AI-Powered Stroke Risk Stratification

This section establishes the core scientific premise for utilizing retinal imaging in cerebrovascular risk assessment. It begins by delineating the profound biological connections that constitute the retina-brain axis, followed by a review of the historical evidence from traditional imaging modalities that has validated this concept and set the stage for more advanced technological approaches.

1.1 The Retina-Brain Axis: A Neurovascular Analogue

The retina is an embryological and anatomical extension of the central nervous system (CNS), making it a unique, non-invasive window through which cerebrovascular and neurodegenerative processes can be observed directly.^{[1, 2, 3]}

Embryological Origins

During embryonic development, the eye develops from the forebrain as an outgrowth of the diencephalon—the same neural structure that gives rise to critical brain components.^{[4, 5]} This shared embryological origin means the retina is composed of actual brain tissue, including CNS neurons and glial cells.^[1]

Anatomical Parallels

The structural similarities between the retinal and cerebral microvasculature provide a strong anatomical basis for the retina-as-a-proxy model.^{[7, 8, 9]} Both vascular systems are end-arterial and protected by highly selective barriers: the blood-retinal barrier (BRB) and the blood-brain barrier (BBB).^{[1, 4, 9]}

From a neural perspective, the retinal ganglion cells (RGCs) are CNS neurons. Their axons form the optic nerve, which is a white matter tract of the CNS, connecting the eye directly to the thalamus and visual cortex.^{[1, 2, 3, 10]}

Physiological Homology

The retinal and cerebral microvasculatures exhibit similar autoregulatory mechanisms to control blood flow in response to metabolic demands.^[11] This homology means that systemic vascular risk factors and diseases exert similar pathological effects on the small vessels of both the eye and the brain.^{[9, 11]}

1.2 Historical Context & Proof of Concept: Evidence from 2D Fundus Photography

Long before advanced imaging like OCT, the concept of using the retina to assess stroke risk was established through traditional 2D color fundus photography. Decades of epidemiological research have demonstrated that visible abnormalities in the retinal vasculature are powerful predictors of future cerebrovascular events.^{[13, 14, 15]}

Systematic reviews confirm strong associations between several key retinal biomarkers and increased stroke risk:^{[14, 15]}

• Retinal Vessel Caliber: Wider retinal venules (CRVE) are significantly associated with higher stroke risk (HR 1.20, 95% CI 1.10–1.31).^[14] Narrower retinal arterioles (CRAE) are also linked to increased risk (HR 1.18, 95% CI 1.04–1.34).^[14]
• Presence of Retinopathy: The appearance of retinopathy lesions is associated with a more than two-fold increase in stroke risk (pooled HR 2.70, 95% CI 1.65–4.43).^{[14, 15]}
• Vascular Network Geometry: Lower fractal dimension and increased arteriolar tortuosity are linked to higher stroke risk.^{[13, 14, 15]}
• Focal Abnormalities: Retinal arteriolar emboli and arteriovenous nicking are predictors of stroke risk.^{[14][14, 15]}

Despite these strong associations, predictive models based on 2D fundus biomarkers have reached a performance plateau. While some automated systems show good discrimination (AUC 0.84 with combined clinical and retinal variables)^[16], no predictive model using fundus imaging has consistently outperformed conventional clinical risk scores in external validation.^[15] This performance ceiling motivates the transition to more advanced imaging modalities.

The performance plateau observed with 2D fundus photography-based models is fundamentally a data limitation problem. The progression of retinal imaging technology from 2D projections to 3D volumetric and functional assessments represents a paradigm shift, unlocking a wealth of new, quantitative biomarkers that were previously inaccessible. This section details the limitations of the older technology and explains how OCT and OCT Angiography (OCT-A) overcome these challenges to enable a more profound analysis of cerebrovascular health.

2.1 Limitations of 2D Fundus Imaging

While foundational, 2D fundus photography has several inherent technological limitations:

• Lack of Depth Information: A fundus photograph is a two-dimensional projection, making it impossible to distinguish between different vascular layers of the retina.^[9]
• Qualitative Assessment: Manual interpretation is inherently qualitative and subject to significant inter- and intra-observer variability.^[18]
• Limited Biomarker Scope: The biomarkers accessible are primarily vascular in nature, providing only a partial picture of neurovascular changes.

