Machine learning-driven approaches to improve clinical diagnostics

: with Matrix-Assisted Laser Desorption/Ionization Time-of-Flight mass spectrometry in microbiology

Abstract

Current antibiotics are limited against superbugs (bacterial strains that have developed resistance to several types of antibiotics) due to the fast mutation rates of these bacteria. These superbugs are frequently drug resistant and can escape into patients’ bloodstreams to cause bacteraemia, making correct and timely prescriptions important. To tackle these concerns, one must first differentiate toxic (defined broadly here as having the toxin producing ability, i.e. toxigenic or to be toxic or having the properties of being toxic) bacteria from non-toxic ones as well as those resistant to certain antibiotics.
This thesis begins by providing the contextual background relating to bacterial infection, mass spectrometry sample preparation and analysis, as well as toxicity and antibiotic resistance. This thesis then provides a review of the principles of MALDI-TOF mass spectrometry in relation to bacterial diagnostics and its current usage within a clinical pathology diagnostic laboratory. The informatic aspects of MALDI-TOF are then reviewed focusing on existing pre-processing, biomarker detection, strain identification and antibiotic resistance approaches to inform clinical prescribing.
A pipeline of approaches for linking MALDI-TOF spectra to bacteria phenotype and genotypes is developed, as is a pipeline for pre-processing the spectra to enable subsequent modelling steps. This thesis presents new Bayesian and sparse Poisson models for diagnosing whether bacteria are toxic or resistant focusing on Staphylococcus aureus toxicity and Beta-lactam resistance in Escherichia coli. We find that when inputted into these models, mass spectrometry data can predict toxicity and resistance with AUC up to 98%, improving on published approaches. The Bayesian model also enables selection of important spectral regions determining toxicity in bacteria as well as statistical quantification of the degree of uncertainty in these regions. For interpretability, the sparse Poisson model provides quantification of the degree of similarity between library set of spectra and a set of unknown query spectra.
Specifically, a functional Bayesian regression was modified and applied to S. aureus toxicity prediction. Finally, a Poisson model based on sparse fitting of a library of previous acquisitions linked to antibiotic susceptibility testing data developed for E. coli resistance demonstrated high AUC up to 0.97 and an acceleration component was added that enabled it to perform 7 times faster than that without acceleration.
These findings provide methodologies and tools that have the potential to be translated to clinical practices and public health surveillance settings. However, future work is required to trial and validate these on larger clinical datasets.

Date of Award	17 Jun 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Matthew B Avison (Supervisor) & Andrew Dowsey (Supervisor)