Cardiovascular Disease Risk Assessment Using Deep Learning and Retinal Imaging in Primary Health Care Settings
Loading...
Citations
Altmetric:
Altmetric::
Download
Associated Records
Authors
File Licence
Editors
Date
2025
Source Title
Journal Title
Journal ISSN
Volume Title
Publisher
Melbourne Authors
Affiliation
Ophthalmology (Eye & Ear Hospital)
Document Type
PhD thesis
Subject Code
Citation
Access Status
This item is embargoed and will be available on 2027-09-15. This item is currently available to University of Melbourne staff and students only, login required.
DOI
Access URL
Open Access URL
Open Access at PMC
Linked Resource URL
NHMRC Grant Code
ARC Grant Code
Description
© 2025 Wenyi Hu
Abstract
Background
Effective cardiovascular disease (CVD) prevention depends on accurate risk stratification, yet most existing tools rely on laboratory tests and clinical data, limiting accessibility. Retinal photography offers a non-invasive “window” into vascular health, and recent deep learning (DL) algorithms can estimate CVD risk from fundus images. Although many such models exist, their real-world use and economic value remain largely unexplored.
Aim
This thesis aims to investigate the real-world needs, feasibility, performance, acceptability, and cost-effectiveness of DL-based retinal photographic CVD risk assessment in primary care.
Methods
A mixed-methods approach was adopted for three original studies.
1. Current practice audit: Medicare data for the Heart Health Check (HHC) were analysed by year and Statistical Area Level 4 (SA4). Linkage to Socio-Economic Indexes for Areas and the Australian Heart Map enabled assessment of demographic correlates using descriptive statistics, correlations, and multivariable linear regression.
2. Pragmatic trial: In two general practice clinics in Victoria, adults who had completed any part of a conventional CVD risk assessment within the previous six months underwent retinal photographic CVD (rpCVD) risk assessment. Agreement with the World Health Organisation (WHO) CVD risk chart (reference standard) was quantified with Pearson correlation, Bland-Altman analysis, and confusion matrices. Predictive accuracy of the rpCVD and WHO CVD scores was validated in the UK Biobank. Feasibility and acceptability were assessed using descriptive statistics and Likert scales.
3. Cost-effectiveness analysis: A decision tree/Markov model was used to compare retinal CVD risk assessment integrated into diabetic retinopathy (DR) screening programmes with the current practice (status quo) and DR screening alone in Australia, Singapore, and the United Kingdom (UK). Parameters were sourced from national datasets, published literature, and primary data. Outcomes included absolute and incremental cost-effectiveness ratios, net monetary benefit (NMB), and benefit-cost ratio (BCR). One-way and probabilistic sensitivity analyses were performed to evaluate uncertainties.
Results
The current practice audit suggested that the HHC remained underutilised, especially in rural and remote areas. Lower uptake was associated with lower socioeconomic status, higher heart failure hospitalisation rates, fewer older adults, and more people with limited English proficiency. These findings highlight the need for enhancing CVD risk assessment in primary care.
Through the pragmatic trial, DL-powered retinal CVD risk assessment was shown to be a feasible, accurate, and acceptable solution. The rpCVD system achieved a 93.9% gradable image rate. It was moderately correlated with the WHO CVD risk score (Pearson correlation coefficient = 0.526, 95% CI: 0.444–0.599) but demonstrated similar predictive performance for 10-year incident CVD (area under the receiver operating characteristic curve = 0.672, 95% CI: 0.658–0.686 vs 0.693, 95% CI: 0.680–0.707). Participant and general practitioner satisfaction rates were 92.5% and 87.5%, respectively.
The cost-effectiveness analysis demonstrated that integrating retinal CVD risk assessment into DR screening would be cost-effective. In Australia, combined screening could save up to A$546.25 million and gain up to 31,539.5 QALYs compared to DR screening alone. Compared with the status quo, the NMB would be up to A$3.77 billion, with a BCR of 9.89. Similar results were observed in Singapore and the UK, despite varying degrees of cost-effectiveness.
Conclusion
Comprehensive CVD risk assessment remains underutilised in Australia. DL-based retinal CVD risk assessment is quick, non-invasive, and achieves prognostic accuracy similar to that of established calculators. Coupling it with existing DR screening programmes is cost-effective in three diverse health systems, supporting its scalable adoption.
Notes
Terms of Use
Keywords
Cardiovascular disease, Risk assessment, Artificial intelligence, Deep learning, Retinal imaging