Many studies have shown that MALDI-TOF MS is not only a faster approach than existing AST methods but also much more cost-effective17. In order to leverage the full capabilities of MALDI-TOF MS in AST, combining it with other analysis techniques, such as machine learning after dimensional reduction, is necessary to extract meaningful information for resistance predictions. With high-performing models, such a workflow would be relevant and reasonably surpass traditional AST methods in practice. Although several other studies on machine learning applications in MALDI-TOF MS have already been conducted on various pathogens, there are no studies at the time of writing that specifically address nosocomial infections such as S. epidermidis. Furthermore, most publications focus only on one antimicrobial class at a time (e.g., glycopeptides like vancomycin) rather than a more diverse group of differing antimicrobials necessary for a holistic evaluation21. Frequent and indiscriminate use of strong antibiotics like vancomycin increases the prevalence of resistant bacteria, so using these drugs only when necessary helps slow the emergence of superbugs24.As such, this study serves as a consideration of the potential for mass spectrometry to be combined with machine learning to rapidly predict AMR for a variety of different antibiotics in nosocomial infections. The high AUROC scores up to 0.95 and AUPRC scores up to 0.97 achieved by this study can be attributed either to the quality and scale of the S. epidermidis dataset or the usage of feature selection as a technique to eliminate irrelevant features and prevent the creation of an overcomplicated model. Indeed, the original study conducted by the authors of the DRIAMS-A dataset achieved maximal AUROC scores of 0.74, 0.74, and 0.80 for Staphylococcus aureus, Escherichia coli, and Klebsiella pneumoniae, respectively4, which are all lower than or equal to the lowest average AUROC score of 0.80 for fusidic acid in the antibiotics modeled for S. epidermidis. In addition, their corresponding AUPRC scores were 0.49, 0.30, and 0.33, respectively, which are significantly overshadowed by the near-perfect AUPRC scores achieved by the best-performing S. epidermidis models. Although the results achieved by this study were significantly more performative than those achieved in the original study, it is essential to remember that the microbe of focus is different in this study, and other factors may have been involved in the performance of these models.RF was specifically chosen for feature selection due to its advantages over other methods. For one, previous studies have shown that the performance of models trained with features filtered by RF has outperformed those filtered by other classifiers like K-Nearest Neighbors, Linear Discriminant Analysis, and SVM20. In addition, although SHAP was later used for model explanation, it was not used for feature selection because the time required for SHAP analysis increases significantly with each added column, making it unsuitable for extremely high-dimensional datasets. RF provides sufficient performance for preliminary feature filtering and reduces the number of features to a size that can be more efficiently handled by SHAP analysis.As shown in Fig. 3, despite high average AUROC scores, the classifiers for penicillin and tigecycline had relatively high standard deviations for AUROC scores between the ten separate train-test splits. Such performance is likely explained by the fact that these classifiers are subject to highly imbalanced class prevalences, with both approaching nearly no susceptible or resistant isolates, respectively. In many cases, imbalanced datasets can be fixed using oversampling techniques such as Synthetic Minority Oversampling TEchnique (SMOTE). However, due to the high dimensionality of each mass spectra instance, consequential data points are often sparsely distributed, making it difficult for SMOTE to generate meaningful synthetic samples. Past studies with other high dimensional datasets like genome data have demonstrated this, showing negligible improvement in model performance after applying SMOTE25. On the other hand, undersampling techniques can lead to significant information loss. Due to the overwhelming difference in numbers between the positive and negative classes within imbalanced datasets in DRIAMS-A, discarding majority class instances to match a small minority can result in a significant loss of diversity and meaningful patterns. As such, holistic metrics like AUROC and AUPRC are crucial for capturing and evaluating model performance based on both the majority and minority classes at the same time.SHAP analysis provided insight into functions of proteins that have already been documented