Machine learning predicts pathological response after bladder chemoradiotherapy using routine clinical factors

In 285 bladder cancer patients, a clinical-variable machine learning model predicted pathological complete response with test AUC 0.829.

KEY POINTS

  • This retrospective single-center analysis included 285 patients with cT2–4N0M0 muscle-invasive bladder cancer who received chemoradiotherapy followed by partial cystectomy (n=148) or radical cystectomy (n=137) between 1997 and 2021. Unlike most bladder-preservation cohorts, all included patients had post-treatment surgical specimens available, allowing pathological complete response (ypT0) to serve as the outcome reference.
  • The institutional treatment approach used 40 Gy in 20 fractions to the true pelvis, with concurrent cisplatin administered in two cycles, followed by response assessment and partial or radical cystectomy. The investigators evaluated nine pretreatment clinical variables, including T stage, tumor size, multiplicity, hydronephrosis, bladder neck involvement, concomitant carcinoma in situ, histological grade, variant histology and progression from previous non-muscle-invasive disease.
  • The study randomly allocated 228 patients to training and 57 to independent testing. Among 102 evaluated machine learning models, the selected stacked ensemble (WeightedEnsemble_L4) achieved a training-cohort AUC of 0.921 and a test-cohort AUC of 0.829 (95% CI 0.706–0.932). Overall, 178/285 patients (62%) achieved pathological complete response.
  • Model calibration remained uncertain despite favorable discrimination. The test-set calibration slope was 0.856 (95% CI 0.469–1.639), and performance decreased when the analysis was repeated across additional independent train/test splits, suggesting sensitivity to dataset partitioning and possible model overfitting.
  • The four highest-ranked clinical predictors of incomplete response were clinical T stage ≥T3, tumor diameter >30 mm, concomitant carcinoma in situ and tumor multiplicity. In the training cohort, pathological complete response occurred in 50.5% versus 70.1% for advanced versus lower T stage, 48.1% versus 69.5% for tumors above versus below 30 mm, and 45.9% versus 72.0% for multifocal versus unifocal disease.
  • The cumulative number of adverse clinical factors strongly stratified pathological response. Across the full cohort, complete response rates were 87% with no factors, 66% with one, 49% with two, and 29% with three or more (p<0.001). These groups also showed differences in five-year metastasis-free survival (87.5%, 86.5%, 70.9% and 59.8%, respectively), although the clinical-factor score was exploratory rather than a prospectively validated treatment decision tool.
  • A central limitation is that all patients underwent surgery after chemoradiotherapy, including those managed within a bladder-preservation protocol. Consequently, neither the prediction model nor the reported survival outcomes directly establishes safe omission of cystectomy or long-term success with RT-based bladder preservation alone; external validation and calibration against contemporary definitive CRT regimens are required.

CLINICAL TAKEAWAY

Routine pretreatment clinical characteristics may help estimate the probability of pathological complete response after bladder chemoradiotherapy, potentially informing future risk-adapted bladder-preservation research. However, the proposed machine learning model remains exploratory, and its predicted probabilities should not be used to withhold cystectomy or select definitive bladder preservation without external validation in appropriately treated cohorts.

SOURCE

Clinical and Translational Radiation Oncology

Browse more research Suggest a correction