Stroke Occurring on the Left or Right Side of the Brain Can Be Identied by the Analysis of Retinal Images: A Case Control Study

Background: The burden of stroke in China has increased dramatically in the past 30 years. Specifying the treatment according to the side of stroke in the brain might be an effective way to reduce the burden. Current imaging tools to identify the side of brain stroke, such as magnetic resonance imaging (MRI), are expensive and time-consuming. Hence, there is a great need for a rapid and inexpensive assessment. In this case, retinal image analysis is a possible approach for stroke side identication. This study aimed at determining the association between retinal characteristics and the stroke side and to establish a predictive model for further investigation. Methods: A total of 168 patients (89 left-sided stroke patients and 79 right-sided stroke patients) were recruited from the Shenzhen Traditional Chinese Medicine Hospital in the study. Retinal characteristics were analysed using an automated retinal image analysis (ARIA) system. Multivariable logistic regression was used to identify and develop predictive models. Results: Each unit increase in the right eye bifurcation coecient of arterioles increased the risk of right-side stroke by 7.523 times (95% CI, 1.823–31.044). Additionally, an elevated bifurcation coecient of venules in the right eye also increased the risk of stroke in the right side of the brain, with an odds ratio (OR) of 7.377 (95% CI, 1.771–30.724). A complex retinal composite score was also associated with a higher risk of right-side stroke (OR, 4.955; 95% CI, 3.061–8.022). Conclusions: This study demonstrated that retinal images can provide useful information for stroke side identication, and specic retinal characteristics may help predict the occurrence of stroke.


Introduction
Stroke has become the leading cause of death in China with an increased disease burden for the last 30 years. The rapidly ageing population is one of the major reasons for this high prevalence. (1,2) The location of stroke in the brain has a strong in uence on the clinical consequences. (3) In most cases, stroke occurs only in one hemisphere of the brain and results in impairment on the opposite side of the body. Studies reveal that stroke in the left hemisphere affects motor control(4) as well as language and speech, (5) whereas right hemisphere stroke affects spatial orientation and posture.(4) Therefore, speci c clinical advice and treatments are required for the patients depending on the stroke side.
Current techniques for stroke diagnosis have been well studied and accepted. Neurological examinations, such as the National Institutes of Health Stroke Scale (NIHSS) and Barthel index, provide information on the severity of stroke and activities of daily living, whereas imaging techniques such as magnetic resonance imaging (MRI) and computed tomography (CT) scans assist in stroke diagnosis by identifying the locations and damages in the brain.(6) However, these techniques have some limitations, CT imaging exposes the patients to a high radiation dose while MRI is an expensive, time-consuming, and environment-restricted procedure.(6, 7) Therefore, there is a great need for new tools that can rapidly differentiate the damaged side of the brain after strokes at a relatively low cost.
The function of the retina makes it one of the most metabolically active tissues in the body and the double vascular network maintains its blood supply. A small part of the circulation can be observed directly through the retina. (8) Previous studies have shown that some characteristics of retinal microvasculature are associated with strokes, which may be predicted by computational retinal analysis. (9)(10)(11)(12)(13)(14) Most studies have investigated only the risks or other aspects of strokes but the association between retinal parameters and the stroke side has not been demonstrated so far. In this study, retinal parameters were extracted from colour fundus retinal images and the retinal characteristics were used via a logistic model to explore the association further.

Data source
In this study, 254 stroke patients from Shenzhen Traditional Chinese Medicine Hospital were enrolled.
Baseline data, including age, sex, medical history, physical examination, laboratory tests, NIHSS, and Barthel index were collected. All patients underwent a detailed cranial MRI scan.
All patients were diagnosed with either left or right hemispheric stroke, aged 30-80 years, with an adequate sitting balance to undergo retinal photography. Patients aged over 80 years were not included in the study because of much higher prevalence of age-related optical opacity and other comorbidities as compared to the younger patients, making them unsuitable for retinal photography and may have caused a bias. Patients with eye diseases in uencing the retinal vasculature, and those who are physically or subjectively unable to comply with MRI scans were excluded. Furthermore, subjects suspected to have cerebral diseases and those with diseases in uencing vessel morphology were also excluded. Ultimately, 168 of the 254 patients who presented with stroke were included in this study.
The clinical and radiological information of the patients along with the stroke side was evaluated and their association with retinal characteristics were analysed.

Retinal characteristics
Retinal characteristics include retinal vessel measurements, arteriole-venous nicking, arteriole occlusion, tortuosity, haemorrhage, exudates, asymmetry of branches, bifurcation coe cients (BC), and bifurcation angles (BA). According to the formula developed by Knudtson et al., the retinal vessel measurements were summarized into central retinal artery equivalent (CRAE) and central retinal vein equivalent (CRVE), representing the diameters of arterioles and venules respectively. (15) The arteriole-to-venule ratio (AVR) was calculated as the ratio of CRAE to CRVE. Vessel tortuosity provides both qualitative and quantitative information by visual grading of one fovea-centred and one disc-centred fundus image. Arteriole venous nicking results in narrowing of the venule when crossed by an arteriole, and arteriole occlusions referred to the blockage of blood ow inside the arterioles when obstructed by an emboli. The branching pattern of retinal vessels, such as bifurcation coe cient (BC); the ratio of the widths of branching vessels to trunk vessels, bifurcation angle (BA); the angle between two branching vessels and asymmetry; the ratio of diameters of two branching vessels demonstrate the relationship between the trunk and the branching vessels. The fractal dimension (FD) of the retinal vasculature measured the complexity of branching patterns. The analysis was performed using three sets of vessels in one retinal image. Haemorrhages and exudates were recorded as present or absent and indicated in probability (0 to 1).

