EXPLAINABLE ARTIFICIAL INTELLIGENCE BASED BIOMARKER PRIORITIZATION FOR ANTIMICROBIAL RESISTANCE PREDICTION IN CLINICALLY SIGNIFICANT BACTERIAL PATHOGENS

Abstract

Multidrug-resistant bacterial infections are becoming a big health issue worldwide, owing to the emergence of antimicrobial resistance (AMR) which is impairing the efficacy of currently available antibiotics. The pathogens most commonly found in the clinical setting, Klebsiella pneumoniae (KP) and Pseudomonas aeruginosa (PA), are often linked to high rates of antimicrobial resistance and to serious hospital-acquired infections. In the present study, a human readable machine learning model was built for predicting resistance to KP and PA as well as predicting biomarkers in the meropenem resistance and ciprofloxacin resistance datasets, respectively. The performances of the supervised machine learning algorithms were compared using the accuracy, precision, recall, F1 score and ROC(Receiver Operating Characteristic) AUC(Area Under the Curve). Exemplarily, ensemble-based models such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost) had the best predictive power on both datasets. Comparative analysis showed that KP has higher predictive stability than PA, suggesting differences in the architecture of resistance among the organisms. SHapley Additive exPlanations (SHAP) analysis was used to rank the features and prioritize the biomarkers for their biological interpretability. Well characterized genes for antimicrobial resistance (e.g. OXA-type carbapenemase, KPC-family genes, NDM genes, and aminoglykoside resistance determinants) were the most important biomarkers in KP, suggesting the gene-driven resistance mechanism. By contrast, PA resistance prediction was linked to proteins involved in regulatory functions, related to efflux, membrane associated proteins, transcription regulation and adaptive stress response elements, indicating a more diverse and system level resistance profile. These patterns of resistance were further confirmed by visualization of biomarkers using heatmap and cross-validated ROC AUC analysis, performed within the two organisms. KP had largely clustered AMR signatures and good classification performance, while PA had a more variable, but biologically relevant, predictive performance. This study shows the importance of explainable AI methods for AMR prediction and discovery of interpretable biomarkers, and highlights the fundamental differences in resistance mechanisms between KP and PA.