The lack of information on the extinction risk of most species is a fundamental challenge in prioritizing conservation strategies and bending the curve of current biodiversity decline. Machine learning methods have shown promising potential to fill this gap, but their applicability remains to be validated at different taxa (especially aquatic species) and spatial scales. We assessed the extinction risk of 1,162 freshwater fish species in China that have not yet been included in the latest IUCN Red List using multiple neural network algorithms based on datasets of species occurrences, biological traits, phylogeny, and relevant environmental layers. The best deep learning models dramatically improved the assessment coverage from 29.9% (496 species) to 93.2–93.9% (1,545–1,557 species) of the whole fauna with an accuracy of 95.4–99.0%. By combining our prediction results with the IUCN Red List, we found that 23.8–26.5% (394–440 species) of Chinese freshwater fishes were identified as possibly threatened species, which is roughly four times the IUCN assessment. Newly assessed species and threatened species were mainly from the orders Cypriniformes (prediction added: 837–846 species; final threatened: 325–350 species), Siluriformes (113–122; 28–37) and Perciformes (74–76; 18–25). The increase in threatened species richness based on predictions was led by the upper reaches of the Pearl and Yangtze. Overall, our findings suggest that deep learning algorithms can provide robust and time-saving assessments of extinction risk for entire freshwater fish fauna on a large national scale, thereby facilitating relevant conservation prioritization.