Rapid prediction of vancomycin-resistant Enterococcus faecium using MALDI-TOF mass spectrometry and machine learning

Area of Science:

• Microbiology
• Computational Biology
• Clinical Diagnostics

Background:

• Vancomycin-resistant Enterococcus faecium (VREfm) is a critical global health threat.
• Current VREfm susceptibility testing methods are time-consuming, delaying treatment.
• Rapid diagnostics are essential for effective antimicrobial stewardship.

Purpose of the Study:

• To develop a rapid and interpretable machine learning classifier for VREfm detection.
• To utilize MALDI-TOF MS spectra for predicting vancomycin resistance in E. faecium.
• To enhance model generalizability using diverse training datasets.

Main Methods:

• Retrospective analysis of 268 E. faecium isolates (VREfm and susceptible).
• MALDI-TOF MS spectra preprocessing and feature selection using extreme gradient boosting.
• Hybrid training strategy combining local isolates with the multi-center DRIAMS repository.
• Temporal validation on independent isolates and performance evaluation using AUROC and Brier score.

Main Results:

• The k-nearest neighbors classifier achieved an AUROC of 0.90 and F1 score of 0.683 in temporal validation.
• Hybrid training models maintained clinically useful discrimination (AUROC > 0.80).
• External-only models showed performance at chance level (AUROC ≈ 0.50).

Conclusions:

• MALDI-TOF MS coupled with machine learning provides a rapid method for predicting VREfm.
• A hybrid local-multi-center training approach addresses limitations of local sample availability.
• This strategy facilitates the clinical implementation of spectral-based antimicrobial resistance surveillance.