Antibiotic stewardship programs (ASP) aim to reduce inappropriate use of antibiotics, but their labor-intensive nature impedes their wide adoption. We developed explainable machine learning (ML) models that can prioritize inpatients who would benefit most from stewardship interventions.
We constructed a cohort of inpatients who received systemic antibiotics and were monitored by a multidisciplinary ASP team at a hospital in the Republic of Korea. We used a dataset of over 130,000 patient-days and extracted more than 160 features from multiple domains, including prescription records, laboratory, microbiology results, and patient conditions. We generated the outcome labels using medication administration history: discontinuation, switching from intravenous to oral medication (IV to PO), and early or late de-escalation. We trained the models using Extreme Gradient Boosting (XGB) and light Gradient Boosting Machine (LGBM) and conducted SHapley Additive exPlanations (SHAP) analysis to explain the model’s predictions.
The models demonstrated strong discrimination when evaluated on a hold-out test set (AUROC - IV to PO: 0.81, Early de-escalation: 0.78, Late de-escalation: 0.72, Discontinue: 0.80). The models identified 41%, 16%, 22%, and 17% more cases requiring discontinuation, IV to PO, early and late de-escalation, respectively, compared to the conventional length of therapy strategy, given that the same number of patients were reviewed by the ASP team. The SHAP results explains how each model makes their predictions, highlighting a unique set of important features that are well-aligned with the clinical intuitions of the ASP team.
The models are expected improve the efficiency of ASP activities by prioritizing cases that would benefit from different types of ASP interventions along with detailed explanations.