From Correlation to Causality: Identifying Potential Environmental Drivers of Pathogenic Antibiotic-Resistant Bacteria in River Water Using Causal Machine Learning.

Pathogenic antibiotic-resistant bacteria (PARB) pose a serious public health threat within the One Health framework, yet identifying their potential environmental drivers in complex aquatic systems remains a challenge. This study systematically compared correlation analysis, explainable machine learning, and causal machine learning within a unified framework. Both Spearman correlation and explainable machine learning identified numerous potentially important factors, notably non-antibiotic pharmaceuticals such as carbamazepine and bezafibrate. However, causal inference via double machine learning, which controls for confounders and interaction effects, revealed a distinctly different driver profile. Under predefined assumptions, this approach estimated potential causal effects for dissolved oxygen, the nitrate-to-ammonium ratio, specific antibiotics (roxithromycin, azithromycin), and non-antibiotic compounds (acenaphthene, 2-chloroanthracene). Taxon-specific analysis further showed that Aeromonas aligned closely with the overall PARB causal profile, whereas Pseudomonas responded primarily to oxidation-reduction potential. Functional profiles suggested potential stress-adaptation mechanisms related to signal transduction and metabolic regulation pathways. By shifting from associative prediction to causal inference, this causal machine learning-guided framework provides a robust analytical basis for identifying environmental drivers and informing targeted management of PARB risks in aquatic ecosystems.