A Multi-Model Comparison of Supervised Machine Learning for Predicting Antimicrobial Resistance in ESKAPE Pathogens ​

Antimicrobial resistance (AMR) causes about 1.3 million deaths annually, yet culture-based susceptibility testing delays treatment by 24–72 hours. Whole genome sequencing (WGS) enables rapid computational AMR prediction, but optimal machine learning approaches remain unclear. We trained and compared three algorithms (Random Forest - RF, Extreme Gradient Boosting - XGBoost, Multilayer Perception - MLP) on 12,668 ESKAPE pathogen genomes using ABRicate-derived AMR gene features. Paired Wilcoxon tests revealed dataset-size-dependent performance: PyTorch multilayer perception (MLP) significantly outperformed RF in large-dataset pathogens (Klebsiella pneumoniae Area Under the Receiver Operating Characteristic Curve (AUROC) 0.823 p=0.0056, Enterococcus faecium 0.782 p=0.0041, Pseudomonas aeruginosa 0.659 p=0.0027) relative to smaller datasets (Staphylococcus aureus p=0.8596, Acinetobacter baumannii p=0.1042). Random Forest excelled when data was limited (Enterobacter spp. n=135, p=0.0702). Feature importances validate established AMR biology. Tigecycline, meropenem, and colistin presented predictive challenges (AUROC <0.63) attributed to heterogenous AMR mechanisms and genetic variants limiting training diversity. The study showed that deep learning suits large surveillance networks (≥1,000 genomes) relative to RF, which suits resource-limited settings.