Machine learning for estimating and comparing clinical rules for treating diarrheal illness with antibiotics

Area of Science:

• Pediatrics
• Infectious Diseases
• Biostatistics

Background:

• Acute diarrheal disease is a major cause of mortality in children under five, particularly in low-resource areas.
• Bacterial infections causing diarrhea may be treatable with antibiotics, but widespread use risks antimicrobial resistance.
• Individualized treatment guidelines are needed to balance treatment benefits against resistance risks.

Purpose of the Study:

• To develop and evaluate a framework for creating individualized treatment rules for childhood watery diarrhea.
• To incorporate diagnostic and clinical information for personalized antibiotic recommendations.
• To explicitly limit overtreatment and mitigate antimicrobial resistance emergence.

Main Methods:

• Utilized a framework for creating and evaluating individualized treatment rules.
• Employed ensemble machine learning and doubly-robust estimation via nested cross-validation.
• Proposed methods to compare rules based on different covariate sets, quantifying biomarker impact.

Main Results:

• Demonstrated appropriate inference in a realistic simulation study.
• Applied the method to real-world data from the AntiBiotics for Children with severe Diarrhea (ABCD) trial.
• Showcased the ability to quantify the impact of additional diagnostic biomarkers.

Conclusions:

• The proposed framework effectively derives and evaluates individualized treatment rules for childhood diarrhea.
• This approach aids in optimizing antibiotic use, reducing resistance risks.
• The methods allow for assessing the value of diagnostic biomarkers in clinical decision-making.