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

Multidrug-resistant (MDR) Klebsiella pneumoniae is a leading cause of healthcare-associated mortality in Africa, yet genomic prediction of resistance has relied almost exclusively on resistance-gene detection validated under random data splits. Whether genomic context—lineage, capsule and O-locus background, and virulence loci, with all resistance determinants excluded—can predict aggregate MDR status, and whether such signal survives geographic transport, remains untested. As a feasibility study, we built an explainable machine-learning framework with leave-one-country-out (LOCO) cross-validation. Phenotypic linkage proved extremely scarce: only 231 of 9,505 strict- K. pneumoniae African NCBI records (2.43%) carry submitter-supplied antibiograms, necessitating a rule-based genotypic MDR proxy label. Country-sufficiency analysis showed LOCO is feasible on the current snapshot—8 countries at n >= 200 genomes—but not on any previously published cohort. In a stratified pilot (175 species-confirmed genomes, 9 countries), tree ensembles reached pooled AUROC 0.85 under random splitting but only 0.66–0.69 under LOCO; this ∼0.15 AUROC geographic-generalization gap suggests that pooled accuracy overstates transportability, though at pilot fold sizes (n <= 20 test genomes) confidence intervals are wide and overlapping. SHAP attributions implicated the ybt virulence locus and O-serotype background, indicating models exploit lineage-associated population structure. The substantive contribution is a leakage-controlled, fully reproducible pipeline indicating that geographic validation, not pooled accuracy, is the operative test for genomic AMR surveillance models.