Detection of systemic cardiovascular illnesses and cardiometabolic risk factors with machine learning and optical coherence tomography angiography: a pilot study

Abstract

Background/Objectives

Optical coherence tomography angiography (OCTA) has been found to identify changes in the retinal microvasculature of people with various cardiometabolic factors. Machine learning has previously been applied within ophthalmic imaging but has not yet been applied to these risk factors. The study aims to assess the feasibility of predicting the presence or absence of cardiovascular conditions and their associated risk factors using machine learning and OCTA.

Methods

Cross-sectional study. Demographic and co-morbidity data was collected for each participant undergoing 3 × 3 mm, 6 × 6 mm and 8 × 8 mm OCTA scanning using the Carl Zeiss CIRRUS HD-OCT model 5000. The data was then pre-processed and randomly split into training and testing datasets (75%/25% split) before being applied to two models (Convolutional Neural Network and MoblieNetV2). Once developed on the training dataset, their performance was assessed on the unseen test dataset.

Results

Two hundred forty-seven participants were included. Both models performed best in predicting the presence of hyperlipidaemia in 3 × 3 mm scans with an AUC of 0.74 and 0.81, and accuracy of 0.79 for CNN and MobileNetV2 respectively. Modest performance was achieved in the identification of diabetes mellitus, hypertension and congestive heart failure in 3 × 3 mm scans (all with AUC and accuracy >0.5). There was no significant recognition for 6 × 6 and 8 × 8 mm for any cardiometabolic risk factor.

Conclusion

This study demonstrates the strength of ML to identify the presence cardiometabolic factors, in particular hyperlipidaemia, in high-resolution 3 × 3 mm OCTA scans. Early detection of risk factors prior to a clinically significant event, will assist in preventing adverse outcomes for people.

Introduction

Optical coherence tomography angiography (OCTA) has been widely studied in various ophthalmic conditions but has also been shown to detect changes in the retinal microvasculature of people with various systemic cardiometabolic factors [1–8]. These studies demonstrate the exciting potential of OCTA and its ability to provide further insights into systemic disease states utilising non-invasive imaging of the retinal microvasculature. Systemic conditions found to be associated with retinal microvascular changes include dyslipidaemia, hypertension, smoking, diabetes, chronic kidney disease, coronary artery disease, and peripheral vascular disease, as well as increasing age [1–8]. Most recently, in a study comparing various sized OCTA scans, associations with various cardiometabolic factors were shown to be progressively attenuated in larger sized OCTA scans (6 × 6 mm and 8 × 8 mm vs 3 × 3 mm), suggesting that cardiometabolic disease may preferentially affect the smallest calibre retinal vessels in the superficial capillary plexus which is best detected on the 3 × 3 mm sized scan [8].

Given these findings, machine learning represents an exciting opportunity to potentially play a role in the screening of retinal microvasculature and identifying changes before clinical manifestations. Artificial intelligence, including machine learning (ML), is becoming an increasingly popular field that is expected to revolutionise future health care. Its use continues to rise rapidly, including within the field of ophthalmology, where up to 11% of artificial intelligence studies are performed [9]. To date, there are no studies assessing retinal vasculature changes and cardiometabolic factors using OCTA and ML. The aims of this study were to assess the feasibility of predicting the presence or absence of systemic conditions (hyperlipidaemia, hypertension and diabetes mellitus) and cardiovascular conditions (ischaemic heart disease, atrial fibrillation, heart failure) using machine learning and OCTA.

Materials and methods

Data collection

We conducted a cross-sectional study at the Royal Adelaide Hospital from January 2019 to December 2019. Participants were recruited from the Department of Cardiology and Ophthalmology after informed and written consent were gained. Exclusion criteria included those with pre-existing or subsequently OCTA-diagnosed retinal vascular disease (diabetic retinopathy, retinal vascular occlusion, degenerative macular diseases), hemodynamic instability, inability to position for scanning, and severely myopic eyes (<−5 dioptres) [10]. Ethics approval was obtained from the Central Adelaide Local Health Network Ethics Committee.

Cardiometabolic profiling

For any participants identified for inclusion, the following data was collected: Age, gender, height, weight, comorbidities as clinically-diagnosed by their treating physicians via chart review (specifically: hypertension, dyslipidaemia, diabetes, heart failure, coronary artery disease, peripheral vascular disease, stroke, chronic kidney disease, obstructive sleep apnoea), smoking status, and current medications. The most recent blood tests were also recorded for each participant including fasting glucose, glycated haemoglobin, creatinine, troponin, total cholesterol, low density lipoprotein, high density lipoprotein, and triglycerides.

