Personalized Antibiograms: Machine Learning for Precision Selection of Empiric Antibiotics

Area of Science:

• Medical Informatics
• Infectious Diseases
• Machine Learning

Background:

• Suboptimal antibiotic use is prevalent in hospitals, contributing to resistance and increased healthcare costs.
• Electronic health records (EHRs) contain vast amounts of data that can be leveraged to improve antibiotic stewardship.

Purpose of the Study:

• To develop and validate machine learning models for optimizing antibiotic selection in hospital settings.
• To identify opportunities for de-escalating antibiotic therapy based on predicted bacterial susceptibility.

Main Methods:

• Machine learning classifiers were trained on EHR data to identify negative microbial cultures (blood and urine) with high precision.
• Predictive models were developed to assess bacterial susceptibility to various antibiotic regimens.
• Decision thresholds were established to identify patient subgroups eligible for narrower-spectrum antibiotics.

Main Results:

• The models accurately flagged negative cultures, reducing unnecessary antibiotic prescriptions.
• Analysis revealed that 14% of Escherichia coli infections treated with piperacillin/tazobactam could have been treated with ceftriaxone.
• Additionally, 13% of these infections could have been treated with cefazolin, demonstrating potential for significant antibiotic de-escalation.

Conclusions:

• Machine learning applied to EHR data offers a powerful tool for optimizing antibiotic use and improving stewardship.
• These models can guide clinicians in selecting appropriate, often narrower-spectrum, antibiotics, thereby reducing wasteful prescribing.
• Implementing these predictive models has the potential to decrease antibiotic resistance and improve patient outcomes.