Related Experiment Video
Updated: Jul 10, 2026
Selecting Multiple Biomarker Subsets with Similarly Effective Binary Classification Performances
Published on: October 11, 2018
stackPredAMR-a stacked random forest approach improves AMR phenotype prediction for multiple species and
Julian Welling1,2, Miriam Balzer1,2, Leah Consten1
1Department of Medicine, University of Duisburg-Essen, Essen, 45147, Germany.
Motivation:
Antimicrobial resistance is a growing global threat, creating a need for rapid and accurate antimicrobial susceptibility testing. Current phenotypic antimicrobial susceptibility testing methods rely on prior isolation and cultivation, making them time-consuming. Whole genome sequencing combined with machine learning offers a faster and cost-effective alternative, but existing approaches are often limited in species coverage, antimicrobial scope, or data availability.
Results:
We developed stackPredAMR, a machine learning framework for predicting resistance to 18 antimicrobial agents in three clinically important bacterial species: Escherichia coli, Klebsiella pneumoniae, and Acinetobacter baumannii. The model uses antimicrobial resistance gene presence as input and incorporates cross-resistance patterns through a stacked architecture with two random forest layers. Benchmarking on more than 2500 publicly available whole genome sequencing datasets with linked phenotypic resistance data showed strong performance, achieving a median accuracy of 0.94, ROC AUC of 0.97, and F1-score of 0.91, outperforming previously published methods. stackPredAMR is freely available and designed to support future extension to additional species and antimicrobial agents.
Availability And Implementation:
Source code and datasets (database-driven reference approach, sample lists, and input features) are available at WIN-KID repository (https://github.com/IKIM-Essen/WIN-KID/tree/v1.0.0.0) and the release page (https://github.com/IKIM-Essen/WIN-KID/releases/tag/v1.0.0.0).
