Data Challenge — ML-Based Clinical Decision Support for Empirical Antibiotic Selection in Klebsiella pneumoniae : A Multi-Regional ATLAS Analysis

Abstract

Antimicrobial resistance (AMR) in Klebsiella pneumoniae is a major global health concern, particularly in low- and middle-income countries (LMICs), where limited access to rapid diagnostics forces clinicians to rely on empirical antibiotic prescribing. This practice often leads to inappropriate use of broad-spectrum antibiotics, accelerating resistance and reducing future treatment options. This study proposes a clinically oriented artificial intelligence framework to support empirical antibiotic selection for K. pneumoniae infections, using the Pfizer ATLAS global antimicrobial surveillance dataset. Unlike conventional machine learning approaches focused solely on resistance prediction, our framework integrates a novel Clinical Decision Scoring Module (CDSM). This module converts predicted resistance probabilities into ranked, actionable antibiotic recommendations tailored to regional resistance patterns, thereby addressing real-world clinical decision-making needs. Three supervised learning models—XGBoost, Random Forest, and Logistic Regression—will be developed and rigorously evaluated. A temporal validation strategy will be applied, training models on data from 2004–2016 and testing on 2020–2022 data to simulate prospective deployment. Additionally, geographic cross-validation will assess model generalizability across LMIC regions, including Sub-Saharan Africa, South Asia, the Middle East and North Africa, and Latin America. Model interpretability will be ensured using SHAP analysis, providing transparent insights into prediction drivers and facilitating clinical trust. The proposed framework will be benchmarked against EUCAST empirical prescribing guidelines, using treatment coverage rate as the primary metric to evaluate clinical utility. This work aims to enhance antimicrobial stewardship, support evidence-based prescribing, and inform regional public health strategies. Importantly, it offers a scalable and practical decision-support solution that can be implemented in resource-limited healthcare settings without requiring advanced laboratory infrastructure or genomic data.