OCTA scanning

Participants underwent OCTA scanning without the use of mydriatics using the Carl Zeiss CIRRUS HD-OCT Model 5000 (Carl Zeiss AG, Oberkochen, Germany) performed by a trained technician. A minimum signal strength of 7/10 was required for inclusion. 3 × 3 mm, 6 × 6 mm and 8 × 8 mm macula OCTA scans were performed on one eye of included participants according to the following protocol: right eye for participants born in even years and left eye for participants born in odd years, the seeing eye in only-eyed participants, and where one eye resulted in an uninterpretable scan the contralateral eye was utilised for analysis. All images were assessed by two graders (MTS and WOC) and scans with artefacts considered to significantly impair quantitative analysis were excluded. Artefacts assessed included projection, motion, displacement, shadowing and vessel doubling [11].

Data pre-processing

Pre-processing and machine learning were conducted with open source Python libraries (namely Sci-Kit Learn and Tensorflow).

All OCT images were resized to 256 × 256 pixels and re-scaled (Fig. 1A). The total dataset was then randomly split into training and testing datasets (75%/25% split). This split was performed once. The holdout test dataset was not used for any analyses aside from performance assessment.

Fig. 1. Example of how each OCT image was processed.

A Example of a 3 × 3 mm OCTA image. B Salience map generated from the OCTA imaging following the application of a neural network aiming to predict the presence or absence of dyslipidaemia.

Additional training data was then generated using random transformations of the images in the existing training dataset. Random transformations included a degree of rotation, width or horizontal shift, and horizontal flipping. These transformations did not change the relative size or orientation of the blood vessels in the images. Ten additional versions of each image in the training dataset were thereby created through such random transformations, which were then used as part of the training dataset.

Model development

In the development of the convolutional neural network (CNN), initially a simple architecture (using one 2D convolutional layer, one 2D maximum pooling layer and one dense layer) was employed.

The second model was developed using transfer learning and MobileNetV2 [12]. Transfer learning is the process through which a previously trained model is then re-trained and applied to a new task, without re-initialising the model’s weights. Initially, this process was begun by providing the MobileNetV2 with a new maximum pooling layer and dense output layer, which were then trained (without training the rest of the model). After 50 epochs, the entire model was then trained on the training dataset for a further 50 epochs.

Model performance assessment and statistical analysis

Once models were developed on the training dataset, their performance was assessed on the unseen test dataset. Area under the receiver operator curve (AUC) was calculated using the trapezoidal rule with Sci-Kit Learn. Youden’s index was used to determine cut-off scores and thereby to calculate sensitivity, specificity, positive predictive value, negative predictive value, accuracy and F1 score [13]. Average-precision score was also calculated using Sci-Kit Learn.

Results

Participants characteristics

A total of 351 participants were screened for inclusion in the study. Subsequently, 104 scans were excluded leaving a final 247 included participants. Of the excluded scans, 4 scans were decentred, 5 patients were found to have retinal vascular disease, 17 scans had a signal strength <7/10, and the remaining 78 were evaluated and agreed upon by two authors to have significant artefact prohibiting accurate Angiotool analyses. The mean age was 67.8 ± 14.4 years and 175 (70.9%) were men. The number with dyslipidaemia was 116 (47.0%). 138 (55.9%) had hypertension and 76 (30.8%) had diabetes mellitus. The numbers of participants with coronary artery disease, and heart failure were 125 (50.6%) and 60 (24.3%) respectively.

Identification with CNN

The CNN performed best in the identification of hyperlipidaemia in 3 × 3 mm scans, with an AUC of 0.74 and accuracy of 0.79 (notably 50% of the cases in the test set had hyperlipidaemia—see Table 1). Figure 1B provides an example of a salience map for the identification of dyslipidaemia with this algorithm. For 3 × 3 mm scans and the identification of diabetes mellitus, the CNN also achieved an AUC of 0.74 and accuracy of 0.79. However, it should be recognised that only 21% of the test set had diabetes mellitus. Modest performance was achieved in the identification of hypertension (AUC 0.58, accuracy 0.67) and congestive heart failure (AUC 0.61, accuracy 0.75). The prediction of ischaemic heart disease and atrial fibrillation both resulted in AUC < 0.5. For 6 × 6 mm scans, there was no significant recognition for any of the cardiometabolic risk factors or cardiovascular conditions with all AUC being 0.5 or less. For 8 × 8 mm scans, the AUC for congestive cardiac failure was 0.58 with an accuracy of 0.69. All other conditions had an AUC of 0.5 or less.

