Artificial Intelligence-Driven Clinical Decision Support for Antibiotic Optimisation
File(s)
Author(s)
Bolton, William
Type
Thesis
Abstract
Artificial intelligence (AI) is poised to revolutionise healthcare in the coming decades. AI-based clinical decision support systems (CDSS) that aim to help clinicians make complex decisions and deliver precision medicine have been an area of significant research. Such technology has been developed for infectious diseases to tackle the growing global challenge of antimicrobial resistance (AMR). To curtail the development of AMR we must optimize antimicrobial use and prolong the therapeutic life of current drugs, which is called antimicrobial stewardship. In this thesis, AI-driven CDSSs that use routinely collected electronic health record (EHR) data to optimise antibiotic use are researched, developed, and evaluated.
This thesis employed multiple research methodologies to investigate key aspects of decision-making that clinicians currently find challenging when prescribing antimicrobials. Specifically, work focused on intravenous-to-oral switching (IVOS) and stopping treatment, aspects of antimicrobial stewardship that have previously been neglected in the literature. Numerous machine learning and deep learning methods were applied with a web-based app, semi-structured interviews, and surveys, used by clinicians to evaluate the IVOS-focused CDSS and determine its acceptability and impact on decision-making.
AI models were able to obtain good performance in stewardship-focused prediction tasks. IVOS models obtained an AUROC of up to 0.80 while not being biased to individuals’ protected characteristics such as sex. Algorithms that investigated stopping treatment accomplished an AUROC of 0.77 for mortality prediction. 42 individuals completed the clinician evaluation study. Often the CDSS did not influence decision-making, with participants able to identify and ignore incorrect advice. Some statistical differences were observed with the AI CDSS influencing clinicians towards not switching the route of antibiotic administration, which aligns with the cautious behaviour reported by 40% of participants.
AI-driven CDSSs hold promise to improve and optimise the complex decisions made by clinicians when prescribing antimicrobials, to benefit healthcare systems, patients, and tackle AMR.
This thesis employed multiple research methodologies to investigate key aspects of decision-making that clinicians currently find challenging when prescribing antimicrobials. Specifically, work focused on intravenous-to-oral switching (IVOS) and stopping treatment, aspects of antimicrobial stewardship that have previously been neglected in the literature. Numerous machine learning and deep learning methods were applied with a web-based app, semi-structured interviews, and surveys, used by clinicians to evaluate the IVOS-focused CDSS and determine its acceptability and impact on decision-making.
AI models were able to obtain good performance in stewardship-focused prediction tasks. IVOS models obtained an AUROC of up to 0.80 while not being biased to individuals’ protected characteristics such as sex. Algorithms that investigated stopping treatment accomplished an AUROC of 0.77 for mortality prediction. 42 individuals completed the clinician evaluation study. Often the CDSS did not influence decision-making, with participants able to identify and ignore incorrect advice. Some statistical differences were observed with the AI CDSS influencing clinicians towards not switching the route of antibiotic administration, which aligns with the cautious behaviour reported by 40% of participants.
AI-driven CDSSs hold promise to improve and optimise the complex decisions made by clinicians when prescribing antimicrobials, to benefit healthcare systems, patients, and tackle AMR.
Date Issued
2024-12-13
Date Awarded
2025-03-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Georgiou, Pantelis
Holmes, Alison
Rawson, Timothy
Sponsor
UKRI
Grant Number
Grant No. P/S023283/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)
Rights Embargo Date
2026-02-28