Answers are grounded in this paper and cite their sources.
Purpose Antimicrobial resistance (AMR) poses a major threat to global public health, with increased morbidity, mortality, costs of health care and treatment failure. Design/methodology/approach This study explored nine supervised ML classifiers to describe pathogen-group trends in modeled AMR burden estimates among children under five years of age in the WHO African Region. A dataset of 768,576 records was sequentially filtered, optimized and evaluated using different performance metrics. SHapley Additive exPlanations (SHAP) and antibiotic-class sensitivity analysis were applied to explore model attribution and structural dependence. Findings The pathogen group had the strongest association with antibiotic class (Cramér's V = 0.5033), and a moderate association with infectious syndrome (Cramér's V = 0.3302). The Multi-Layer Perceptron (MLP) model had the highest F1 score (0.838; 95% confidence intervals: 0.834–0.841) and ROC-AUC of 0.931, followed by Gradient Boosting, Decision Tree, Random Forest, Extra Trees and KNN, with closely comparable F1 scores (around 0.835–0.836). Five-fold cross-validation showed stable internal performance, with no significant Holm-adjusted pairwise differences between MLP and the other top classifiers. Explainability analysis indicated that tuberculosis and several antibiotic classes are the major contributors to the model predictions. Removal of antibiotic class decreased F1-score by approximately 0.167–0.192 across all models, indicating strong structural reliance on this predictor. Originality/value This study provides an application-oriented, robust explainable ML benchmark with leakage control and effective model interpretation, which can assist in improving AMR surveillance, AMR stewardship and antibiotic-informed clinical decision support.
No takes yet. Share an insight, caveat, or question.
Usanase et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: