Predicting risk of late age-related macular degeneration using deep learning

Abstract

By 2040, age-related macular degeneration (AMD) will affect ~288 million people worldwide. Identifying individuals at high risk of progression to late AMD, the sight-threatening stage, is critical for clinical actions, including medical interventions and timely monitoring. Although deep learning has shown promise in diagnosing/screening AMD using color fundus photographs, it remains difficult to predict individuals' risks of late AMD accurately. For both tasks, these initial deep learning attempts have remained largely unvalidated in independent cohorts. Here, we demonstrate how deep learning and survival analysis can predict the probability of progression to late AMD using 3298 participants (over 80,000 images) from the Age-Related Eye Disease Studies AREDS and AREDS2, the largest longitudinal clinical trials in AMD. When validated against an independent test data set of 601 participants, our model achieved high prognostic accuracy (5-year C-statistic 86.4 (95% confidence interval 86.2-86.6)) that substantially exceeded that of retinal specialists using two existing clinical standards (81.3 (81.1-81.5) and 82.0 (81.8-82.3), respectively). Interestingly, our approach offers additional strengths over the existing clinical standards in AMD prognosis (e.g., risk ascertainment above 50%) and is likely to be highly generalizable, given the breadth of training data from 82 US retinal specialty clinics. Indeed, during external validation through training on AREDS and testing on AREDS2 as an independent cohort, our model retained substantially higher prognostic accuracy than existing clinical standards. These results highlight the potential of deep learning systems to enhance clinical decision-making in AMD patients.

Keywords: Eye manifestations; Image processing; Machine learning; Prognostic markers.

Fig. 1. The two-step architecture of the framework.

a Raw color fundus photographs (CFP; field 2, i.e., 30° imaging field centered at the fovea). b Deep classification network, trained with CFP (all manually graded by reading center human experts). c Resulting deep features or deep learning grading. d Survival model, trained with imaging data, and participant demographic information, with/without genotype information: ARMS2 rs10490924, CFH rs1061170, and 52-SNP Genetic Risk Score. e Late age-related macular degeneration survival probability.

Fig. 2. Creation of the study data sets.

To avoid ‘cross-contamination’ between the training and test data sets, no participant was in more than one group.

Fig. 3. Prediction error curves.

Prediction error curves of the survival models in predicting risk of progression to late age-related macular degeneration on the combined AREDS/AREDS2 test sets (601 participants), using the Brier score (95% confidence interval).

Fig. 4. A screenshot of our research prototype system for AMD risk prediction.

a Screenshot of late AMD risk prediction. 1, Upload bilateral color fundus photographs. 2, Based on the uploaded images, the following information is automatically generated separately for each eye: drusen size status, pigmentary abnormality presence/absence, late AMD presence/absence, and the Simplified Severity Scale score. 3, Enter the demographic and (if available) genotype information, and the time point for prediction. 4, The probability of progression to late AMD (in either eye) is automatically calculated, along with separate probabilities of geographic atrophy and neovascular AMD. b Four selected color fundus photographs with highlighted areas used by the deep learning classification network (DeepSeeNet). Saliency maps were used to represent the visually dominant location (drusen or pigmentary changes) in the image by back-projecting the last layer of neural network.