Statistical Analysis
Descriptive statistics, chi-square tests, and two-tailed independent samples t-tests were used to compare the demographics and retinal characteristics between the left and right stroke sides. A p-value < 0.05 was considered statistically signi cant. Odds ratios (ORs) and corresponding 95% con dence intervals (95% CIs) were obtained by logistic regression to control for confounding. Stepwise multivariable logistic regression was employed to select the best model. The classi cation accuracy and the area under the receiver operating characteristic (ROC) curve were measured. Leave-one-out classi cation was used to validate the developed model. All the data was analysed using SPSS 25.0 software (IBM, New York, USA).
The fully automatic retinal image analysis method was developed using R (University of Auckland, Auckland) and MATLAB (MathWorks, Massachusetts, USA) computer software. The detailed procedure of the automatic retinal imaging analysis can be found in Zee. (16)

Descriptive demographic variables
In total, 168 patients (133 men and 35 women) with mean age of 55.72 years, 89 left-side strokes and 79 right-side strokes were analysed in this study. Among these patients, 113 (67.3%) had ischemic strokes and 50 (29.8%) had haemorrhagic strokes. Smokers and drinkers in this analysis include both current and former status. A summary of the demographic variables is listed in Table 1. In general, there was no signi cant difference in demographics between the patients with stroke on the left side and the right side.   Using left-side stroke as the reference, the results from the multivariable logistic regression analysis of retinal characteristics are shown in Table 3. As shown in this table, two right-eye retinal characteristics and the retinal composite score were strongly associated with the stroke side. Patients with an increased bifurcation coe cient of arterioles in the right eye (RBCA) were at seven times higher risk of right-sided strokes, with an OR of 7.523 (95% CI, 1.823-31.044; P = 0.005). In addition, a higher value of bifurcation coe cient of venules in the right eye (RBCV) was also associated with a higher risk of right-side strokes, with an OR of 7.377 (95% CI, 1.771-30.724; P = 0.006). Furthermore, a retinal composite score including complex effects containing tortuosity, haemorrhage, and fractal dimension of the vessel network, were also positively associated with right-side stroke (OR, 4.955; 95% CI, 3.061-8.022; P < 0.001). The area under the ROC curve was 0.842 (95% CI, 0.782-0.901) for the multivariate logistic regression with a sensitivity of 79.8% and speci city of 73.4% using a cutoff of 0.5 probability, indicating an excellent potential predictive ability. (Fig. 1)

Validation
The validation of the logistic model was performed via leave-one-out cross-validation. Using RBCA, RBCV, and retinal composite score, 76.2% of left-side strokes and 74.4% of right-side strokes were correctly classi ed with a total accuracy of 75.3%.

Discussion
Previous studies have revealed the association between retinal characteristics and stroke in many aspects, but here we further investigated the association of the brain side with stroke. It might be an effective way to reduce the burden by specifying the treatment according to the side of the stroke. Hence, identifying the stroke side was one of the critical roles. The current imaging tool and the gold standard would be MRI, but since it is an expensive and a time-consuming procedure, its use clinically and in research is limited. Other imaging tests, such as CT and angiography, help in identifying the location hours after strokes, but one of the problems using CT is the high radiation dose. Since the changes in retinal vessels can re ect some abnormalities in the cerebrovascular system, retinal characteristics might also provide information for diagnosis and treatment of brain diseases. To our knowledge, no study has investigated the association between retinal characteristics and stroke sides. Therefore, this study is the rst to demonstrate that the eyes on one side might not correspond to the same side of the brain, but the retinal characteristics of both the eyes together would be able to determine the side of stroke. We further established a prediction model that may provide an additional source for stroke side identi cation. The retinal characteristics provided a classi cation of the left and right stroke sides with 76.8% accuracy based on the current study.
Our study found that an increased bifurcation coe cient for arterioles and venules was associated with a higher risk of right-side stroke, which indicated changes in the branching pattern of retinal vessels. In normal conditions, the pattern of retinal vasculature would provide an optimal route for blood ow with minimal energy costs. Although there is no literature available to support this nding, it might re ect that the side of the brain where the vessels are damaged by strokes could be identi ed via bifurcation coe cient in the retina. Since the biological effects are thought to be more complex, we explored the interactions between retinal characteristics. In this study, we found that as a series of retinal components increased at the same time, the risk of right-side stroke increased.
This study has several strengths and limitations. One of the major strengths of this study is that it is the rst study to investigate the association between retinal characteristics and the stroke side and it also establishes a preliminary prediction model for identi cation. Some limitations should also be noted. First, the small sample size and missing values affect the statistical power of the model; for instance, exudate is a useful retinal parameter, but due to a large number of missing values, it cannot be presented in the model. In addition, a cross-sectional design makes it di cult for causal inferences. As a preliminary study, the underlying mechanisms for the associations are still unclear. In the future, we need to increase the sample size and design a better study model to investigate the issue. Moreover, to ful l the need for clinical application, this model should also be validated.

Conclusion
In conclusion, our study provides an evidence of the association between retinal characteristics and stroke side. With the prediction model, in the future, we would be able to standardize the retinal image assessment automatically for the management of patients with stroke. This might be useful to improve the treatment and rehabilitation.

Authors' contributions
LnJ, LeJ, ZB and ZY were involved in the study design. ZY and ZW acquired the date. YH, HX and WJ provided clinical assistance. LeJ and ZB calculated retinal characteristics using the ARIA system. LnJ, ZY, LeJ, and ZB prepared the draft of the manuscript. YZ and YH revised the manuscript. All authors have read and approved the nal manuscript.