Background: Low-grade noninvasive (LGTa) bladder cancer is a relatively quiescent but heterogenous malignancy, characterized by downstream recurrences requiring repeated transurethral resections and frequent surveillance. Investigations to elucidate drivers of recurrence have been sparse, but will help risk-stratify patients with LGTa and allow augmentation of follow up protocols. Methods: Patients with LGTa index tumors were stratified by those with no downstream recurrences (nonrecurrent) vs. those with later recurrences (recurrent). RNA sequencing identified differentially expressed genes (DEGs), deconvoluted for cell-type using xCell. Pathologic analysis was performed by a genitourinary pathologist, then a deep-learning artificial intelligence (AI) platform was leveraged to correlate recurrence risk and recurrence-free survival (RFS) based on deep-learning algorithm of segmented nuclei. Results: Thirty index bladder tumors/patients were identified, 18 (60%) of which had later recurrence (Table). There were 238 DEGs recognized, with recurrent tumors expressing signatures for epithelial mesenchymal transition, myogenesis, TNFα signaling via NFκB, and angiogenesis. Recurrent tumors also demonstrated a higher tissue micoenvironment, stroma, and cancer-associated fibroblast score. Pathologic TME analysis validated these findings, with recurrent tumors demonstrating a higher frequency of inverted growth pattern and a higher median stroma percentage. Finally, the AI-derived signature was predictive of recurrence and risk-stratified the cohort (HR= 5.43 [95% CI 1.1-26.76]) for predicting high vs. low risk of recurrence. Patients in the high risk group had a 87.5% recurrence rate while those in the low risk group had a 28.5% recurrence rate (p<0.01). Conclusions: Using a multi-disciplinary approach, we identified key signatures in recurrent LGTa bladder cancer. Characterization of these factors is a critical first step in the risk-stratification of LGTa tumors, and may allow risk-stratification of surveillance protocols and identification of possible targets for chemoprevention trials.