Machine Learning Based Diagnostic Utility for Predicting Ertapenem Resistance in K.pneumoniae Using MALDI-TOF MS Spectra

Overview

Klebsiella pneumoniae is a gram-negative bacterium responsible for over 600,000 deaths annually worldwide, a number that rivals HIV/AIDS (630,000 deaths, UNAIDS 2024), exceeds malaria (610,000 deaths, WHO 2024), and approaches the upper range of seasonal influenza (290,000–650,000 deaths, WHO/Lancet), yet receives a fraction of the global research attention directed at those diseases. Carbapenems, a powerful class of β-lactam antibiotics, serve as last-resort treatments when other options fail; however, the emergence of carbapenem-resistant strains renders treatment options extremely limited. Ertapenem, a widely used carbapenem, is of particular importance because resistance to it frequently precedes broader carbapenem resistance, making it a critical early warning marker for treatment decisions. Traditional resistance detection methods, including Polymerase Chain Reaction, Whole-Genome Sequencing, and Gene Expression Analysis, remain constrained by long processing times, high infrastructure costs, and an inability to reliably detect novel or uncommon resistance mechanisms. While MALDI-TOF MS (Matrix-Assisted Laser Desorption/Ionization Time-of-Flight Mass Spectrometry) enables rapid bacterial identification within minutes, determining ertapenem susceptibility still requires additional time-consuming culture-based testing. In this study, four supervised machine learning algorithms, namely Logistic Regression, K-Nearest Neighbors (KNN), Multi-Layer Perceptron (MLP), and Random Forest, were integrated with MALDI-TOF MS spectra from the DRIAMS dataset, comprising over 300,000 clinical samples from four Swiss medical institutes, filtered to 397 Klebsiella pneumoniae isolates. Class imbalance was addressed through systematic evaluation of resampling strategies including undersampling, SMOTE, and ADASYN, with undersampling selected for its superior macro recall and lowest false-negative rate, a critical priority. Logistic Regression achieved the strongest performance, with 86% test accuracy, precision of 0.81, recall of 0.82, F1-score of 0.81, and an AUROC of 0.83, outperforming prior MALDI-TOF-based antimicrobial resistance prediction baselines. Unlike previous studies that examined broad carbapenem resistance using computationally intensive models, this work targets ertapenem specifically using lightweight classifiers deployable on standard computers. A web application prototype was also developed, with the idea of enabling clinicians to upload patient MALDI-TOF MS spectral data and receive real-time resistance predictions, reducing diagnostic turnaround from days to seconds and supporting faster, more targeted treatment decisions.

Competition history

• CSEF 2026 Microbiology (Senior Division) · Entry S-16-08

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Microbiology projects
• CSEF Microbiology projects
• Every CSEF 2026 project