Rapid Screening of Methicillin-Resistant Staphylococcus aureus Using MALDI-TOF MS and Machine Learning: A Randomized, Multicenter Study

Abstract

Methicillin-resistant Staphylococcus aureus (MRSA) is a major cause of healthcare-associated infections including bacteremia. The rapid detection of MRSA is essential for prompt treatment and improved outcomes. However, traditional MRSA screening and confirmatory tests based on bacterial cultures with antimicrobial susceptibility tests and/or molecular diagnostics are time-consuming (>2 days), labor-intensive, and costly. We report that AMRQuest software, which was developed using logistic regression-based machine learning and matrix-assisted laser desorption/ionization-time-of-flight spectra of S. aureus isolates, can be successfully implemented in clinical microbiology laboratories to screen MRSA and identify bacterial species simultaneously, with the cefoxitin disk diffusion test as a reference. Analytical sensitivity, specificity, percent agreement, and Cohen's kappa values were calculated to determine the accuracy of the AMRQuest software. The minimum sample size of the testing set for statistical analysis was determined considering the local prevalence of MRSA infections. MRSA screening was performed using 537 consecutive S. aureus isolates, including 231 MRSA and 306 methicillin-susceptible S. aureus isolates, from three tertiary-care hospitals. The results from the AMRQuest software were similar to those obtained using the reference method, cefoxitin disk diffusion testing, making it a powerful method for the rapid detection of MRSA prior to traditional antibiotic resistance testing.

1

Workflow of the AMRQuest software which is the MALDI-TOF MS-based methicillin-resistant S. aureus (MRSA) screening system. (a) Processing of MALDI-TOF mass spectra for machine learning. (b) Training process and structure of AMRQuest software. AMRQuest software was designed as a module that can use MALDI-TOF spectra after S. aureus identification without additional MALDI process. (c) Randomized, single-blind study design for clinical evaluation of AMRQuest software. Minimum number of MRSA and methicillin-susceptible S. aureus (MSSA) isolates were calculated by using meta-analysis of the disease prevalence in Korea.

2

Contribution of feature m/z ranges to MRSA screening. (a) Shapley additive explanations (SHAP) values of the 20 most impactful feature m/z ranges. A positive SHAP value represents the contribution to the determination of MRSA. (b) ANOVA results of the top 20 features ordered by magnitude of upregulation in MRSA or MSSA and their SHAP plot. * Previously identified m/z features. ^§ m/z features in the top 20 for both ANOVA and SHAP analysis.

3

MRSA screening performance of AMRQuest software. (a) Confusion matrix, (b) Violin plot, (c) the receiver operating characteristic curve, and (d) the precision-recall curve of analysis results using AMRQuest software from testing set including MRSA (n = 231) and MSSA (n = 306). The whiskers in violin plot indicated the 5th and 95th percentiles. The horizontal dotted line (score = 0.5) represents the cutoff AMRQuest score. The score range of 0.4 to 0.6, indicated by the gray zone, represents the low-confidence prediction range. The ROC curve and the PR curve were plotted considering that the S. aureus isolates with the AMRQuest score in the gray zone were false positives or false negatives. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve.