Machine learning and synthetic outcome estimation for individualised antimicrobial cessation

Area of Science:

• Clinical Informatics
• Pharmacology
• Artificial Intelligence in Medicine

Background:

• Optimizing antimicrobial treatment duration is critical to prevent treatment failure, adverse events, and antimicrobial resistance (AMR).
• Current decision-making for antimicrobial cessation is complex and lacks robust research support.
• Both under- and over-treatment contribute to the growing challenge of AMR.

Purpose of the Study:

• To develop and validate a predictive model for estimating patient outcomes under different antimicrobial treatment durations.
• To assess the impact of earlier antimicrobial cessation on intensive care unit (ICU) length of stay (LOS) and mortality.
• To provide a foundation for a clinical decision support system for personalized antimicrobial stewardship.

Main Methods:

• Utilized electronic health record data from the MIMIC-IV database for 18,988 patients admitted to the ICU.
• Employed a recurrent neural network autoencoder combined with a synthetic control approach to model patient outcomes.
• Estimated ICU length of stay (LOS) and mortality under scenarios of stopping versus continuing antibiotic treatment.

Main Results:

• The developed model demonstrated reliable estimations of patient outcomes for both stopping and continuing antibiotic scenarios.
• Earlier cessation of antibiotic therapy was associated with a statistically significant reduction in ICU length of stay (mean reduction of 2.71 days).
• No significant impact on patient mortality was observed when antibiotic treatment was stopped earlier.

Conclusions:

• A novel model can reliably estimate patient outcomes under contrasting antimicrobial treatment durations.
• Retrospective analysis suggests that shorter antibiotic durations are often non-inferior to longer ones, aligning with existing clinical evidence.
• This approach, when developed into a clinical decision support tool, can aid in individualized antimicrobial cessation decisions, reduce antibiotic overuse, and mitigate AMR.