Background: Neuroendocrine (NE) carcinoma is a rare and aggressive variant of muscle invasive bladder cancer (MIBC). Molecular subtyping studies found 5-15% of bladder tumors had transcriptome profiles consistent with NE carcinoma but lacked NE histology (Robertson 2017, Sjödahl 2017). Identifying NE variants may have prognostic implications and modify treatment recommendations. In this study, we present a robust genomic classifier trained to identify NE carcinoma. Methods: Transcriptome-wide expression profiles were generated for 576 MIBC patients collected from seven institutions. Model training included profiles generated from TURBT and RC specimens from 320 patients prior to treatment with NAC or chemo-radiation. The validation cohort consisted of 256 RC specimens (no prior systemic treatment). Using 10 MIBC-related gene sets, a GLMNET model was built to predict patients with a NE carcinoma expression profile. Uni- and multi-variable survival analyses were used to characterize outcomes of the predicted NE tumors. Results: In the training set, hierarchical clustering using a panel of 54 genes showed a cluster of 17 patients (5.3%) with a NE carcinoma expression profile. These patients had significantly worse 1 year progression free survival (65% vs 82% for NE vs overall; p = 0.046). In the validation set, 7 tumors were classified (2.7%) as NE with 4 (57%) patients dying from the disease at 1 year after RC. Within 3 years of RC, 100% (7/7) of patients with NE tumors had died. After adjusting for various clinical and pathological factors, patients with predicted NE tumors had a 6.40 increased risk of all-cause mortality (p = 0.001). Conclusions: We have developed a gene expression signature that predicts a particularly high-risk group that may need treatment intensification, alternative chemotherapy or clinical trials. Validation will be required to assess the potential clinical utility of this NE carcinoma classifier.