Effect of frailty status and histopathological features on upstaging of non-muscle-invasive bladder cancer: a critical analysis based on machine learning.

AIM: This study aimed to develop a machine learning (ML)-assisted model to predict the risk of upstaging (subsequent higher stage on repeat pathology) in non-muscle-invasive bladder cancer (NMIBC).

METHODS

A retrospective cohort study was conducted on 380 patients with NMIBC. Comprehensive data on demographics, comorbidities, frailty (modified frailty index, mFI), nutritional status (nutritional risk index, NRI), inflammatory indices, and detailed histopathological features were collected. The primary outcome was pathological upstaging, defined as a diagnosis of a higher T stage (from Ta to ≥ T1 or from T1 to ≥ T2) on a subsequent procedure. Multivariate logistic regression and six machine learning models were developed and evaluated using tenfold cross-validation, and their performances in predicting upstaging were directly compared.

RESULTS

Pathological upstaging occurred in 84 patients (22.1%). Multivariate logistic regression identified initial T1 stage (OR: 13.54), high tumor grade (OR: 4.60), lymphovascular invasion (LVI) (OR: 3.67), frailty (mFI ≥ 2, OR: 2.68), and lower estimated glomerular filtration rate (eGFR) as factors independently associated with upstaging (AUC: 0.702). Machine learning models, particularly the support vector machine (SVM), demonstrated superior predictive performance (AUC: 0.796). Analysis using supervised learning algorithms confirmed tumor grade as the strongest associated factor, followed by initial T stage, the presence of lymphovascular invasion on initial TURBT (transurethral resection of the bladder tumor), frailty, eGFR, and occupational exposure.

CONCLUSIONS

This study demonstrates that artificial intelligence models provide a superior framework for predicting subsequent pathological upstaging in NMIBC compared to traditional multivariate logistic regression. The ML-driven analysis adequately validated established clinical risk factors associated with upstaging. Among all variables evaluated, high tumor grade emerged as the most powerful and clinically significant factor associated with upstaging. Thus, ML-assisted tools may feasibly be integrated into clinical practice to enhance risk stratification.