Retinal vascular measures from diabetes retinal screening photographs and risk of incident dementia in type 2 diabetes: A GoDARTS study

Emanuele Trucco, Frsa, Fiapr, Fbmva, Faaia, Alexander S. F. Doney, Aditya Nar, Yu Huang, Emanuele Trucco, Tom J. MacGillivray, Peter Connelly, Graham P. Leese, Gareth J. McKayFrontiers Digit. Health 2022 · DOI

TL;DR

This study explores whether retinal vascular measures, obtainable from standard diabetes retinal screening photographs, can predict the risk of developing dementia in individuals with type 2 diabetes. The findings suggest that these easily accessible retinal images hold significant potential as biomarkers for stratifying dementia risk, paving the way for improved prevention strategies.

What Problem Does This Paper Address?

Predicting dementia risk is a critical challenge, especially given the increasing prevalence of neurodegenerative diseases and the failure of many clinical trials to identify effective treatments early enough. Individuals with type 2 diabetes are known to have a higher risk of dementia, making them a key population for risk stratification. Current methods for dementia risk prediction often rely on clinical and genetic factors, but there's a need for more accessible and cost-effective biomarkers. The retina, being an embryological derivative of the brain, offers a unique opportunity to image brain health non-invasively. This paper investigates whether standard retinal photographs, already collected for diabetic retinopathy screening, can provide valuable insights into future dementia risk.

Methods

The researchers utilized data from the GoDARTS bioresource, a large prospective cohort of individuals with type 2 diabetes. They extracted retinal vascular measures (RVMs) from digital retinal screening photographs using the VAMPIRE software. These RVMs included:

  • Retinal vascular calibre: Central Retinal Arteriolar Equivalent (CRAE) and Central Retinal Venule Equivalent (CRVE), measuring the width of blood vessels.
  • Retinal vascular tortuosity: Arterial/venular tortuosity (TORTA/V), quantifying the windingness of blood vessels.
  • Retinal vascular fractal dimension: Arteriole/venule fractal dimension (FDA/V), measuring the complexity of the vascular branching pattern.

They then employed Cox's proportional hazards models and competing risk models to assess the association between these RVMs and the risk of incident all-cause dementia (ACD), Alzheimer's disease (AD), and vascular dementia (VD). The analysis accounted for competing risks of death and adjusted for clinical covariates (age, sex, diabetes duration, CVD history, BMI, blood pressure, cholesterol, HbA1c, smoking status) and ApoE4 genotype. To evaluate the predictive value of RVMs, they performed Receiver Operating Characteristic (ROC) analyses, calculating Area Under the Curve (AUC), Net Reclassification Improvement (NRI), and Integrated Discrimination Improvement (IDI).

Datasets

The study used data from the GoDARTS bioresource. This bioresource comprises approximately 8,500 patients with type 2 diabetes and 7,500 individuals without diabetes, all recruited from Tayside, Scotland. Comprehensive electronic medical records (EMRs) are available for these participants, with up to 20 years of follow-up data. For this study, 6,111 patients with type 2 diabetes had their retinal vascular measures successfully extracted and analyzed.

Key Insights and Ideas

  • The retina as a "window to the brain": This study reinforces the idea that easily accessible retinal images can provide valuable information about brain health and future disease risk.
  • Differential impact on dementia subtypes: The study highlights that different retinal vascular measures have distinct associations with Alzheimer's disease (AD) and vascular dementia (VD), suggesting unique underlying pathophysiological processes.
  • Fractal dimension as a key predictor: Retinal arteriolar fractal dimension (FDA) emerged as a significant predictor of increased risk for all-cause dementia (ACD) and AD. Conversely, venular fractal dimension (FDV) showed a complex relationship, associated with reduced AD risk but increased VD risk.
  • Vascular calibre matters: Wider arteriolar calibre (CRAE) was linked to a reduced risk of ACD, while wider venular calibre (CRVE) was associated with a reduced risk of AD.
  • Predictive power beyond traditional factors: RVMs significantly improved the accuracy of dementia risk prediction when added to models already including clinical and genetic factors (ApoE4 genotype).

Key Results

  • Increased arteriolar fractal dimension (FDA) was associated with a 17% increased risk of incident ACD (csHR 1.17; 1.08-1.26) and a 33% increased risk of AD (csHR 1.33; 1.16-1.52).
  • Increased venular fractal dimension (FDV) was associated with a 15% reduced risk of AD (csHR 0.85; 0.74-0.96) but a 22% increased risk of VD (csHR 1.22; 1.07-1.40).
  • Wider arteriolar calibre (CRAE) was associated with a 10% reduced risk of ACD (csHR 0.9; 0.83-0.98).
  • Wider venular calibre (CRVE) was associated with a 13% reduced risk of AD (csHR 0.87; 0.78-0.97).
  • The inclusion of RVMs led to a significant increase in AUC for ACD prediction (from 0.7855 to 0.7896, p=0.022). While increases were observed for AD and VD, they did not reach statistical significance due to wider confidence intervals.
  • RVMs significantly improved risk stratification, as indicated by positive Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) for all dementia types.

Limitations and Future Work

  • Healthy survivor bias: The authors acknowledge the potential for healthy survivor bias, where individuals with more severe underlying health issues might have already died before being included in the study or before RVM measurement. However, they mitigated this by using competing risk models.
  • Algorithm dependency: The study notes that different algorithms for measuring RVMs, particularly fractal dimension, can lead to variations in results across studies.
  • Diabetes-specific population: The findings are specific to individuals with type 2 diabetes, and their generalizability to other populations may be limited due to the unique microvascular complications associated with diabetes.
  • Clinical interpretability of deep learning: While deep learning approaches can enhance prediction accuracy, they often lack clinical interpretability. The authors suggest that identifying RVMs that are both predictive and clinically interpretable is crucial.
  • Future work: The authors propose that further development of disease-specific imaging biomarkers from retinal images could significantly enhance dementia risk stratification and facilitate the development of future prevention strategies. They also suggest exploring the interplay of RVMs with genetic factors and environmental exposures to gain a more comprehensive understanding of dementia risk.

Summary updated

BibTeX
@article{faaia2022retinal,
  author = {Emanuele Trucco, FRSA, FIAPR, FBMVA, FAAIA and Alexander S. F. Doney and Aditya Nar and Yu Huang and Emanuele Trucco and Tom J. MacGillivray and Peter Connelly and Graham P. Leese and Gareth J. McKay},
  title = {{Retinal vascular measures from diabetes retinal screening photographs and risk of incident dementia in type 2 diabetes: A GoDARTS study}},
  journal = {Frontiers Digit. Health},
  year = {2022},
  doi = {10.3389/fdgth.2022.945276},
}
Figure