Predicting antimicrobial resistance using historical bacterial resistance data with machine learning algorithms

Abstract

Abstract Antibiotic resistance of bacterial pathogens is considered by the World Health Organization as a major threat to global health aggravated by the misuse of antibiotics. In clinical practice results of bacterial cultures and antibiograms can take several days. In the meantime, prescribing an empirical antimicrobial treatment constitutes a challenge in which the practitioner has to strike a balance between antibiotics spectrum and expected susceptibility probability. In this contribution, we report the development and testing of a machine-learning-based system that early predicts the antimicrobial susceptibility probability and provides explanations of the contribution of the different cofactors at 4 different stages prior to the antibiogram (sampling, direct examination, culture, and species identification stages). A comparative analysis of different state of the art machine learning and probabilistic methods was carried out using 7 years of historical bacterial resistance data from the Hôpital Européen Marseille, France. Our results suggest that dense neural network-based models and Bayesian models are suitable to early predict antibiotics susceptibility (average AUC 0.91 at the species identification stage) even for the less frequent situations.