Abstract
Purpose :
Deep learning system (DLS) is a machine learning technology and has gained tremendous traction in medical imaging recently. In the field of ophthalmology, it has also been shown that DLS can extract new knowledge from retinal fundus images. The aim of this study was to evaluate the performance of a newly developed DLS in estimating Haemoglobin A1c (HbA1c), an important marker for diabetes.
Methods :
In this study, we retrospectively included 17,422 multi-ethnic Asian individuals from 5 population-based and clinical eye studies in Singapore. Retinal photos and clinical HbA1c measurement from serum sample were collected for all participants. Of the total set, 13,937 individuals’ data (25,637 fundus images) were used for training and development of the DLS; another independent set of 3,485 individuals’ data (6380 fundus images) were used for validation of the DLS. Only macular-centred retinal photos of sufficient quality were selected for DLS training and subsequent validation. Serum HbA1c measurement was used as ‘ground truth’ data in this instance. The mean (standard deviation) of serum HbA1c in training set was 6.31% (1.32), with a wide range of 3.9 to 15.1%. The ResNet and DenseNet convoluted neural networks were used for training of the DLS model. Agreement between DLS-predicted and actual serum HbA1c measurements were evaluated with mean absolute error (MAE), Bland Altman (BA) plots (mean difference and limits of agreement [LOA]), and intraclass correlation coefficient (ICC).
Results :
In the validation dataset, the overall MAE between DLS-predicted and serum HbA1c measurements was 0.87%, the BA plot mean difference was 0.18% [95% LOA -2.39 to 2.76%], and ICC was 0.60 [95% CI 0.57-0.62, indicating fair agreement between the 2 measurements with slight overestimation by the DLS. Among diabetics, the overall MAE between the 2 measurements was 1.19%, with a BA plot mean difference of -0.57% [95% LOA -3.76 to 2.61%], showing slight underestimation by the DLS in this subgroup. On the other hand, among non-diabetic individuals, the overall MAE between the 2 measurements was 0.62%, with a BA plot mean difference of 0.69 [95% LOA -0.62 to 2.02%], showing slight overestimation by the DLS among healthy individuals.
Conclusions :
We developed a novel DLS which shows early promising performance in estimating HbA1c. When further validated, this new modality may act as a new point-of-care tool for diabetic screening and monitoring.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.