We found two deviant groups that were unpredictable with clinical models predicting bladder cancer metastasis. The group G consists of patients at high risk of pN+ , but they have pN0. The group P consists of patients at low risk of pN+ , but they have pN+ . We aimed to determine the genetic differences between these two groups. 1603 patients from SEER database were enrolled to build a multivariate model. This model was applied to patients from the TCGA database to distinguish groups G and P. Differentially expressed genes between the two groups were identified. RT-qPCR was used to validate the results in a cohort from FUSCC. Two deviant groups were identified both in the SEER population and the TCGA population. Expression of 183 genes was significantly different between the two groups. 18 genes achieved significant statistical power in predicting lymph node metastasis excluding these two deviant groups. The 18-gene signature outperformed 3 other bladder cancer lymph node prediction tools in 2 external GEO datasets. RT-qPCR results of our own cohort identified NECTIN2 (P = 0.036) as the only gene that could predict metastasis. Our study showed a novel gene screening method and proposed an 18-gene signature highly predictive of bladder cancer metastasis.