Prediction of antibiotic resistance from antibiotic susceptibility testing results from surveillance data using machine learning

Antimicrobial resistance is a growing global health threat, and artificial intelligence offers a promising avenue for developing advanced tools to address this challenge. In this study, we applied various machine learning techniques to predict bacterial antibiotic resistance using the Pfizer ATLAS Antibiotics dataset. This comprehensive dataset includes patient demographic data, sample collection details, antibiotic susceptibility test results, and resistance phenotypes for 917,049 bacterial isolates. The dataset was divided into two subsets: Phenotype-Only and Phenotype + Genotype, excluding and including 589,998 isolates with genotype data, respectively. Both subsets underwent exploratory data analysis, preprocessing, machine learning model training, validation, and optimization. XGBoost consistently outperformed other models, achieving AUC values of 0.96 and 0.95 for the Phenotype-Only and Phenotype + Genotype sets, respectively. Hyperparameter tuning yielded slight accuracy improvements, while data balancing techniques notably increased recall. Across all models, the antibiotic used emerged as the most influential feature in predicting resistance outcomes. The SHAP summary plots generated provide insights into model interpretability. Our findings provide valuable insights into global AMR patterns and demonstrate the potential of AI-driven approaches for resistance prediction to help inform clinical decision-making and support the formulation of effective AMR mitigation policies, subject to the availability of highly granular datasets.