UCtracker: A Deep Learning–Based DNA Methylation Model for Noninvasive Diagnosis and Recurrence Surveillance of Urothelial Carcinoma in a Prospective Study

Noninvasive diagnosis and longitudinal surveillance of urothelial carcinoma (UC) remain clinically challenging. Here, we developed and prospectively validate UCtracker, a urine DNA methylation-based deep learning model for UC detection and postoperative recurrence monitoring. UC-specific differentially methylated regions (DMRs) were identified by high-depth whole-genome bisulfite sequencing of UC tissues, paired adjacent normal tissues, and non-tumor urine samples. UCtracker was constructed using the top 2000 hypomethylated DMRs and a convolutional neural network-bidirectional long short-term memory architecture. In an internal validation cohort (n = 165), UCtracker achieved a sensitivity of 94.6% and a specificity of 94.4%, and maintained high performance in an independent multicenter cohort (n = 55), with a sensitivity of 90.6% and a specificity of 91.3%. UCtracker showed higher sensitivity than UroVysion fluorescence in situ hybridization for T1 tumors, high-grade tumors, and bladder UC. Subsampling analyses demonstrated stable diagnostic performance even at ultralow sequencing depths. In postoperative surveillance, longitudinal urine profiling of 131 samples from 48 UC patients detected 94.1% of recurrence events and identified recurrence up to 250 days before clinical confirmation. These findings support UCtracker as a highly accurate and cost-effective urine-based tool for UC diagnosis, postoperative surveillance, and personalized patient management.