Applications of AI-assisted interpreter Raman spectroscopy platform for rapid and accurate cefotaxime-resistant Escherichia coli classification

The emergence of cefotaxime resistance in Escherichia coli (E. coli) poses a significant challenge in clinical practice, driving a demand for rapid diagnostics. Here, we introduce a label-free approach combining Raman spectroscopy with deep learning to rapidly distinguish between cefotaxime-resistant and susceptible E. coli strains. Using clinically isolated samples, we developed classification models based on AlexNet, ResNet, and PCA-LDA algorithms. Evaluation on the internal test set showed that the AlexNet model achieved superior performance, with an accuracy of 88.13% and an area under the receiver operating characteristic curve (AUC) of 0.944. ResNet and PCA-LDA attained accuracies of 81.25% and 78.34%, respectively, with AUC values above 0.85, highlighting the advantage of deep learning in Raman-based resistance identification. To interpret the decision mechanism, we integrated Binary Stochastic Filtering (BSF) with SHapley Additive exPlanations (SHAP) analysis, identifying discriminative spectral features (tentatively assigned to candidate biomolecules such as phenylalanine, tryptophan, and nucleic acids) associated with E. coli linked to cefotaxime resistance. Especially, an exploratory 1D Transformer model achieved 94.38% accuracy with 8% gain and an AUC of 0.991 with 0.07 gain for Raman spectral classification, through introducing global feature modeling. In conclusion, this work offers a promising tool for rapid resistance screening (<0.5 h), while leveraging interpretable AI to reveal candidate spectral biomarkers, thereby aiding in elucidating resistance mechanisms and guiding targeted therapy.