AI-powered rapid detection of multidrug-resistant Klebsiella pneumoniae with informative peaks of MALDI-TOF MS

TL;DR

This study uses AI and mass spectrometry to quickly predict antibiotic resistance in Klebsiella pneumoniae, a dangerous bacteria, using interpretable spectral markers.

Contribution

The novel contribution is the development of interpretable machine learning models using MALDI-TOF MS data to predict resistance to specific antibiotics in K. pneumoniae.

Findings

Machine learning models achieved 78.58% accuracy in predicting resistance to ciprofloxacin, cefuroxime, and ceftriaxone in K. pneumoniae isolates.

Resistance-associated m/z signals like 3657, 4341, and others were consistently enriched in resistant isolates, providing interpretable spectral markers.

Model performance was stable over time but declined across hospitals, indicating geographic variability in resistance profiles.

Abstract

Klebsiella pneumoniae is a highly virulent superbug with rising antibiotic resistance worldwide. While matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) has transformed microbial identification, its application to antimicrobial resistance prediction remains underexplored, particularly for large clinical cohorts. In this study, we developed machine-learning models with feature-level interpretability using MALDI-TOF MS data to rapidly predict resistance to ciprofloxacin (CIP), cefuroxime (CXM), and ceftriaxone (CRO) in K. pneumoniae.

Using more than 28 000 isolates from two hospitals, the best-performing models reached an independent test accuracy of 0.7858, with sensitivity of 0.7289 and specificity of 0.8127. Several resistance-associated m/z signals—including 3657, 4341, 4519, 4709, 5070, 5409, 5921, 5939, and 6516—were consistently enriched in resistant isolates, offering interpretable spectral markers linked to resistance. Performance remained stable in time-based validation but declined across hospitals, suggesting sensitivity to geographic variability in resistance profiles. Overall, this study demonstrates that combining MALDI-TOF MS with machine learning enables rapid and interpretable prediction of resistance to commonly used fluoroquinolone and cephalosporins in K. pneumoniae. These findings highlight the clinical potential of such models for supporting empiric therapy and emphasize the importance of incorporating local data or adaptive strategies to improve generalizability across healthcare settings.

Data available on request from the authors.

Figures