Machine Learning Prediction of Cardiovascular Risk in Type 1 Diabetes Mellitus Using Radiomic Features from Multimodal Retinal Images

Abstract

Purpose: To develop a machine learning (ML) algorithm capable of determining cardiovascular (CV) risk in multimodal retinal images from patients with type 1 diabetes mellitus (T1DM), distinguishing between moderate, high, and very high-risk levels.

Design: Cross-sectional analysis of a retinal image data set from a previous prospective OCT angiography (OCTA) study (ClinicalTrials.gov NCT03422965).

Participants: Patients with T1DM included in the progenitor study.

Methods: Radiomic features were extracted from color fundus photographs (CFPs), OCT, and OCTA images, and ML models were trained using these features either individually or combined with clinical data (demographics and systemic data, OCT + OCTA commercial software metrics, ocular data, blood data). Different data combinations were tested to determine the CV risk stages, defined according to international classifications.

Main outcome measures: Area under the receiver operating characteristic curve mean and standard deviation for each ML model and each data combination.

Results: A data set of 597 eyes (359 individuals) was analyzed. Models trained only with the radiomic features achieved area under the curve (AUC) values of (0.79 ± 0.03) to identify moderate risk cases from high and very high-risk cases, and (0.73 ± 0.07) for distinguishing between high and very high-risk cases. The addition of clinical variables improved all AUC values, obtaining (0.99 ± 0.01) for identifying moderate cases, and (0.95 ± 0.02) for differentiating between high and very high-risk cases. For very high CV risk, radiomics combined with OCT + OCTA metrics and ocular data achieved an AUC of (0.89 ± 0.02) without systemic data input. The performance of the models was similar in unilateral and bilateral eye image data sets.

Conclusions: Radiomic features obtained from retinal images are helpful to discriminate and classify CV risk labels, differentiating risk categories. The addition of demographics and systemic data combined with ocular data differentiate high from very high CV risk cases, and interestingly OCT + OCTA metrics with ocular data identify very high CV risk cases without systemic data input. These results reflect the potential of this oculomics approach for CV risk assessment.

Financial disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Keywords: Cardiovascular risk; Diabetes mellitus type I; Machine learning; Optical coherence tomography angiography; Radiomics.

Figure 1

Consolidated standard for outcome reporting trials (CONSORT) flowchart diagram of included and excluded patients and eyes.

Figure 2

Retinal images collected for each individual eye included in the study data set. A, Color fundus photographs (CFP); B: Structural OCT macular scan; C: OCT angiography (OCTA) 3 × 3-mm superficial capillary plexus (SCP); D: OCTA 3 × 3-mm deep capillary plexus (DCP); E: OCTA 6 × 6-mm SCP; F: OCTA 6 × 6-mm DCP. (CFP: Topcon DRI-Triton, Topcon Corp; OCT and OCTA: Cirrus, Carl Zeiss Meditec) (adapted from Carrera-Escale et al⁹).

Figure 3

Data groups scheme of the attributes used for the classification tasks. DCP = deep capillary plexus; DM = diabetes mellitus; DR = diabetic retinopathy; eGFR = estimated glomerular filtration rate; HbA1c = glycosylated hemoglobin; HDL = high-density lipoprotein; LDL = low-density lipoprotein; OCTA = OCT angiography; SCP = superficial capillary plexus.

Figure 4

Pipeline of the methodology followed to train and optimize the machine learning models. This process is repeated for all the proposed classification tasks, data combinations, and models. AUC = area under the receiver operating characteristic curve; RF = random forest.

Figure 5

Machine learning models performance for all the classification tasks and data combinations, representing area under the curve (AUC) values. Left: AUC values for distinguishing between moderate risk and high or very high risk. Right: AUC values for distinguishing between high risk and very high risk. All these results are extracted using both eyes per patient. B = blood analysis data; D = demographics and systemic data; LR = logistic regression; O = ocular data; R = radiomics of retinal images; RF = random forest; S = commercial software OCT and OCT angiography data; SVC = support vector classifier.

Figure 6

Receiver operating characteristic (ROC) curves of models’ performance for all the classification tasks and data combinations. Left column: ROC curves discerning between moderate risk and high or very high risk. Right column: ROC curves for distinguishing between high risk and very high risk. Top row: logistic regression model. Second row: support vector classifier using linear kernel (SVC-linear) model. Third row: support vector classifier using radial basis function kernel (SVC-rbf) model. Bottom row: random forest (RF) model. All these results are extracted using both eyes per patient. B = blood analysis data; D = demographic and systemic data; O = Ocular data; R = radiomics of retinal images; S = data from the commercial software.

Figure 7

Machine learning models performance for both classification tasks, evaluated using the best data combinations but changing the type of images used in each case. It presents area under the curve (AUC) values. Top: AUC values for distinguishing between moderate risk and high or very high risk. Bottom: AUC values for distinguishing between high risk and very high risk. All these results are extracted using both eyes per patient. O = ocular data; R = radiomics of retinal images; S = commercial software OCT and OCT angiography data.

Figure 8

Multiclass receiver operating characteristic (ROC) curves of the final models, evaluated on the most relevant data combinations. Class 0 refers to moderate risk, class 1 to high risk, and class 2 to very high risk. Left: radiomics only. Middle: ophthalmic data (radiomics, ocular, and commercial software). Right: all available data. AUC = area under the curve.