Stroke risk assessment for the community by automatic retinal image analysis using fundus photograph

Background: Primary prevention of stroke is vital for saving lives and disabilities, and retina characteristics have been investigated as potential tools for stroke risk assessment. This study reports the development of a statistical model for stroke risk assessment using manually digitized retinal image characteristics obtained from a case-control study. We further report the results of a fully automatic version of the analysis (ARIA-stroke) on the study. The model was then validated using a separate dataset to show that it can be applied in a primary care setting. Methods: We have carried out a case-control study with 244 subjects (122 strokes and 122 controls). About 66% of each group was diabetes patients. A manual digitization process was used to measure retinal characteristics including central retinal artery equivalent (CRAE), central retinal vein equivalent (CRVE), arteriole-venule ratio (AVR), bifurcation coefficients, bifurcation angles, and bifurcation asymmetries, arteriole-venous nicking, tortuosity, hemorrhage, exudates, and arteriole occlusions. Logistic models were developed to evaluate both the clinical and retinal characteristics. A fully automatic approach for the analysis of the retinal images was developed and the method was validated using a separate data set with 412 subjects (138 normal controls, 198 hypertensions and 76 stroke cases) Results: The manual analysis shows that retinal characteristics are valuable in stroke risk assessment with AUC of 0.78 (95% C.I. 0.72-0.84) for retinal characteristics alone versus AUC of 0.66 (95% C.I. 0.59-0.73) for clinical variables alone. The combined model with both clinical and retinal characteristics has an AUC of 0.84 (95% C.I. 0.78- 0.89) outperformed model using clinical or retinal variables alone. For the automatic ARIA-stroke model, the average probability of stroke for the control group was 0.141 (95% CI: 0.126-0.156), and the case group was 0.847 (95% CI: 0.839-0.855). When we looked at the patient subgroups with and without diabetes, the average probability of stroke for the control without diabetes was 0.054 (95% CI: 0.046-0.063), control with diabetes was 0.185 (95% CI: 0.170-0.199), stroke without diabetes was 0.853 (95% CI: 0.841-0.866), stroke with diabetes was 0.843 (95% CI: 0.833-0.854). The sensitivity and specificity was 100% in the case-control study using a probability cutoff of 0.5. We have also estimated the retinal Benny Zee 115 Background Stroke is a disease with high mortality and debilitating even for survivors. It generates great financial burden on survivors’ families and the health care system worldwide. Krishnamurthi et al. reported that the global burden of ischemic and hemorrhagic stroke increased significantly between 1990 and 2010 in terms of the absolute number of cases, number of deaths, and disability-adjusted life years (DALY) lost.1 They found that the global burden of strokes increased in low-income and middle-income countries as opposed to high-income countries. This has become an important global health issue. Various interventions for stroke prevention are available and some have shown to be effective, but the challenge is on the ability to provide a more specific and accurate classification. From an individual-based prevention perspective, there are various ways to assess the risk of stroke. They include ultrasound, computed tomography angiography (CTA), and magnetic resonance angiography (MRA). Ultrasound can assess stenosis and blood velocity of vessels in relative superficial surfaces and is widely used to evaluate the carotid stenosis. More than 70% stenosis is indication for carotid endarterectomy. However, stroke caused by carotid stenosis accounts for only 4% of all stroke cases.2 CTA and MRA can detect abnormality of larger cerebral vessels, but these techniques are costly, inconvenience and invasive. From a population-based prevention perspective, we can substantially reduce the burden of stroke if we reduce blood pressure, promote physical activity, increase smoking cessation, and a healthy diet.3 However, tools to estimate stroke risk for an individual are not well developed. Feigin et al. suggested the use of Stroke Riskometer App in addition to other tools such as Framingham and QSTROKE stroke risk prevention algorithms. The mobile app-based approach is promising and may increase general awareness of the importance of stroke risk reduction, but the accuracy remains to be proven.4 Cerebral vascular change is one of the major pathology causes of stroke. Retina vessel circulation shares similar morphology, function, and pathologic changes with cerebral vascular system. Since retina is the only place throughout the body where a small part of the vascular system can be observed directly, cerebral vascular changes can be explored through retinal image to determine the risk of strokes. Previous studies have shown that a number of retinal characteristics were significantly associated with strokes.5-9 However, none of them demonstrated they were adequate for stroke risk estimation. In this paper, we extracted the retinal parameters from color fundus images and identified risk factors associated with stroke cases; we further explored the use of retinal characteristics in a multivariate model for stroke risk assessment. Furthermore, we employed a novel method to automate the analysis of the retinal image for stroke risk assessment and to estimate the retinal parameters using data from a case-control study. We then validated the methodology using a separate data set. METHODS In the initial case-control study, 122 stroke cases were entered from an Acute Stroke Unit in collaboration with the diabetic retinopathy screening program in Hong Kong. The patients were diagnosed with either ischemic stroke or hemorrhagic stroke and had adequate sitting balance to carry out the retinal photography. There were 81 stroke cases with diabetes and 41 stroke cases without diabetes. Patients who were age 80 years or older were not included, since this age group is likely to have optical opacity and other complication that was not suitable for capturing color retina photo and may introduce bias of other sources. Patients with eye disease that had influence on the retinal vessel structures or spot characteristics and those with stroke subtypes of cardioembolic stroke, and subarachnoid hemorrhage were excluded. Patients suspected to suffer from cerebral diseases and those with disease that influence vessel morphology were also excluded. 122 control subjects matched with age and diabetic status were selected. Controls subjects without stroke were recruited from Eye Outpatient Clinics or diabetic retinopathy screening program. Only patients with routine eye checkup, recovered central serous chorioretinopathy, mild quiet age related maculopathy confirmed by fluorescein and indocyanine green angiography were included as controls. The mean length of follow-up period from the date of taking the retinal image was 4.3 years. All the controls were aged from 50 to 80 years old and have no retinal disease or with only mild diseases without influencing vessel structure in color retina images, such as mild dry age-related maculopathy, central serous chorioretinopathy, post-cataract extraction, retinal pigment epithelial detachment. Written informed consent was obtained, and the project was done according to the guidelines of the Declaration of Helsinki and approved by the Joint CUHK-NTEC Clinical Research Ethics Committee. Clinical risk factors Stroke risk factors including age, gender, hypertension, diabetes, hyperlipidemia, smoking status, histories of ischemic heart disease and atrial fibrillation were recorded in the study. Hypertension was defined as systolic blood pressure greater than 140 mm Hg, diastolic blood pressure above 90 mm Hg, or use of antihypertensive medication during the previous 2 parameters that are potentially useful for interpretation of the results. The observed data have significantly high correlations with the estimated values showing high goodness-of-fits. The validation study using a separate data set with normal controls, hypertension controls, and stroke cases have confirmed the results with a cutoff probability of 0.5, the sensitivity is 97% and specificity is 100%. Conclusion: This study demonstrated that retinal images contain valuable information for stroke risk assessment in addition to conventional clinical variables. A fast and fully automatic method can be used to estimate risk of stroke based on fundus photographs alone. We have also shown that a number of retinal characteristics may provide insights on clinical interpretation of the risk estimate and this method may be used in community setting or population screening.