ARTIFICIAL INTELLIGENCE–DRIVEN PREDICTION OF ANTIMICROBIAL RESISTANCE USING INTEGRATED CLINICAL, WHOLE-GENOME SEQUENCING, AND METAGENOMIC DATA: A PROSPECTIVE MULTICENTER COHORT STUDY
ABSTRACT
Background: Antimicrobial resistance (AMR) is a major global health challenge that compromises effective treatment of infectious diseases and increases morbidity, mortality, and healthcare costs. Artificial intelligence (AI) has emerged as a promising approach for rapid prediction of resistance phenotypes using large-scale biological and clinical datasets. The objective of the study is to develop and validate a multimodal AI framework integrating clinical, whole-genome sequencing, and metagenomic data for accurate prediction of antimicrobial resistance in a prospective multicenter cohort.
Materials and Methods: This prospective multicenter study enrolled 2,492 patients with confirmed bacterial infections from five tertiary-care hospitals. Clinical characteristics, antimicrobial susceptibility testing results, whole-genome sequencing data, and metagenomic profiles were collected. Multiple machine-learning models, including Logistic Regression, Random Forest, XGBoost, and Deep Neural Networks (DNNs), were trained and evaluated. Model performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1-score. External validation was performed in an independent cohort of 742 patients.
Result: Among 2,492 patients, 1,014 (40.7%) exhibited antimicrobial-resistant infections. Prior antibiotic exposure (adjusted OR=3.12, 95% CI: 2.61–3.74), ICU admission (adjusted OR=2.46, 95% CI: 1.98–3.06), blaNDM positivity (adjusted OR=5.28, 95% CI: 4.02–6.93), and reduced microbiome diversity (adjusted OR=2.39, 95% CI: 1.91–2.99) were independent predictors of resistance. The multimodal DNN achieved the highest performance with an AUC of 0.957 (95% CI: 0.947–0.967), accuracy of 91.8%, sensitivity of 92.6%, specificity of 90.9%, and F1-score of 0.917. External validation confirmed excellent predictive performance (AUC=0.943; accuracy=90.4%).
Conclusion: Integration of clinical, genomic, and metagenomic data through a multimodal AI framework enables highly accurate prediction of antimicrobial resistance. Such models may facilitate earlier therapeutic decision-making and strengthen precision antimicrobial stewardship strategies.
Keywords: Antimicrobial resistance, Artificial intelligence, Machine learning, Whole-genome sequencing, Metagenomics, Deep learning, Predictive modelling, Precision medicine, Antimicrobial stewardship, Multi-omics.
Background: Antimicrobial resistance (AMR) is a major global health challenge that compromises effective treatment of infectious diseases and increases morbidity, mortality, and healthcare costs. Artificial intelligence (AI) has emerged as a promising approach for rapid prediction of resistance phenotypes using large-scale biological and clinical datasets. The objective of the study is to develop and validate a multimodal AI framework integrating clinical, whole-genome sequencing, and metagenomic data for accurate prediction of antimicrobial resistance in a prospective multicenter cohort.
Materials and Methods: This prospective multicenter study enrolled 2,492 patients with confirmed bacterial infections from five tertiary-care hospitals. Clinical characteristics, antimicrobial susceptibility testing results, whole-genome sequencing data, and metagenomic profiles were collected. Multiple machine-learning models, including Logistic Regression, Random Forest, XGBoost, and Deep Neural Networks (DNNs), were trained and evaluated. Model performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1-score. External validation was performed in an independent cohort of 742 patients.
Result: Among 2,492 patients, 1,014 (40.7%) exhibited antimicrobial-resistant infections. Prior antibiotic exposure (adjusted OR=3.12, 95% CI: 2.61–3.74), ICU admission (adjusted OR=2.46, 95% CI: 1.98–3.06), blaNDM positivity (adjusted OR=5.28, 95% CI: 4.02–6.93), and reduced microbiome diversity (adjusted OR=2.39, 95% CI: 1.91–2.99) were independent predictors of resistance. The multimodal DNN achieved the highest performance with an AUC of 0.957 (95% CI: 0.947–0.967), accuracy of 91.8%, sensitivity of 92.6%, specificity of 90.9%, and F1-score of 0.917. External validation confirmed excellent predictive performance (AUC=0.943; accuracy=90.4%).
Conclusion: Integration of clinical, genomic, and metagenomic data through a multimodal AI framework enables highly accurate prediction of antimicrobial resistance. Such models may facilitate earlier therapeutic decision-making and strengthen precision antimicrobial stewardship strategies.
Keywords: Antimicrobial resistance, Artificial intelligence, Machine learning, Whole-genome sequencing, Metagenomics, Deep learning, Predictive modelling, Precision medicine, Antimicrobial stewardship, Multi-omics.
View PDF Back