Diabetic Retinopathy: EPC Mitochondrial Function and Progression

Two posters presented at AAO 2019 will highlight some of the latest research pertaining to diabetic retinopathy.

The first, “EPC Mitochondrial Function: An Indicator for Angiogenesis in PDR,” sought to determine whether there was any relationship between endothelial progenitor cell (EPC) mitochondrial function, Wnt signaling pathway activity, and EPC number in patients with diabetic retinopathy. To do this, the researchers collected and analyzed plasma from patients with and without diabetic retinopathy. They used flow cytometry to measure the number and mitochondrial function of EPC, and enzyme-linked immunosorbent assay was used to acquire the plasma Dickkopf-1 level. This information allowed the researchers to observe any relationships between EPC mitochondrial function, EPC number, and Wnt activity.

The analysis presented a positive correlation between EPC mitochondrial function and EPC numbers (r = 0.4443, P = 0.0036) and the Wnt regulator Dickkopf-1 level (r = 0.3676, P = 0.0181) in human plasma.

“Our results reveal the potential of EPC mitochondrial function and metabolic profile as a promising prognosis for pathological vascular formation, especially in [diabetic retinopathy] and other neovascularization ocular diseases,” the researchers concluded.

The second scientific poster, “Machine Learning for Prediction of Diabetic Retinopathy Progression,” evaluated the potential of machine learning to forecast diabetic retinopathy progression through teleophthalmology fundus images. This study included 2,094 patient screenings who were re-screened within 2.5 years. The researchers combined feature extraction with pretrained convolutional neural networks (CNNs) or support vector machines (SVMs).

When applied to original images, CNNs had an accuracy rate between 70% and 75% when faced with three classes (stable, improved, worsened); with two classes, the accuracy rate ranged between 82% and 86%. With modified images that isolated retinal vessels autonomously, accuracies improved by 1 to 10 percentage points. The highest performing model achieved a 75.9% specificity rate. However, because of false negatives, sensitivity was 33.3%.

The study authors concluded that “CNNs and SVMs show potential in identifying retinal images at risk of DR progression. Performance improves with isolation of retinal vessels and addition of patient parameters.”