Data Challenge: Temporal Trends and Machine Learning Prediction of Phenotypic Carbapenem and Third-Generation Cephalosporin Resistance in Gram-Negative Bacteria Using Two Decades of ATLAS Global Surveillance Data

Abstract

Antimicrobial resistance (AMR) in Gram-negative bacteria continues to rise globally, undermining the effectiveness of key last-line antibiotics such as carbapenems and third-generation cephalosporins. While large-scale surveillance systems like ATLAS have generated extensive longitudinal data, most existing studies remain largely descriptive, focusing on reporting resistance prevalence rather than understanding how resistance evolves, interacts across antibiotics, and progresses into future risk patterns. A major gap in current AMR research is the limited ability to integrate multi-dimensional surveillance data into predictive and clinically actionable intelligence. Traditional analyses typically examine temporal trends in isolation, without jointly modeling pathogen characteristics, geographic variation, specimen sources, MIC distributions, and co-resistance structures. As a result, current evidence is often retrospective and insufficient for anticipating emerging resistance threats. This study addresses this gap by leveraging the ATLAS dataset to develop a predictive and analytical framework for phenotypic resistance in Gram-negative bacteria. The focus is on carbapenem and third-generation cephalosporin resistance across pathogens, regions, and clinical contexts, with an emphasis on capturing both temporal dynamics and underlying resistance relationships. Methodologically, the study moves beyond classical trend analysis by integrating machine learning models capable of learning non-linear interactions across multiple features, including time, geography, pathogen type, antibiotic class, MIC distributions, and co-resistance patterns. These models will be used not only for prediction but also for uncovering hidden resistance structures and improving generalizability across regions and time periods. Model evaluation will follow time-aware validation strategies, and interpretability techniques will be applied to extract clinically meaningful patterns from complex model outputs. The expected outcome is a robust predictive framework that identifies emerging resistance risks earlier than conventional surveillance approaches, while also revealing interaction patterns that inform empirical treatment strategies and antimicrobial stewardship.