Early antimicrobial resistance prediction - Using urinary proteomics and machine learning

Abstract

Antimicrobial resistance (AMR) is a growing global health challenge, requiring improved methods for rapid and accurate resistance prediction. This study investigates whether urinary proteomics data can be used to predict AMR phenotypes using machine-learning approaches. Mass spectrometry-based proteomics data from urine samples associated with E.coli urinary tract infections were preprocessed, and analysed using multiple machinelearning frameworks. Both multi-label and single-label classification approaches were evaluated using stratified cross-validation, followed by feature selection and stability-based analysis to reduce dimensionality while improving interpretability. In addition, antibiotic-specific models were developed to account for variation in resistance prevalence and class imbalance across antibiotics. The results show that proteomics-derived protein abundance profiles contain predictive information related to AMR phenotypes, although performance varied across antibiotics and modelling strategies. Single-label classification generally outperformed the multi-label framework in terms of stability and interpretability. Feature selection substantially reduced the dimensionality of the dataset while maintaining comparable predictive performance, suggesting that a limited subset of proteins captures a significant proportion of the predictive signal. However, class imbalance and small sample size limited the ability to reliably predict minority resistance phenotypes. The findings demonstrate the potential of urinary proteomics combined with machine learning for AMR phenotype prediction, while highlighting key methodological challenges such as class imbalance, high dimensionality, and limited generalisability. These results support further development of proteomics-based predictive models and validation on larger, independent clinical cohorts.