Table 1.

Results of the application of machine learning algorithms to 3 × 3 mm OCTA scans to predict the presence of systemic and cardiovascular conditions.

Model	Disease	AUC	TP	FN	TN	FP	Sensitivity	Specificity	PPV	NPV	F1 Score	Average-precision score	Accuracy	Proportion of test dataset that is positive class
CNN	Hyperlipidaemia	0.74	10	2	9	3	0.83	0.75	0.77	0.82	0.80	0.71	0.79	0.50
	Diabetes Mellitus	0.74	4	1	15	4	0.80	0.79	0.50	0.94	0.62	0.50	0.79	0.21
	Hypertension	0.58	12	0	4	8	1.00	0.33	0.60	1.00	0.75	0.54	0.67	0.50
	Ischaemic heart disease	<0.5	16	1	1	6	0.94	0.14	0.73	0.50	0.82	0.71	0.71	0.71
	Congestive cardiac failure	0.61	4	3	14	3	0.57	0.82	0.57	0.82	0.57	0.38	0.75	0.29
	Atrial fibrillation	<0.5	3	1	5	15	0.75	0.25	0.17	0.83	0.27	0.15	0.33	0.17
MobileNetV2	Hyperlipidaemia	0.81	7	5	12	0	0.58	1.00	1.00	0.71	0.74	0.87	0.79	0.50
	Diabetes Mellitus	0.6	4	1	10	9	0.80	0.53	0.31	0.91	0.44	0.31	0.58	0.21
	Hypertension	0.69	11	1	3	9	0.92	0.25	0.55	0.75	0.69	0.69	0.58	0.50
	Ischaemic heart disease	0.59	8	9	6	1	0.47	0.86	0.89	0.40	0.62	0.81	0.58	0.71
	Congestive cardiac failure	<0.5	1	6	17	0	0.14	1.00	1.00	0.74	0.25	0.36	0.75	0.29
	Atrial fibrillation	0.55	4	0	7	13	1.00	0.35	0.24	1.00	0.38	0.22	0.46	0.17

Identification with MobileNetV2

Similarly to CNN, the MobileNetV2 performed best for 3 × 3 mm scans, particularly in the identification of hyperlipidaemia (AUC 0.81, accuracy 0.79). Hypertension was identified with an AUC of 0.69 and accuracy of 0.58. Diabetes mellitus had an AUC of 0.6 and accuracy 0.58 and heart failure had AUC < 0.5. The MobileNetV2 achieved an AUC > 0.5 for both ischaemic heart disease (AUC 0.59, accuracy 0.58) and atrial fibrillation (AUC 0.55, accuracy 0.46). However, both groups still had an accuracy lower than that of the most prevalent class in the test dataset. Again, similarly to the CNN, the identification of cardiovascular conditions and cardiometabolic risk factors was less significant for 6 × 6 mm and 8 × 8 mm scans. For the 6 × 6 mm scans, diabetes mellitus, hypertension and hyperlipidaemia had AUC of 0.59, 0.56 and 0.54 with accuracies of 0.73, 0.62 and 0.58 respectively. All other risk factors had an AUC < 0.5. For the 8 × 8 mm scans, hyperlipaemia and hypertension both had an AUC of 0.61 with accuracies of 0.69 and 0.62 respectively.

Discussion

The results of this study support the potential feasibility of using OCTA and machine learning to help identify certain systemic conditions. In particular, the performance of these algorithms using the 3 × 3 mm scan in identifying hyperlipidaemia support this idea with moderately high AUC and accuracy results for both models.

OCTA and its combined use with ML has been examined in a small number of other studies either using ML to assess image quality or in retinal vascular diseases, all of which have shown promising results [14]. For example, Lauermann et al. examined the ability of deep learning to assess image quality of 3 × 3mm superficial vascular plexus OCTA images. Their deep learning algorithm showed promising results with an accuracy of 90% achieved [15]. Aslam et al. examined the differentiation in diabetic status. They found that ML had the potential to recognise the difference on OCTA images of those without diabetes, with diabetes but without retinopathy and with diabetes with retinopathy [16]. Other conditions examined to date include diabetic retinopathy, the identification of choroidal neovascularisation and ability to differentiate between arteries and veins [17–20].

