One of the most critical clinical scenarios impacted by antimicrobial resistance (AMR) is the management of sepsis. Mortality increases by up to 8% for every hour delay until effective antimicrobial initiation, so timely effective empirical antibiotic treatment is paramount. In many LMICs, limited diagnostic capacity prevents timely microbial identification and antimicrobial susceptibility testing (AST). Even in high-income countries, AST results typically take 24–72 hours, necessitating empirical therapy in the interim period.
When empirical first-line therapy fails, clinicians must escalate therapy swiftly. However, traditional antibiograms, which guide empirical antimicrobial selection, are based on population-level susceptibility estimates. They rarely account for patient-specific factors and offer no guidance on selecting second-line agents following presumed first-line resistance. This lack of escalation guidance and personalised strategies highlights a critical gap in clinical decision-making, contributing to both delays in effective treatment and unnecessary use of broad-spectrum antibiotics.
We aim to estimate probabilities of resistance to commonly used antibiotics among bloodstream infections, conditional on resistance to first-line agents and patient characteristics. The goal is to support the development of a clinical decision-support tool for more effective and timely antibiotic escalation.
We will make use of the Antimicrobial Testing Leadership and Surveillance (ATLAS) dataset, which contains AST results from blood isolates across diverse geographies and timeframes. We will apply a mixture of Bayesian modelling and machine learning to capture complex relationships between patient-level features and AMR, with an aim to generalise and transfer learning to specific LMIC settings where data availability is sparse.
This study will provide evidence required to provide personalised guidance on empirical antibiotic selection and escalation when comprehensive patient-level AST data is pending or lacking, thereby reducing time to effective antibiotic therapy while avoiding overreliance on broad-spectrum antibiotics.