Despite heavy alterations to freshwater ecosystems, the reasons for species’ sensitivity and resistance to imperilment are unclear. To address this need, we develop a machine learning framework to predict global imperilment status for 10,631 freshwater fish species using a comprehensive set of environmental, socioeconomic, and intrinsic species-level predictors. Using updated IUCN Red List data, we train and validate Random Forest classifiers to distinguish imperiled (Vulnerable, Endangered, Critically Endangered) from non-imperiled species. We examine the relative influence of 52 variables derived from 12 global sources describing extrinsic environmental and socioeconomic factors and intrinsic species-specific characteristics. Our models achieve higher accuracy for non-imperiled species (90.1%) compared to imperiled species (81.8%), reflecting the greater heterogeneity of threats and conditions driving imperilment. Across models, key predictors include habitat variables, taxonomic order, hydrological characteristics, and disturbance indicators, underscoring the interplay between ecology, geography, and human pressures. This integrative, reproducible approach demonstrates the utility of machine learning for guiding proactive conservation and provides a scalable framework for global biodiversity risk assessment.

Murphy C.A., Olivos J.A., Arismendi I., García-Berthou E., Johnson S.L. & Dunham J. 2026. Environment, taxonomy, and socioeconomics predict non-imperilment in freshwater fishes. Nature Communications 17: 1661.   https://doi.org/10.1038/s41467-025-68154-w