Genomic Context as a Predictor of Multidrug Resistance in African Klebsiella pneumoniae: A Feasibility Study with Leave-One-Country-Out Validation

A feasibility study has been conducted to determine if genomic context can predict multidrug resistance (MDR) in Klebsiella pneumoniae from African countries. The study utilized a leave-one-country-out (LOCO) cross-validation approach with an explainable machine-learning framework. However, a scarcity of phenotypic linkage data limited the research, as only 2.43% (231 out of 9,505) of the African NCBI records contained antibiogram information.

To overcome this limitation, a rule-based genotypic MDR proxy label was employed. The study found that LOCO cross-validation was feasible using data from 8 African countries (n = 200 genomes), but not applicable to any previously published cohort. In a stratified pilot study of 175 species-confirmed genomes from 9 countries, tree ensembles achieved an area under the receiver operating characteristic curve (AUROC) of 0.85 when data were split randomly.

However, LOCO cross-validation yielded AUROC values ranging from 0.66 to 0.69, indicating a geographic-generalization gap of approximately 0.15 AUROC. This suggests that pooled accuracy may overestimate the transportability of genomic context for predicting MDR in Klebsiella pneumoniae across different African countries, although the pilot fold sizes were relatively small (n = 200).