# The Evaluation of Machine Learning Models Using Matrix‐Assisted Laser Desorption/Ionization Time‐of‐Flight Mass Spectrometry (MALDI–TOF–MS) Spectra for the Prediction of Antibiotic Resistance in Klebsiella pneumoniae

## TL;DR

Machine learning models using MALDI–TOF–MS spectra can quickly and accurately predict antibiotic resistance in Klebsiella pneumoniae, potentially improving infection treatment and reducing delays.

## Contribution

Demonstrates that ML models using MALDI–TOF–MS spectra can predict antibiotic resistance in K. pneumoniae with high accuracy and rapid turnaround.

## Key findings

- ML models using MALDI–TOF–MS spectra achieved AUROC values above 0.90 for predicting antibiotic resistance in K. pneumoniae.
- Ensemble and deep learning models reached up to 97% accuracy and 0.99 AUROC for carbapenem resistance prediction.
- Diagnostic turnaround time was reduced from days to minutes using existing MALDI–TOF–MS equipment.

## Abstract

Antimicrobial resistance in Klebsiella pneumoniae poses a major clinical challenge, driving development in rapid, diagnostic strategies that extend beyond conventional susceptibility testing. Twenty‐three studies demonstrate that using matrix‐assisted laser desorption/ionization time‐of‐flight mass spectrometry (MALDI–TOF–MS) spectra to create machine learning (ML) models yields rapid and accurate predictions of antibiotic resistance in K. pneumoniae. Across these studies, most models focused on carbapenem resistance and achieved Area Under the Receiver Operating Characteristic Curve (AUROC) values consistently above 0.90, with ensemble algorithms, particularly Random Forest, XGBoost, and Light Gradient Boosting Machine, and deep learning models such as Convolutional Neural Networks attaining accuracies as high as 97% and even AUROCs reaching 0.99 or higher. Sample sizes ranged from 35 to over 15,000 isolates, reinforcing the robustness of these findings across diverse clinical settings. In addition to high discrimination performance, this evaluation reports that ML models developed using MALDI–TOF–MS spectra shorten diagnostic turnaround from days (48–96 h with conventional methods) to minutes or hours, using existing MALDI–TOF–MS equipment for economical implementation. However, ML diagnostic tools remain constrained by limited external validation, spectra preprocessing protocols, and variability between different MALDI–TOF–MS platforms. These limitations may restrict model generalizability and clinical translation, highlighting the need for standardized workflows and larger multicenter evaluations.

Machine learning models developed from matrix‐assisted laser desorption/ionization time‐of‐flight mass spectrometry spectra can rapidly predict antimicrobial resistance in Klebsiella pneumoniae. By transforming routine spectral data into accurate resistance profiles within minutes, this approach could reduce diagnostic delays and support earlier targeted therapy, improving infection management and antimicrobial stewardship. However, current models are limited by external validation data sets, which may restrict model generalizability on unseen data, a problem compounded by different lab workflows used to obtain spectra.