Evolutionary Algorithm
September 1, 2022
Theodoros Zanos
September 1, 2022Article of the Month – September 2022
Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning
Ryan Poplin et al.
Objectives
Cardiovascular disease is the leading cause of death globally, affecting roughly 18.2 million adults aged 20 and older. As the vulnerability of patients with cardiovascular disease increases, so does the demand for cardiovascular disease risk calculators. Current cardiovascular disease risk calculators lack certain parameters, failing to identify necessary signals for cardiovascular risk factors. Certain markers of cardiovascular disease, such as hypertensive retinopathy, can be evident through the eyes. This study focuses on exploring whether additional signals for cardiovascular risk can be extracted from retinal images using deep convolutional neural networks (deep learning), which can be obtained quickly, cheaply, and non-invasively in an outpatient setting.
Methods
This study recruited 500,000 participants, aged 40-69 years. Two datasets were used in this study. The first dataset consisted of obtaining participants’ health measurements (smoking status, resting blood pressure). In addition, participants underwent paired retinal fundus and optical coherence tomography for fundus images. This dataset was divided into a development dataset to develop models and a validation dataset to assess the models’ performance. The second dataset-EyePACS-consisted of acquiring retinal fundus images from clinics’ routine clinical care for diabetic retinopathy screening. The images from both datasets were pre-processed for training and validation. Three separate neural network models were trained for predicting different risk factors: classification, regression, and MACE models. The performance of the neural network models was then assessed. First, MAE, the coefficient of determination, AUC, and Cohen’s k were used to evaluate the models’ performance for continuous predictions. The non-parametric bootstrap procedure and one-tailed binomial test were used to assess the statistical significance of the results. Finally, a deep-learning technique called soft attention (Fig. 2) was used to better understand how the neural-network models arrived at the predictions.
Results
The models were able to predict cardiovascular risk factors (age, systolic blood pressure, BMI, and HbA1c) better than baseline. The mean absolute error (MAE) for predicting the patient’s age was 3.26 years versus baseline’s 7.06 years. As shown in figure 1, the predicted age and actual age have a linear relationship. In addition, in 78% of the cases, the predicted age was within a 5-year margin of the actual age, whereas baseline predictions only fell into the 5-year margin 44% of the time. Results also displayed that major adverse cardiovascular events (MACE) within five years were predicted better when using risk calculators/models that combined risk factors. The results using the deep-learning technique, soft attention, showed that the algorithm used the blood vessels and optic disc to make its predictions.
Conclusions
This study indicates that the application of deep learning to retinal fundus images alone can be used to predict multiple cardiovascular risk factors, including age, gender, and systolic blood pressure. Therefore, neural network models can improve cardiovascular risk stratification.
Relevance to Healthcare Field
As cardiovascular disease quickly becomes the leading cause of death globally, there is a demand for methods to rapidly and accurately recognize risk factors to provide proper medical attention to patients. Forms of artificial intelligence (AI) can produce highly accurate algorithms that diagnose diseases from medical images with comparable accuracy to that of physicians. The adoption of these methods can significantly aid patients and physicians in multiple ways.