# Machine learning for personalized antimicrobial susceptibility breakpoints: (Adaptive clinical breakpoint interpretation)

## TL;DR

This study explores using machine learning to predict UTI diagnoses and personalize aminopenicillin dosing recommendations based on EUCAST guidelines.

## Contribution

The novel use of machine learning to adaptively interpret antimicrobial susceptibility breakpoints for personalized treatment.

## Key findings

- XGBoost models achieved an AUC of 0.62 for predicting complicated UTI and UTI in patients with Enterobacterales infections.
- Adjusting probability thresholds improved appropriate aminopenicillin regimen recommendations to 96.6% in bacteriuria cases.
- Machine learning shows potential for personalized EUCAST breakpoint implementation in antimicrobial susceptibility testing.

## Abstract

Infection diagnoses are critical to the personalized interpretation of EUCAST aminopenicillin breakpoints for Enterobacterales, but microbiology laboratories cannot predict diagnosis when specimens are received. Here, we assess whether machine learning could facilitate personalized antimicrobial susceptibility breakpoint reporting by predicting urinary tract infection (UTI) diagnoses.

XGBoost models were trained using open-source electronic healthcare record data to predict complicated UTI in patients with Enterobacterales bacteriuria and to predict UTI in patients with Enterobacterales bacteraemia. These models were validated and used to provide simulated aminopenicillin dosing/regimen recommendations based on antimicrobial susceptibility results for patients with bacteriuria and bacteraemia in a holdout dataset. The main outcomes were the proportions of patients recommended appropriate aminopenicillin dosages/regimens according to EUCAST guidelines based on their diagnosis.

The area under the receiver operating characteristic curve was 0.62 for predicting both complicated UTI in patients with bacteriuria and UTI in patients with bacteraemia. In the simulation study, 79.3% (n = 276) and 72.7% (n = 8) of patients with ampicillin-susceptible Enterobacterales bacteriuria and bacteraemia, respectively, were recommended appropriate aminopenicillin dosages/regimens for their infection diagnosis according to EUCAST guidelines. Adjusting the probability threshold for predicting complicated UTI increased the proportion of appropriate recommendations in bacteriuria to 96.6% (n = 336).

Using machine learning models to predict the probability of complicated UTI in patients with bacteriuria and the probability of UTI in patients with bacteraemia resulted in appropriate aminopenicillin dosages/regimens being recommended in most cases. These results provide proof-of-concept for how machine learning could facilitate the personalized implementation of EUCAST aminopenicillin breakpoints.