Statistical Modelling in Diabetic Retinopathy Research

: An in-depth aetiological analysis and predictive modelling study to enhance diabetic retinopathy detection and reduce global health inequalities in people with type 2 diabetes

Abstract

Public healthcare systems need innovative, cost-effective, and accessible approaches to identify people with diabetic retinopathy (DR) because the cost and training to deliver current retinal screening services to detect DR are major barriers especially in low-and-middle-income-countries (LMICs) or resource restricted settings. A risk-based approach could be a viable alternative to universal screening, however, the lack of convincing evidence on the ability of models to guide risk stratification in sight threatening diabetic retinopathy (STDR) poses a challenge for the implementation of risk-based screening strategies in LMICs. Research projects that span across developed and developing countries provide opportunities for global translation. The ORNATEIndia project was designed to develop and test diverse translatable health strategies in India and the United Kingdom to tackle the global burden of diabetes-related vision impairment (VI) and reduce health inequalities among people with diabetes (PwD). This thesis comprises of eight publications undertaken as part of the ORNATE-India project. The objectives of my thesis were to critically evaluate the statistical methods in each publication that ranged from traditional statistics to cutting-edge machine learning, with distinct aims of addressing aetiological research as well as risk prediction for DR. All statistical models employed likelihood concepts from a frequentist inferential framework. Chapter 1 reviewed the global prevalence of DR (2015-2019), critiquing screening methods, the lack of studies in LMICs and the global differences in DR prevalence. Chapter 2 (with 2 publications) explored the burden of VI and blindness in PwD. Publication 2 estimated the national prevalence of VI and blindness in PwD in India using survey weighted methods, showing a higher prevalence in the lower socioeconomic strata. Publication 3 considered the age-standardised incidence of VI over 10 years in proliferative DR (PDR) patients undergoing Panretinal Photocoagulation (PRP) based on direct standardization methods, highlighting the need for prompt diagnosis and treatment of PDR. Chapter 3 explored the UK ethnic disparities in DR and STDR incidence using Cox proportional hazards models. Higher risk of DR and STDR in Black and South Asians were observed compared to their White counterparts. Moreover, risk factors of kidney function decline in PwD were found to be similar to those for STDR, substantiating the need for holistic approaches to prevention of microvascular complications. Chapter 4 investigated diagnostic biomarkers of DR and STDR using various statistical weighting procedures. In publication 5, an environmental wide association study (EWAS) conducted on the National Health and Nutrition Examination Survey (NHANES) datasets, highlighted hyperglycemia as a key factor in DR and STDR. While publication 6, collected primary data from UK and Indian participants to assess 13 blood biomarkers for STDR screening. Cystatin-C, not collected in publication 5, emerged as a top biomarker in comparison to those investigated, emphasising the association between renal disease and STDR. While diagnostic tools can be used to identify existing STDR, prognostic models are crucial for prevention, improving long-term health outcomes and reducing treatment costs. Chapter 5 presents a universally applicable STDR risk tool, with model coefficients robustly verified using cox models and under the assumption of interval-censoring, demonstrating strong performance in internal (c-statistics ranging 0.778-0.832) and external validation (c-statistics ranging 0.685-0.823). The tool overcame a major barrier for implementation as it required no blood tests or technical examinations. Chapter 6 demonstrates the development and validation of resource-driven chronic kidney disease (CKD) risk models for PwD, using fractional polynomials to model non-linear relationships and novel decision curve analysis to assess utility. These models stratify individuals based on the availability of tests, with the least invasive model eliminating the need for blood tests or technical examinations other than kidney markers eGFR and albumin to creatinine ratio (ACR). Finally, chapter 7 summarises the entire doctoral work, evaluating the ORNATE-India project’s impact and my contributions that led to its success.

Date of Award	Apr 2024
Original language	English
Awarding Institution	Coventry University
Sponsors	NIHR Moorfields Biomedical Research Centre
Supervisor	Robyn Tapp (Supervisor), A. Toby Prevost (Supervisor) & Sobha Sivaprasad (Supervisor)