Machine learning-based inference of carbapenemase phenotypic traits in Klebsiella pneumoniae using routine antimicrobial susceptibility profiles: a dual-center study

Rapid differentiation of serine carbapenemase- and metallo-β-lactamase (MBL)-associated phenotypes in carbapenem-resistant Klebsiella pneumoniae (CRKP) is important for early clinical risk stratification. This study aimed to develop and externally evaluate machine-learning models using routinely generated antimicrobial susceptibility testing (AST) data to infer phenotypically defined carbapenemase groups and explore the limitations of AST-based classification in dual-positive isolates. This retrospective two-center study included 522 non-duplicate CRKP isolates. Isolates from the Sixth Affiliated Hospital of Sun Yat-sen University constituted the derivation cohort ( n = 277), while isolates from Zhujiang Hospital of Southern Medical University formed an independent external validation cohort ( n = 245). Carbapenemase phenotypes were defined using inhibitor-based enhancement testing with 3-aminophenylboronic acid and EDTA. Ten log₂-transformed minimum inhibitory concentration features were analyzed. Eight machine-learning algorithms were compared using stratified 10-fold cross-validation. Two parallel binary tasks were developed to identify serine-positive and MBL-positive phenotypes, together with an exploratory four-class model. In internal task-specific stratified 10-fold cross-validation, random forest (ranger) achieved an AUROC of 0.897 and AUPRC of 0.965 for the Serine task and an AUROC of 0.840 and the highest AUPRC of 0.676 for the MBL task. In the independent SMU-ZJ cohort, ranger achieved a Serine AUROC of 0.760 (95% CI, 0.664–0.860) and AUPRC of 0.951 (0.931–0.972), and an MBL AUROC of 0.734 (0.634–0.831) and AUPRC of 0.420 (0.263–0.629). The external cohort contained only 27 MBL-positive isolates. The exploratory four-class model achieved 87.8% overall accuracy but correctly identified only 2 of 24 MBL-only isolates and none of the three dual-positive isolates. Machine-learning models based on routine AST profiles demonstrated moderate cross-center discrimination for phenotypically defined carbapenemase groups and may provide a preliminary screening signal without requiring additional input variables beyond the completed AST panel. However, low MBL sensitivity at the fixed 0.50 cutoff, reduced precision-recall performance in external validation, and poor recognition of dual-positive isolates indicate that these models should complement, rather than replace, confirmatory phenotypic or molecular testing. Prospective multicenter studies incorporating genomic validation and prespecified clinical thresholds are warranted.