as well as evidence for uncharacterized biomarkers—if a specific mass value corresponds to a protein belonging to the same or related species that is scarcely annotated, then it could indicate a resistance protein for future investigation. For example, the feature bin of 1,916 in the ciprofloxacin model contributed significantly to model performance and was identified to be a known but minimally documented protein at 6,812 Da. Moreover, the second most relevant feature for predicting ciprofloxacin resistance was calculated to be the feature bin of 1,604, which corresponds to UPF0337 protein SE_0604 at 6,812 Da. A different MALDI-TOF MS study conducted on S. aureus found that another UPF0337 protein that underwent a mutation reduced antibiotic binding affinity through a molecular docking simulation, confirming that mutations may also play a major role in antibiotic resistance26. On the other hand, another possibility that explains the significance of particular feature bins is the horizontal transfer of resistance-encoding genes between Staphylococci or other bacterial species, such as the closely related S. aureus. For instance, another relevant feature for the ciprofloxacin model corresponding to the feature bin of 2,025 was not identified for S. epidermidis in UniProt but was identified at 8,077 Da as a pathogenicity island family protein in S. aureus.In the case of gentamicin, the most relevant feature bin of 1,917 was identified as transposase at 7,753 Da. Transposase is responsible for the movement of transposons—which may encode antibiotic resistance—from one plasmid to another27. Another interesting feature identified by SHAP analysis was the feature bin of 326, for which, although no documented protein exists for S. epidermidis in UniProt, it corresponds to a delta-hemolysin biomarker at 2,979 Da in S. aureus. Delta-hemolysin is one of the primary virulence agents of S. aureus, is known to increase bacterial resistance to antibiotics indirectly, and has also been identified in some isolates of S. epidermidis28,29. This illustrates the significance of gene transfers between Staphylococci in AMR while also emphasizing the ability of MALDI-TOF MS machine learning models to capture relevant biomarkers rather than noise.Finally, the same pathogenicity island family protein identified by SHAP analysis in ciprofloxacin was also the most relevant feature of the rifampicin model, demonstrating that AMR resistance is multifaceted and similar resistance mechanisms exist between different families of antibiotics. Additionally, the rifampicin model also detected the presence of DNA-directed RNA polymerase subunit omega—encoded by the rpoZ gene—at 7,766 Da for the feature bin of 1,922. Recently, the rpoZ gene and the subunit encoded by it were validated to be a significant factor in biofilm formation, motility, and antibiotic resistance, although in E. coli and not S. epidermidis30.However, interpretation of the roles of the features within the bacteria should be made cautiously due to limitations. For one, if a potential biomarker is identified, it may not necessarily be the protein conferring resistance, but rather a protein that is encoded on the same plasmid as the resistance protein31. Previous studies have demonstrated that the primary mutation responsible for ciprofloxacin resistance are mutations in the gyrA and parC genes, which have masses of 100,144 Da and 91,145 Da, respectively32. Such proteins are outside of the detection range and are therefore not considered when building the model, demonstrating that although model performance may not be optimized to detect the main resistance biomarker directly, highly accurate results can still be achieved by detecting other smaller but related proteins. Additionally, due to the restrictions of MALDI-TOF MS, it is more challenging to detect proteins with larger masses because the current dataset is limited to 20,000 Da.There are several other limitations in the present study. First, dealing with imbalanced datasets is challenging due to the dimensionality of mass spectra. As such, constructing models to analyze MALDI-TOF MS may not be adequate when an antibiotic is nearly always resisted or effective against S. epidermidis. In addition, models trained on one site may not easily transfer to another—even though processes can be nearly identical between sites. There are often other factors, such as differing calibrations of the same instruments, that lead to compounding differences between the mass spectra generated at one site compared to the mass spectra generated at another. This can be seen in the ciprofloxacin resistance model, where the model trained on DRIAMS-A performed relatively well on DRIAMS-C and DRIAMS-D but was not as adequate in DRIAMS-B.
