Culture-Based Assessment of Presumptive Resistant Bacterial Taxa in the Urban Danube River near Novi Sad: Environmental Associations Revealed by Machine Learning

Background/Objectives: Environmental dissemination of antimicrobial resistance (AMR) is increasingly driven by wastewater-impacted aquatic systems, yet the key factors controlling multidrug-resistant (MDR) bacterial distribution remain unclear. This study evaluated environmental factors associated with MDR bacteria in the urban Danube River (Novi Sad, Serbia) using a machine learning framework. Methods: Surface-water and wastewater samples were collected during summer and autumn 2024. Bacterial isolates were obtained through membrane filtration onto chromogenic media and identified using MALDI-TOF MS. Physicochemical parameters (including COD, BOD5, turbidity, pH, and temperature) were used as predictors in seven machine learning models (ANN, RF, SVM, XGB, MARS, TREE, NB). Model performance was assessed using AUC, accuracy, and error metrics. Results: Wastewater samples showed higher bacterial abundance and taxonomic richness than river surface-water samples, with frequent recovery of Klebsiella pneumoniae, Escherichia coli, Aeromonas veronii, and Pseudomonas spp. Tree-based models (RF, XGB) performed best. Organic pollution indicators, turbidity, pH, and water temperature were the most prominent factors. Conclusions: Wastewater-related pollution gradients, reflected by organic load parameters, turbidity, pH, and water temperature, were associated with the occurrence of selected bacterial taxa recovered on selective media. These associations were more pronounced in wastewater samples, while river surface-water samples showed lower abundance and taxonomic richness but still contained selected environmental and opportunistic taxa. Because antimicrobial susceptibility testing and molecular confirmation of resistance determinants were not performed, the findings should be interpreted as culture-based and exploratory. Machine learning approaches may support environmental screening and hypothesis generation for AMR-oriented surveillance, but future studies should include standardized phenotypic and molecular confirmation.