and urinary data. The digital-pathology CHAI platform, validated across 12 international centres, stratified recurrence, progression and BCG-unresponsive disease; it also showed the only reported treatment-by-biomarker signal, with a significant BCG versus gemcitabine/docetaxel interaction (P = 0.029). Integrative models outperformed standard calculators with modest gains, but prospective validation was absent.

Key findings

• CHAI stratified high-grade recurrence (hazard ratio [HR] 2.08), progression (HR 3.87) and BCG-unresponsive disease (HR 2.31).
• CHAI showed a significant BCG versus gemcitabine/docetaxel interaction (P = 0.029), the review’s only signal of predictive value.
• PROGRxN-BCa and DeepSurv had C-indices of 0.79 and 0.881, respectively; gains over standard calculators were modest (ΔC-index 0.05-0.10).
• Only 40% of studies had external validation, none were prospective, and five were judged at high risk of bias.

Why it matters

More reliable identification of patients unlikely to benefit from BCG could inform evaluation of treatment alternatives as these enter practice. The review suggests digital pathology may help stratify outcomes and offers an early treatment-selection signal, while integrative models may improve risk prediction. However, the evidence remains insufficient for routine adoption without prospective validation and better calibration reporting.

What to keep in mind

No included study was prospectively validated; only 40% performed external validation, and five studies were at high risk of bias.