Artificial Intelligence to Stratify Severity of Age-Related Macular Degeneration (AMD) and Predict Risk of Progression to Late AMD
Age-related macular degeneration (AMD) is the leading cause of vision loss in those older than age 50 years in the developed world.1–5 The number of people with AMD is expected to increase 1.5-fold over 10 years because of our aging population, hypertension, and other causes.6,7 At present, there is no treatment for late dry Age-related macular degeneration (geographic atrophy).8,9 Although treatment with antivascular endothelial growth factor is often effective in maintaining or improving vision in the neovascular form of advanced AMD (i.e., wet AMD), it does not provide a cure.
It is also often too late to mediate the issue by the time a person visits an ophthalmologist as the treatment cannot regenerate the vision.10,11 Further, such treatments are costly (typically costs about $9000–$65,000 for one eye, depending on which drug is used, for a 2-year course of treatment 12) and may be associated with significant cardiovascular risks 13 or even development of GA.14 Although the total (direct and indirect) cost of AMD is $225 billion per year15 and is expected to increase (AMD cases will be increased2), the indirect cost is even greater due to injury, depression, and social dependency resulting from blindness.16
The Age-Related Eye Disease Study (AREDS), the largest Age-related macular degeneration (AMD) study, showed that specific antioxidants and vitamin supplementation (called AREDS supplements) reduce the risk of progression from intermediate-stage AMD to late AMD that can allow for preventive strategies.17 Approximately 7.6% of the United States population over the age of 60 is estimated to have advanced or intermediate AMD.18 Recent publications looking at the 10-year experiences of appropriately selected patients taking the AREDS formulation demonstrates that it is effective at slowing disease and improving visual acuity in approximately 25% of patients.19 For this reason, identification of people at risk for late AMD is very important because it could enable timely treatment such as photobiomodulation20 and laser intervention21) and AREDS supplements. Recent studies21 showed that sub-threshold laser (or sub-threshold nanosecond laser) reduces the progression of intermediate AMD except in the case of reticular pseudo-drusen (RPD), in which case it worsens.
Motivated by this need, we reviewed the early screening of AMD and the prediction of AMD progression, which are a necessity to prevent late AMD. We found that:
• Systems have been built from existing datasets for automatic Age-related macular degeneration Artificial Intelligence screening/prediction, but none are ready for clinical deployment
• No method includes RPD.22 RPD double the risk of progression to advanced wet AMD over soft drusen alone.22,23
• No method has been proposed for telemedicine-based automated AMD screening in remote/underserved areas.
Our review found prediction models24,25 based on manual evaluations of drusen and pigment abnormalities that achieved 75.6% accuracy for 10-year-time (in contrast, our fully automated prediction model herein achieved 86.36% accuracy). AREDS report 826 showed on a population basis that for subjects aged 55 to 80 years followed 6.3 years, treatment with antioxidants plus zinc yielded a significant odds reduction for the development of advanced AMD compared with placebo. Genetic, ocular variables (manual analysis of fundus image), and sociodemographic parameter-based prediction of late AMD is reported in,27,28 and recently improved with additional genetic modeling. A number of AMD screening methods have been reported elsewhere,29–32 which can only determine the disease status, not predict late Age-related macular degeneration AMD. For example, Grassmann et al.31 reported an ensemble deep learning-based classifier of 12 different AREDS categories based on pathology, but not a predictor. We have first proposed a fully automated late AMD prediction model, which was presented at ARVO 2018.33 Recently, Burlina et al. proposed a deep learning (DL)-based model34 for 5-year late Age-related macular degeneration AMD progression but did not demonstrate the late dry and late wet Age-related macular degeneration prediction. However, in Burlina et al., one DL model essentially performs image classification by the AREDS nine-step severity scale, as in Grassman et al., and then relies on the published AREDS probabilities for progression at 5 years, rather than AI, to calculate progression risks. An alternate DL model, with regression directly from the image to risk prediction, as we propose here, had poorer overall performance than those that rely on the AREDS statistics. Our model is more complex and finely tuned than any of those, exploiting both DL for classification and machine learning for prediction as well as other retinal and demographic factors. In addition, we include in our training data abrupt transitions (early to late AMD in 1–2 years), and also predict late dry and wet AMD, which is unique.
Here, we propose the first color fundus photo-based noninvasive screening and prediction model for late Age-related macular degeneration AMD for the 1- or 2-year incident with dry and wet form categorization. It is novel in many respects: different input sizes for neural network architectures for learning scale variant and invariant image features; a logistic model tree35 for building a final classifier after assembling different deep learning models, which is a new approach in retinal image classification also proposed by Grassmann et al; a single value risk of conversion produced from the 12-point Age-related macular degeneration AMD severity scale36 utilizing deep convolution neural networks.
The proposed noninvasive technology will identify higher volumes of at-risk patients and determine whether an individual, including early Age-related macular degeneration AMD subjects, will develop late AMD in 1 to 2 years and should be referred to an ophthalmologist (Schematic for the overall screening and prediction of late AMD, Figure 1). AREDS Report 1737 mentioned, and we confirmed, that in AREDS, 36 subjects converted from early to late AMD within a year, and 50 within 2 years. Using the prediction score, the ophthalmologist can perform further testing and/or have higher confidence about immediate treatment (e.g., photobiomodulation20 or laser intervention 21) or advise more frequent follow-up visits.