Machine learning for personalized antimicrobial susceptibility breakpoints.

Area of Science:

• Clinical microbiology and infectious diseases
• Machine learning applications in healthcare
• Pharmacodynamics and antimicrobial stewardship

Background:

• Accurate infection diagnosis is essential for interpreting European Committee on Antimicrobial Susceptibility Testing (EUCAST) breakpoints for aminopenicillins.
• Current laboratory methods cannot predict infection diagnoses upon specimen receipt, hindering personalized treatment.
• Enterobacterales are common causes of urinary tract infections (UTIs) and bacteraemia.

Purpose of the Study:

• To evaluate the utility of machine learning (ML) in predicting UTI diagnoses.
• To assess if ML-driven UTI prediction can facilitate personalized antimicrobial susceptibility breakpoint reporting.
• To enable tailored aminopenicillin dosing and regimen recommendations based on predicted diagnoses.

Main Methods:

• XGBoost ML models were developed using electronic healthcare record data.
• Models predicted complicated UTI in patients with Enterobacterales bacteriuria and UTI in patients with Enterobacterales bacteraemia.
• Model performance was validated, and simulated aminopenicillin recommendations were generated for a holdout dataset.

Main Results:

• The ML models achieved an area under the receiver operating characteristic curve of 0.62 for predicting both complicated UTI and UTI.
• In simulation, 79.3% of bacteriuria and 72.7% of bacteraemia cases received appropriate aminopenicillin recommendations.
• Adjusting the prediction threshold improved appropriate recommendations for bacteriuria to 96.6%.

Conclusions:

• ML models effectively predict UTI probabilities, leading to appropriate aminopenicillin dosage recommendations in most cases.
• This study demonstrates the potential of ML to personalize the application of EUCAST aminopenicillin breakpoints.
• ML facilitates personalized antimicrobial susceptibility reporting by predicting infection diagnoses.