Although diabetic retinopathy has been examined in multiple studies, to the best of our knowledge, ours is the first study to investigate the identification of cardiometabolic risk factors using OCTA and ML. OCTA images have previously been analysed to identify the association between retinal microvascular changes and various cardiometabolic risk factors. This includes a reduction in retinal capillary density in participants with poorly controlled hypertension as measured using AngioVue 6 × 6 mm OCTA scans [2], and a reduction in vessel density and perfusion in hypertensive participants as compared to controls using Zeiss Cirrus 5000 3 × 3 mm OCTA scans [3]. Reduced vessel densities using AngioVue 3 × 3 mm scans have been shown in participants with chronic kidney disease [4], while pre-clinical diabetic changes in vessel density have also been demonstrated using Optovue 6 × 6 mm OCTA [5]. Smoking is associated with increased vessel density, vessel length, and junction density in 3 × 3 mm Zeiss OCTA The EYE-MI pilot study demonstrated reduced Angioplex measured vessel density in participants with impaired left ventricular ejection fraction and those with higher heart risk scores using a 3 × 3 mm Zeiss OCTA [7].

Within our study, both models performed best for 3 × 3 mm scans for all cardiometabolic factors. In particular, both models were strongest at being able to identify hyperlipidaemia. As the models were applied to both 6 × 6 mm and 8 × 8 mm scans, both the AUC and accuracy reduced significantly. This is consistent with previous literature where it has been reported that associations of vessel abnormalities with cardiometabolic factors were progressively weakened and statistically attenuated in 6 × 6 mm and 8 × 8 mm OCTA scans. These differences were attributed to calibre-specific differences in vessel detection with OCTA scans of varying sizes, with the highest-resolution 3 × 3 mm scan able to better quantify the smallest calibre retinal vessels [8]. Our findings support the idea that the earliest retinal microvascular changes associated with cardiometabolic conditions are most readily detectable in the small-calibre retinal vessels best captured using the highest-resolution OCTA scan.

As a pilot study, this project has significant limitations. In particular, the small sample size, with a therefore small test set of individuals, limits the generalisability of the results. The study was also conducted at a single centre. The recruitment methods employed may have resulted in a degree of selection bias (cardiology inpatients and ophthalmology outpatients) affecting the prevalence of conditions, such as congestive cardiac failure, and male predominance in the study cohort. Given the close interlink between cardiometabolic risk factors and their role in retinal vascular disease, the exclusion of such patients may have altered the population captured. This could be further explored in future studies.

Future research in this area may aim to utilise a larger cohort to derive more accurate models and externally evaluate model performance. In this study, binarization and skeletonization of OCTA images was not performed, due to previously identified issues with repeatability [21]. However, subsequent analyses may seek to evaluate OCTA pre-processed with these techniques. Subsequent studies may also examine the combination of OCTA and non-OCTA data (such as age and gender), to determine whether multi-modality inputs may improve model performance. The ability of machine learning and OCTA to predict the likelihood of other illnesses, such as future stroke, or parameters that may serve as biomarkers, such as predicted age, may also be areas for further study. Further studies are required to gauge how accurate such methods may become when provided with larger datasets.

In conclusion, the rising use of OCTA and ML brings an exciting and promising future for efficient and non-invasive screening of people for cardiometabolic risk factors. This would be particularly useful if such risk factors could be identified and managed prior to a clinically significant event. This study has demonstrated the strength in both ML models to identify cardiometabolic factors, in particular hyperlipidaemia, in high-resolution 3×3mm OCTA scans.

Summary

What was known before

• Optical coherence tomography angiography (OCTA) is able to identify changes in the retinal microvasculature of people with various cardiometabolic factors

• Machine learning has previously been applied to ophthalmic imaging

What this study adds

• Machine learning can identify various cardiometabolic risk factors in 3 × 3 mm OCTA scans, particularly hyperlipidaemia

• This study represents the exciting possibility of being able to non-invasively screen people for cardiometabolic risk factors with OCTA and ML, prior to a clinically significant event

Author contributions

SH, SB, WC, CXW, MTS were all responsible for data collection, designing the study, analysing results, creation of tables, drafting of the study and review of the final manuscript. DS and CM were responsible for result interpretation and review of the final manuscript.

Data availability

The data that support the findings of this study are not openly available due to reasons of sensitivity. They are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.