Characterizing frailty and its drivers in bladder cancer using machine learning-derived vulnerability and resource phenotypes.

Abstract

724 Background: Frailty predicts surgical complications and mortality in bladder cancer. Guidelines endorse evaluating for frailty, ideally via Comprehensive Geriatric Assessment (CGA) prior to determining a treatment plan. The objective of this study was to apply artificial intelligence to prospectively collected CGA data to not only quantify frailty but also to characterize its drivers in a clinically interpretable framework that can inform management. Methods: Urothelial cancer patients enrolled from a multidisciplinary bladder cancer clinic (9/2020–7/2021) completed a CGA incorporating validated assessments of functional status, multimorbidity, nutrition, cognition, and mental health, augmented with CT-derived body composition assessments. A random forest classifier was trained to identify predictive features of frailty. These predictive features were then transformed into positive (resources) and negative (vulnerabilities) scores and analyzed both individually and by domain. A principal component analysis (PCA) biplot was generated to provide an interpretable visualization of frailty phenotypes, with quadrants corresponding to distinct vulnerability domains. A Cox proportional hazards model was constructed using patient death and follow-up time as outcomes, adjusted for pathological stage, to assess prognostic significance. Results: The cohort included 67 patients (median age 71, 16.4% female), the majority with muscle-invasive disease (77.6%). Random forest modeling identified comorbidity burden, ECOG status, grip strength, and psychosocial resilience factors among the most strongly associated features. The PCA biplot stratified patients into distinct vulnerability/resource phenotypes, with physical performance and comorbidity burden mapping to vulnerability quadrants, and psychosocial resilience features (hope, optimism, self-kindness) mapping to resource quadrants. In Cox modeling, comorbidity score was associated with increased mortality risk (HR 1.15, p =0.04), whereas demographic resource score was protective (HR 0.78, p =0.024), independent of pathological stage (C-index 0.71). Conclusions: In this prospective observational cohort study, we leveraged machine learning to generate a paradigm on which an individual's personalized vulnerability versus resource profile can be mapped that highlight actionable drivers of frailty that may be targeted with interventions to mitigate those drivers or buttress resources prior to anticipated medical or surgical interventions. The resultant frailty atlas offers a novel, interpretable framework that distinguishes physical, psychosocial, and multimorbidity-related vulnerabilities, with demonstrated prognostic significance. This approach moves beyond simple frailty classification to provide clinically relevant, data-driven insights that can inform personalized prehabilitation and treatment planning in bladder cancer. Findings also provide proof of concept of the complex heterogeneity of patient frailty.