Feasibility study on rapid identification of methicillin-resistant Staphylococcus aureus by different algorithms of MALDI-TOF mass spectrometry / 中华检验医学杂志

ABSTRACT

Objective:To explore the feasibility of rapid identification of methicillin-resistant Staphylococcus aureus using different algorithms of the matrix-assisted laser desorption/ionization time of flight (MALDI-TOF) mass spectrometer. Methods:Totally 314 clinical isolates of Staphylococcus aureus were selected from the bacterial bank at Beijing Tongren Hospital from January 2017 to June 2019. The samples were identified by MALDI-TOF MS, and screened by cefoxitin disk method (inhibition ring diameter ￡21 mm) and PCR mecA gene. The strains were divided into a methicillin-resistant Staphylococcus aureus (MRSA) group (130 strains) and a methicillin-susceptible Staphylococcus aureus (MSSA) group (184 strains). Then, after collecting the spectrograms of these samples using formic acid extraction, the MRSA group and MSSA group were divided into three subgroups each, namely MRSA-1 (43 strains), MRSA-2 (42 strains), MRSA-3 (45 strains) and MSSA-1 (60 strains), MSSA-2 (61 strains) and MSSA-3 (63 strains). The groups were studied using genetic algorithm (GA), fast classification algorithm (QC) and supervised neural network algorithm (SNN) in the ClinProTools software on the Bruker MALDI-TOF mass spectrometer, and the convolutional neural network algorithm (CNN) in the Ex-SmartSpec software on the Zhongyuan Hui-Ji mass spectrometer. These studies were repeated for 3 rounds. The first round with MRSA-1 and MRSA-2, MSSA-1 and MSSA-2 being model groups, MRSA-3 and MSSA-3 being validation groups. The validation groups were rotated for each round. The areas under the receiver operating characteristic (ROC) curve expansions of the four algorithms were used to confirm each program′s performance. Then, 38 MRSA strains and 40 MSSA clinical strains were selected from the bacterial bank of the Laboratory of Beijing Tongren Hospital from July 2019 to December 2019, and were put through the formic acid extraction method to collect their spectra. These samples were tested independently with their convolutional neural network models. Results:After three rounds of modeling and verification, the areas under the ROC curves of the three Bruker ClinProTools programs were as follows for genetic algorithm, the areas were 0.89, 0.74, and 0.64 respectively; for fast classification algorithm, the areas were 0.77, 0.95, and 0.94 respectively; and for supervised neural network algorithm, the areas were 0.90, 0.98, and 0.98 respectively. The areas under the ROC curves of the convolutional neural network algorithm with Zhongyuan Huiji mass spectrometer′s Ex-SmartSpec software were 0.95, 0.99, and 0.99 respectively. The independent test results of convolutional neural network algorithm showed that these results have an accuracy, specificity, sensitivity and AUC of 88.82% (810/912), 81.15% (779/960), 84.88% (1 589/1 872) and 0.92 respectively.Conclusions:The supervised neural network algorithm of Bruker′s ClinProTools and the convolutional neural network algorithm of Zhongyuan Hui-Ji mass spectrometer′s EX-Smartspec is clinically acceptable for rapid identification of MRSA performance indicators. Using convolutional neural network algorithm and MALDI-TOF mass spectrometry, MRSA strains can be identified quickly, providing timely advice for clinical medications.