Leveraging artificial intelligence for clinical decision support in resource-constrained settings

<p dir="ltr">Artificial intelligence (AI), especially machine learning (ML) algorithms, promise to transform clinical decision-making, improving the safety and quality of care delivery and patient outcomes. AI is increasingly being embedded into contemporary clinical decision support (CDS) systems, but few have been implemented and evaluated in routine care, particularly in resource-constrained clinical settings that are characterised by limited access to medical expertise and equipment. </p><p dir="ltr">This thesis examines the different facets of leveraging AI-based CDS to improve care delivery in resource-constrained clinical settings. The research consisted of three studies utilising a mixed-methods, exploratory sequential design to examine the application of AI for improving the management of cardiovascular disease, a leading global health issue due to its high mortality and morbidity. Indonesia was selected as a focus setting, providing insights into various conditions associated with limited resources. </p><p dir="ltr">The first study examined contemporary AI-based CDS and their effects on decisionmaking, care delivery, and patient outcomes. A scoping review identified 32 studies evaluating various types of AI-based CDS in healthcare settings. All were undertaken in developed countries and largely in secondary and tertiary care settings (91%). The review confirmed a gap in evaluating AI-based CDS for resource-constrained settings. The most common clinical tasks supported by AI were image recognition and interpretation (38%) and risk assessment (28%). Most systems were assistive (72%), requiring clinicians to confirm or approve CDS recommendations. These findings informed the design and hypothesis of the interviews and experiment undertaken as subsequent studies. </p><p dir="ltr">The second study explored the sociotechnical context of leveraging AI-based CDS to support cardiovascular disease management using semi-structured interviews with Indonesian doctors working in resource-constrained settings (n=27). Doctors reported challenges in dealing with a high patient volume and clinical decisions were largely based on experience, as there was limited access to confirmatory examinations. In these settings, doctors generally used CDS on mobile devices and indicated a preference for data entry to be automated. Participants highlighted a critical need for CDS to support the assessment of atherosclerotic cardiovascular disease (ASCVD) risk to help prioritise and facilitate better clinical management of high-risk patients. These findings informed the design of the third study to assess the effect of AI-based CDS on ASCVD risk assessment and management. </p><p dir="ltr">The third and final study was a within-subject randomised controlled experiment that used simulated patient cases to assess the potential effects of AI-based CDS on 10- year ASCVD risk assessment and management. One hundred and two doctors were recruited and asked to complete 9 patient cases online, with and without AI assistance that was available via an emulated mobile device. The results showed that AI-based CDS significantly improved risk assessment (+27%, <i>p</i><0.001) and prescription of statins (+29%, <i>p</i><0.001). Cases assisted by AI-based CDS took less time than the control (<i>p</i>=0.017). Doctors generally had positive perceptions about the use of AIbased CDS. </p><p dir="ltr">This thesis contributes new knowledge about the potential utility of AI-based CDS in improving the assessment and management of cardiovascular disease risk in resource-constrained clinical settings. It demonstrates a problem-driven approach to co-design and evaluate AI with consideration for the distinct sociotechnical context of resource-constrained settings. The experiment showed a plausible AI intervention, providing access to a better risk calculator on a mobile device with automated data acquisition. As such, this has the potential to enhance primary prevention by improving decision-making and care delivery. Further research is needed to ascertain if improvements observed in the clinical simulation translate to real-world clinical settings. </p><p dir="ltr">A 60-second video as part of Visualise Your Thesis^TM is available in Supplement A.7</p>