Abstract
Antimicrobial resistance (AMR) is one of the most pressing public health challenges of the 21st century. This study aims to evaluate the efficacy of mass spectral data generated by VITEK(®) MS instruments for predicting antibiotic resistance in Staphylococcus aureus, Escherichia coli, and Klebsiella pneumoniae using machine learning algorithms. Additionally, the potential of pre-trained models was assessed through transfer learning analysis. A dataset comprising 2229 mass spectra was collected, and classification algorithms, including Support Vector Machines, Random Forest, Logistic Regression, and CatBoost, were applied to predict resistance. CatBoost demonstrated a clear advantage over the other models, effectively handling complex non-linear relationships within the spectra and achieving an AUROC of 0.91 and an F1 score of 0.78 for E. coli. In contrast, transfer learning yielded suboptimal results. These findings highlight the potential of gradient-boosting techniques to enhance resistance prediction, particularly with data from less conventional platforms like VITEK(®) MS. Furthermore, the identification of specific biomarkers using SHAP values indicates promising potential for clinical applications in early diagnosis. Future efforts focused on standardizing data and refining algorithms could expand the utility of these approaches across diverse clinical environments, supporting the global fight against AMR.