Explainable deep learning assisted SERS detection of cathinones in environmental water

Synthetic cathinones are a class of artificially synthesized central nervous system stimulants, classified as new psychoactive substances, which possess potent psychoactive effects and pose significant health risks. Due to their molecular-structure similarities, synthetic cathinones pose significant challenges for traditional detection methods in achieving effective differentiation. This study proposes an interpretable deep convolutional network-assisted surface-enhanced Raman spectroscopy (SERS) detection technique for the identification of cathinones in environmental water. First, Raman and SERS spectra of four standard cathinone substances were obtained, and their characteristic vibrational modes were analyzed using density functional theory calculations. Furthermore, a highly sensitive SERS detection method was developed for spiked samples, achieving high sensitivity detection of four synthetic cathinones in environmental water, with a minimum detection concentration of 2 ng/mL. Finally, one-dimensional convolutional neural network (1D-CNN) classification model incorporating channel and spatial attention mechanism was successfully constructed, achieving a classification accuracy of 99.8% for the aforementioned four substances and negative samples. Additionally, the gradient-weighted class activation mapping (Grad-CAM) algorithm was employed to identify SERS peaks representing the most critical discriminative features, effectively elucidating the classification mechanism of the 1D-CNN model and enhancing the model's transparency and interpretability.