Background: Over 25% of urothelial carcinoma have actionable genetic alterations in FGFR. The clinical trial of the pan-FGFR inhibitor requires thousands of patients with FGFR alterations to be screened; however, the cost is prohibitively expensive. A deep learning model can help predict the occurrence of specific genetic alterations and effectively screen patients with FGFR actionable alterations in bladder cancer. Methods: The prediction model of FGFR oncogenic alterations was trained by using 440 H&E-stained bladder cancer diagnostic slides (from 373 patients) from the OrigiMed Database and validated with 448 slides (from 376 patients) from the additional TCGA-BLCA dataset. The model employs an attention-based, weakly supervised learning method to identify tissue subregions most likely to have the mutations of FGFR1-4, so as to classify the entire slide. Every patch is assigned an attention score, which represents the corresponding contribution to slide classification. The slide-level aggregation rule of attention-based pooling computes a final attention score by weighted average. Results: The FGFR oncogenic alteration predictive AI (FGFRonaP) model was trained, validated and tested on the OrigiMed real-world cohort. The FGFRonaP model could identify patients with FGFR1-4 gene oncogenic aberration on the test set accurately (Precision = 60.0%, Recall = 69.2%, AUROC = 0.839). When the model was tested on the additional TCGA independent cohort, 11 negative samples were included mistakenly while 323 true negative slide and 19 true positive slides were identified correctly. The FGFRonaP model showed an acceptable classification precision (63.3%), much higher than the occurrence rate (26.9%) of FGFR oncogenic alterations, in spite of an imperfect AUROC (0.636) and recall (0.167). Conclusions: FGFRonaP is feasible to serve as an in-house or monocentric application to pre-screen the potential FGFR-altered patients. It could be useful and efficient in preliminary screening to reduce workloads of genetic testing, and thus help pharmaceutical companies improve the proportion of candidates with FGFR-alterations in all patients who undergo genetic testing and cut expensive costs for enrolling patients.