Interpretable hybrid attention CNN for SERS detection of five etomidate analogs in environmental water

SERS detection of five etomidate compounds in environmental water was achieved. • The minimum observable concentration of the five etomidate compounds in environmental water was 2 ng/mL. • CNN-AT model achieved 99.27% accuracy with the lowest standard deviation of 0.33%. • A visual analysis was conducted on the features extracted by the CNN-AT model. Environmental water bodies are directly related to human health. Monitoring emerging pollutants such as etomidate analogs in environmental water can provide important support for combating illegal drug abuse. In this study, an interpretable classification strategy was developed to distinguish etomidate analogs with highly similar structures and highly overlapping surface-enhanced Raman spectroscopy (SERS) spectra in spiked environmental water, achieving a minimum observable concentration of 2 ng/mL. Density functional theory (DFT) was further employed to assign substance-specific molecular vibrational fingerprints and to elucidate the relationship between substituents and spectral features. Finally, we developed a convolutional neural network model integrating a hybrid attention mechanism (CNN-AT) for classifying five structurally similar substances, achieving an accuracy of 99.27%. The results show that the developed model outperforms decision tree (DT), partial least squares discriminant analysis (PLS-DA), K-nearest neighbors (KNN), support vector machine (SVM), long short-term memory (LSTM), and Transformer models in accuracy. Moreover, analysis using gradient-weighted class activation mapping (Grad-CAM) visualization indicates that the model's classification focuses on characteristic peaks related to the molecular structure of the drugs, thereby enhancing the model's interpretability and transparency. Download: Download high-res image (100KB) Download: Download full-size image