PredKlebAMR: A machine learning framework prioritizing sensitivity in Klebsiella pneumoniae resistance profiling

Abstract

The global increase in carbapenem-resistant Klebsiella pneumoniae (CRKP) has undoubtedly changed the management of infectious diseases. Meropenem resistance in K. pneumoniae is a response to bacterial strategies, these includes production of carbapenemases (KPC, NDM, OXA-48), a reduction in membrane permeability through mutation of porin (OmpK35/36), and genomic plasticity via gene amplification and/or plasmid exchange. This study developed a machine learning model that targets “Very Major Errors” (VMEs) and predicts antimicrobial resistance (AMR) for prognosis. Genomic datasets (6000) were retrieved from the Bacterial and Viral Bioinformatics Resource Center (BV-BRC) database and preprocessed. Gene-based feature engineering was performed using Kleborate, and a Random Forest model was trained and deployed via an interactive R Shiny web application. Internal validation of the tool PredKlebAMR was done using a stratified 80/20 split of the training dataset, and a 5-fold cross-validation was implemented to optimize the Random Forest hyperparameters. While benchmarking against ResFinder 4.1 and AMRFinderPlus 4.2.7 was done using an independent external cohort consisting of three-tiered datasets from the National Center for Biotechnology Information (NCBI) database. PredKlebAMR , trained using 5,942 K. pneumoniae genomes (38 Kleborate-derived predictors), achieved a strong performance on a held-out global test set (accuracy 91.92%, sensitivity 94.67%, specificity 89.47%; Kappa 0.839), with key predictors including fluoroquinolone mutations, high-risk sequence types, KPC-2/KPC-3, and OmpK35/36 porin changes. When benchmarked against AMRFinderPlus 4.2.7 and ResFinder 4.1, PredKlebAMR delivered a higher sensitivity and fewer VMEs in the global cohort (95.45% vs 90.91% | 86.36% sensitivity; VMEs 1 vs 2 | 3), the African cohort (95.56% vs 91.11% | 82.2% sensitivity; VMEs 2 vs 4 | 8), and the Thailand cohort (95.62% vs 92.5% | 63.12% sensitivity; VMEs 7 vs 12 | 59) with statistically significant sensitivity improvement in Africa (McNemar’s p = 0.041). PredKlebAMR, with its high sensitivity to VMEs, and via an interactive dashboard ( https://igmr.org/software/predklebamr ) that translates predictions into biological rationale, will support clinicians in making life-changing decisions while managing patients.