Clinical Human-Derived Pathogen Signatures Captured by SERS and Deep Learning for Environmental Exposure Risk Assessment

Microbial contamination and antibiotic-resistant bacteria (ARB) pose significant threats to environmental ecosystems, particularly in regions lacking adequate sanitation. Urinary pathogens of human origin, frequently detected in contaminated water and surfaces, represent early indicators of ARB-related exposure. However, current exposure assessment approaches rely heavily on standard strains, limiting their effectiveness in real-world scenarios. In this study, a label-free platform was developed by integrating surface-enhanced Raman spectroscopy (SERS) with a convolutional neural network (CNN) for the ARB exposure assessment. A comprehensive spectral database consisting of 368 clinical urinary isolates was established. The CNN achieved the highest classification accuracy (97.6%), surpassing random forest (93.1%) and PCA-SVM (91.2%). Robust performance was further confirmed in wastewater samples (92.2%) and independent urine specimens (90.3%). Importantly, SHAP-based interpretation revealed key discriminatory features in the 724-738 cm^-1 region, associated with adenine and tryptophan, thereby enhancing model interpretability. Together, these findings establish a practical and transparent framework for microbial surveillance and ARB exposure assessment. This approach supports real-time monitoring and provides a practical tool for environmental health management.