Development of machine learning algorithms for scaling-up antibiotic stewardship

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems
• Machine Learning for Antibiotic Stewardship

Background:

• Antibiotic stewardship programs (ASP) are crucial for reducing inappropriate antibiotic use.
• The labor-intensive nature of ASPs limits their widespread adoption.
• Explainable machine learning (ML) models offer a solution to prioritize patients for interventions.

Purpose of the Study:

• To introduce explainable ML models for prioritizing inpatients for antibiotic stewardship interventions.
• To enhance the efficiency of ASP activities by identifying patients who would benefit most.

Main Methods:

• Trained Extreme Gradient Boosting (XGB) and light Gradient Boosting Machine (LGBM) models on a large cohort of inpatients (over 130,000 patient-days).
• Utilized over 160 features including prescription, laboratory, microbiology, and patient condition data.
• Employed SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• Models demonstrated strong predictive performance (AUROC for IV to PO: 0.81, Early de-escalation: 0.78, Late de-escalation: 0.72, Discontinue: 0.80).
• Identified significantly more cases for discontinuation (41%) and IV to PO switch (16%) compared to conventional strategies.
• SHAP analysis provided clinically relevant explanations for model predictions, aligning with ASP team expertise.

Conclusions:

• Explainable ML models can significantly improve ASP efficiency.
• These models prioritize patients for targeted interventions like discontinuation or de-escalation.
• The integration of ML offers a scalable approach to antibiotic stewardship.