Deep Learning Models Predict Age, Sex and Race from OCT Optic Nerve Head Circle Scans

Abstract

Purpose : To develop and evaluate the accuracy of deep learning algorithms for predicting age, sex and race from Spectralis optical coherence tomography (OCT) retinal nerve fiber layer (RNFL) circle scans.

Methods : RNFL circle scans acquired from healthy individuals and glaucoma suspect and glaucoma patients from the Diagnostic Imaging in Glaucoma Study (DIGS) and the African Descent and Glaucoma Evaluation Study (ADAGES) were randomly assigned to training (85%), validation(5%) and testing(10%) datasets by patient. Three deep learning models were trained on unsegmented RNFL circle scans for predicting age, sex and self-reported race. The dataset for age included 10096 images from 676 individuals. Age at the time of imaging was determined retrospectively by subtracting birthdate from OCT imaging date, ranging from 18 to 96 years. The dataset for biological sex included 8986 images from 608 individuals. The dataset for race included 5594 images from 382 individuals of European descent and 3356 images from 219 individuals of African descent. Other racial groups were excluded from the model due to low sample sizes. The deep learning model was evaluated using mean absolute error (MAE) or area under the receiver operating characteristic curve (AUROC) and mean probability.

Results : The individual deep learning models accurately predicted age with a MAE (95% CI) within 4.5 years (3.9, 5.2) and a strong (R² (95% CI)) association between predicted and actual age of 0.73 (0.69, 0.81). Diagnostic accuracies (AUROC (95% CI)) for predicting race and sex were 0.96 (0.86, 0.99) and 0.70 (0.57, 0.80), respectively. With a value of 0% for European descent, and 100% for African descent, the mean probability for ED was 10.1% (95% CI: 4.6-15.6%) while for AD was 94.9% (95% CI: 90.7-99.2%). The average number of images for ED was 43 (9,126), and for AD was 44 (18,108). With a value of 0% for Female, and 100% for Male, the mean probability for Female was 25.9% (95% CI: 19.5-32.4%) while for Male was 43.9% (95% CI: 34.2-53.7%).

Conclusions : These results suggest that there is information in raw OCT scans that varies by demographic variables including age, race and sex. Moreover, deep learning models can accurately pick up features not apparent to human reviewers. While there are easier ways to determine demographics, the research implies there is more knowledge to be discovered and data to be derived from retinal OCT images.

This is a 2020 ARVO Annual Meeting abstract.