e16574 Background: Circulating tumor DNA (ctDNA) dynamics have emerged as a promising biomarker of prognosis and therapeutic response in urothelial cancer (UC). However, the optimal approach to characterize the relationship between serial ctDNA changes and overall survival (OS) across various UC clinical disease states remains unclear. This study systematically evaluated multiple methods to describe ctDNA kinetics, including qualitative, quantitative, and data-driven machine learning (ML) strategies, to identify relationships between longitudinal ctDNA patterns and OS outcomes. Methods: We performed a single-institution retrospective cohort study of patients with UC, including variant histologies, with ≥2 tumor-informed ctDNA assays (Signatera) available. A 6-month landmark analysis was employed to minimize immortal time bias. Three ctDNA kinetic classification strategies were applied: (1) Qualitative—binary detectability patterns (clearance, conversion, persistent positive/negative); (2) Quantitative—percent change from baseline (major response ≥90% reduction, partial 30-89%, stable 0-29%, progressive ≥20% increase); (3) ML—unsupervised k-means clustering (k = 5) based on descriptive and kinetic ctDNA features. The primary endpoint was discrimination of OS from 6-month landmark using Harrell's C-index. Pairwise differences were tested via bootstrap resampling (500 iterations). Sensitivity analyses were performed stratified by treatment setting. Results: Among 215 evaluable patients, therapy settings included: neoadjuvant 38 (17.7%), adjuvant 51 (23.7%), metastatic 71 (33.0%), and no systemic therapy 85 (39.5 %). OS C-indices were: ML 0.902 (95% CI 0.855-0.934), Qualitative 0.883 (0.843-0.919), Quantitative 0.852 (0.818-0.899). Pairwise comparisons showed no significant differences (all p > 0.10). Subgroup analysis showed no statistically significant differences by treatment setting (all p > 0.10), although quantitative performance was numerically lower in metastatic setting (C-index 0.709 vs 0.799 qualitative; 0.839 ML), with a −0.13 difference relative to ML (95% CI −0.203–0.042; p = 0.376). Conclusions: All three ctDNA kinetic classification approaches demonstrated strong OS discrimination (C-index > 0.85) with no significant performance differences across treatment settings, which suggests that multiple analytic strategies may be suitable depending on clinical context. Larger, prospective studies are needed to define optimal approaches and establish context-specific thresholds.
Elikan et al. (Thu,) studied this question.