2.2 Optical Coherence Tomography (OCT): Structural Precision

OCT represents a fundamental advancement in retinal imaging, providing high-resolution, cross-sectional images of the retina with micrometer-level precision.^{[20, 21]} Unlike fundus photography, which captures a 2D projection, OCT generates a series of 2D cross-sectional images (B-scans) that can be reconstructed into a 3D volume.

The key advantage of OCT is its ability to precisely segment and measure individual retinal layers. For stroke risk assessment, the most relevant layers are:

• Retinal Nerve Fiber Layer (RNFL): Composed of ganglion cell axons, RNFL thickness is a direct measure of axonal integrity.^{[22, 23]}
• Ganglion Cell-Inner Plexiform Layer (GCIPL): This layer contains the cell bodies of retinal ganglion cells and their synaptic connections.^{[24, 25]}

2.3 OCT Angiography (OCT-A): Functional Vascular Imaging

OCT-A represents the most recent advancement in retinal imaging for cerebrovascular assessment. Unlike traditional fluorescein angiography, which requires dye injection, OCT-A generates high-resolution images of the retinal vasculature by detecting motion contrast from flowing blood cells.^{[26, 27]}

The key innovation of OCT-A is its ability to provide depth-resolved imaging of the retinal vasculature, separating the different vascular plexuses that were previously superimposed in 2D fundus images. This capability has revealed novel biomarkers:

• Vessel Density: OCT-A can quantify the density of perfused vessels in specific layers.^[28]
• Foveal Avascular Zone (FAZ) Morphology: The FAZ is the central region of the macula that lacks blood vessels.^[29]
• Capillary Non-perfusion: OCT-A can identify areas where capillaries have been lost or occluded.

The high dimensionality of OCT/OCT-A data presents both opportunity and challenge. While these volumetric scans contain richer information than 2D images, they necessitate more complex AI architectures and larger datasets for robust training.

2.4 Novel Quantitative Biomarkers Enabled by OCT/OCT-A

The technological capabilities of OCT and OCT-A have unlocked a new class of quantitative biomarkers that provide a more granular assessment of neurovascular health:

• Vascular Density and Perfusion: OCT-A allows for the calculation of Vessel Density (VD) and Perfusion Density (PD), enabling precise localization of capillary dropout and ischemic areas.
• Foveal Avascular Zone (FAZ) Morphology: OCT-A can precisely measure FAZ area, perimeter, and circularity index, which are sensitive indicators of foveal ischemia.
• Neural Layer Thickness: Structural OCT provides precise measurements of RNFL and GCIPL thickness, signifying neurodegeneration.
• Vascular Geometry: Metrics such as fractal dimension and vessel tortuosity can be quantified with greater precision within specific vascular layers.

The true power lies in assessing both neural and vascular markers simultaneously, enabling investigation of the complex interplay between microvascular damage and neurodegeneration.

Table 1: Comparative Analysis of Retinal Imaging Modalities

Table 1: Comparative Analysis of Retinal Imaging Modalities

The evolution in imaging technology has been paralleled by an evolution in computational methods. This section provides a critical assessment of the current artificial intelligence (AI) landscape for stroke risk stratification.

3.1 Current AI Landscape: Performance on 2D Fundus Images

Deep learning, particularly using convolutional neural networks (CNNs), has been extensively applied to 2D fundus photographs for the prediction of cardiovascular disease (CVD) and stroke.^{[13, 14, 28]}

Regarding stroke prediction, the performance has been promising but variable. Studies have reported AUC values ranging from 0.57 to 0.78 for predicting or detecting stroke.^[13] More recent approaches, such as the MVS-Net using multi-view input from both eyes, have achieved an AUC of 0.84 for stroke and TIA detection.^[13]

A critical question is whether these AI models outperform established clinical risk scores. The evidence is mixed but encouraging. Several studies suggest AI models can achieve performance comparable to or superior to conventional risk scores like the Framingham Risk Score.^{[29, 30, 31]} However, a systematic review concluded that no fundus-based AI model has yet demonstrated consistent superiority over conventional risk scores in external validation.^[15]

Table 2: Performance of AI Models on 2D Fundus Images vs. Clinical Risk Scores

Table 2: Performance of AI Models on 2D Fundus Images vs. Clinical Risk Scores