Machine learning based prediction of antimicrobial resistance.: a five-year retrospective study

Abstract

Introduction species are well-recognized pathogens implicated in both healthcare-associated and community-acquired infections, and they contribute substantially to the global burden of antimicrobial resistance. In this study, we investigated the epidemiology, temporal resistance trends, multidrug resistance profiles, and the application of machine learning (ML) approaches to predict antimicrobial susceptibility among Klebsiella isolates in Al-Kharj, Saudi Arabia. Methods A retrospective analysis conducted using routine microbiology laboratory data collected between 2019 and 2024. Antimicrobial susceptibility testing included multiple agents across major antimicrobial classes, allowing classification of isolates into defined resistance phenotypes. Multidrug-resistant (MDR), extensively drug-resistant (XDR), and pan-drug-resistant (PDR) profiles were determined using standard class-based definitions. In parallel, several supervised machine learning models were developed and evaluated, including Random Forest, Support Vector Machine (SVM), Gradient Boosting, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Extra Trees Classifier, Classifier Chains (multilabel classification), Deep Neural Network (DNN), Voting Ensemble, and Convolutional Neural Network (CNN). Models were trained using the training dataset and subsequently evaluated on the testing dataset to assess predictive performance and generalizability. Results Over the five-year study period,species accounted for 2,646 isolates (13.7%) of all clinical isolates, withrepresenting the predominant species. Urinary tract infections were the most frequent source (45.5%), followed by blood cultures (15.4%), respiratory samples (14%), and wound and soft tissue specimens (12.2%). High resistance rates were observed for ceftazidime, whereas most other antibiotics demonstrated moderate resistance levels. Tigecycline, and Ceftriaxone showed the lowest resistance rates. Amongspp. isolates, 57.8% were classified as multidrug-resistant, 1.9% as extensively drug-resistant, and no pan-drug-resistant isolates were identified. Among the evaluated models, the best overall model was XGBoost with an accuracy of 0.70, precision of 0.65, recall of 0.64, an F1-score of 0.64, and a ROC-AUC of 0.74. Conclusions Collectively, these findings underscore the increasing clinical burden ofinfections and demonstrate the potential of boosting-based machine learning algorithms, particularly XGBoost and LightGBM, for accurate prediction of antibiotic resistance. Integration of these models into clinical decision-support systems and antimicrobial stewardship programs may facilitate timely and appropriate antimicrobial therapy, thereby improving patient outcomes and promoting more rational antibiotic use.