Machine learning reveals drivers of microplastic bioaccumulation in fish from a freshwater reservoir ecosystem.

Microplastics (MPs) are increasingly detected in freshwater ecosystems, yet the factors influencing their bioaccumulation in fish, particularly in reservoir systems, remain insufficiently understood. We investigated MPs bioaccumulation in 150 individuals of three commercially important freshwater fish species with contrasting feeding strategies (carnivorous, omnivorous, and filter-feeding) collected from the Aras Reservoir in northern Iran. MPs were detected in 59% of specimens, with a mean abundance of 1.15 ± 0.96 particles per fish, predominantly polypropylene fibers (55%) in the 300-1000 μm size range. To move beyond descriptive monitoring, a two-phase multilevel machine learning approach was applied to identify key drivers of MPs bioaccumulation. In the first phase, Random Forest, Decision Tree, and Support Vector Machine models highlighted feeding strategy and body-size class as interacting biological predictors, revealing complex relationships not captured by conventional statistical analyses. In the second phase, particle characteristics (shape, color, and richness) were incorporated into ensemble models (RF, XGBoost, and GBM), substantially improving explanatory characterization of MP burden. Because these particle-level attributes are intrinsic properties of the ingested assemblage and are not independent of the response variable, the Phase 2 results should be interpreted as exposure profiling rather than ecological prediction from external drivers. Beyond the specific study area, this approach provides a generalizable, data-driven framework for assessing and predicting MPs bioaccumulation in reservoir systems.