This project was presented at IDWeek, a leading infectious diseases conference in the U.S. You can read the abstract here.
The full paper is currently under review in a peer-reviewed journal.
Delayed and inappropriate antibiotic use in emergency departments contributes to the spread of resistant bacteria and worse outcomes. Extended-Spectrum β-Lactamase-producing Enterobacterales (ESBL-E) infections are a growing threat. Rapid, accurate risk stratification is essential to ensure early carbapenem treatment when appropriate — without overusing these powerful drugs and fueling resistance.
We built a machine learning model to predict ESBL-E infections at the time of antibiotic ordering in the Emergency Department using real-world EHR data. The model was trained on over 129,000 adult ED encounters at Johns Hopkins (2019–2024). Features included:
The algorithm — a gradient boosting model — was trained on data before July 2023 and validated on the remainder.
This model shows strong potential to support real-time antibiotic decisions in the ED using existing EHR infrastructure. With further refinement, it can guide early, targeted interventions — improving outcomes and preserving the effectiveness of last-line antibiotics.