Research Question: Can statistically significant upward shifts in the upper percentiles (75th/90th) of MIC distributions for Klebsiella pneumoniae and Escherichia coli against carbapenems forecast breakpoint-defined resistance 6–12 months in advance at the national level?
Clinical management of infections caused by antimicrobial-resistant (AMR) pathogens often relies on identifying resistance after therapeutic failure. Resistance is typically defined only when minimum inhibitory concentration (MIC) values cross fixed breakpoints. However, this binary classification may miss early warning signals of emerging resistance, resulting in delayed or suboptimal treatment and increased patient morbidity.
We propose an unsupervised machine learning approach to detect statistically significant shifts in the upper percentiles (75th and 90th) of MIC distributions for Klebsiella pneumoniae and Escherichia coli to carbapenems (imipenem and meropenem). Drawing inspiration from Sakagianni et al., we will apply change-point detection techniques to these percentile trends to identify early shifts indicative of brewing resistance. Our central hypothesis is that these subtle distributional changes precede clinically defined resistance events, offering an earlier alert signal.
We will use the ATLAS_Antibiotics dataset to extract longitudinal MIC data, focusing geographically on South Africa and Israel. These countries represent distinct AMR trajectories and laboratory capacities, enabling comparative insight into model generalizability.
Our key outcomes will include:
- Lead time between MIC shifts and confirmed resistance emergence
- Performance metrics of the change-point detection algorithm (sensitivity, specificity, false positives)
- A reproducible, open-source pipeline that can be generalized to other pathogens or drug classes
By transforming granular MIC trends into actionable early-warning tools, this project could strengthen AMR surveillance and empower clinicians to personalize treatment decisions, shifting AMR response toward prevention rather than reaction.