Background: The detection of bladder cancer in patients presenting with hematuria or for recurrence surveillance is routinely accomplished with a combination of urine cytology and cystoscopy, each with its inherent limitations. Cystoscopy is an invasive procedure, whereas urine cytology lacks the desired level of sensitivity. This study describes the development of four gene expression classifiers (GECs) optimized for the non-invasive detection of both high-grade and low-grade urothelial carcinoma in patients presenting with hematuria or for bladder cancer recurrence surveillance. Methods: Biomarker identification was carried out via microarray analysis of patient urine samples collected at 9 sites in a training set (n=241) composed of urine samples from patients undergoing cystoscopy for hematuria or recurrence surveillance. Identified mRNA biomarkers (n=171) were transferred to the NanoString nCounter platform and an assay was developed that involves the direct analysis of urine sediment lysates, without mRNA purification. NanoString data collected on 261 patient samples was used to generate four GECs optimized for the detection high-grade and low-grade urothelial carcinoma in hematuria or recurrence surveillance patients. Results: The GEC developed for the detection of high-grade urothelial carcinoma in patients presenting with hematuria (n=123) performed with a cross-validated ROC AUC of 0.96, while the GEC for low-grade performed with an AUC of 0.81. In the recurrence surveillance cohort (n=120) the GEC developed for the detection of high-grade performed with an AUC = 0.86 and low-grade with an AUC = 0.61. Conclusions: These results establish the feasibility of using a urine-based gene expression classifier to detect urothelial carcinoma and also to distinguish between high-grade and low-grade. An ongoing multicenter clinical trial will allow us to validate test performance on a larger independent test set of prospectively collected urine samples.