Using SD-OCT Biomarkers and Big Data Analyses to Identify Eyes at Risk for AMD Progression
New imaging algorithms and a statistical risk assessment model may be beneficial for earlier identification and treatment of eyes with AMD.
Understanding risk factors for progression of age-related macular degeneration (AMD) has implications for how frequently patients may be followed and when to initiate treatment. The concomitant presentation of dry AMD and bilateral large drusen may increase the risk for progression to advanced AMD, so patients with these clinical characteristics may warrant closer follow-up. However, if patients with intermediate dry AMD are at low risk for progression, then it is likely that they can be seen less often. Perhaps more important, stratifying the risk of progression may be helpful for clinical decision-making in terms of when to initiate treatment so as to reduce the risk of conversion to the neovascular form of the disease.
Earlier treatment of eyes with newly converted wet AMD results in better visual acuity outcomes.1 How, then, can retina specialists identify eyes most likely to progress?
TRADITIONAL SURVEILLANCE TECHNIQUES
Based on data from AREDS, there is a 5-year progression risk of 12% to 50% in eyes with intermediate dry AMD, depending on whether one or both eyes have pigment change in addition to bilateral large drusen.2 Based on this, the observation strategy most likely to detect wet AMD conversion would involve seeing patients every 3 months for a dilated fundus examination and spectral-domain optical coherence tomography (SD-OCT) imaging. Combined with at-home daily Amsler grid testing, this would allow earlier identification of new wet AMD cases and minimal delay to first anti-VEGF injection.
Unfortunately, this strategy is not without shortcomings. It is costly, both in terms of the health care dollars spent on evaluation and management examinations and imaging CPT codes, and in the time patients spend in doctors’ offices. There is also a maximum number of patients that can be seen in our clinics, and seeing all intermediate dry AMD patients every 3 months would reduce the available slots for non-AMD retina patients with active disease. Further, it is not efficient to see every patient whose eyes may be at high risk for progression.
Adding genetic testing to the clinical examination is one way to boost the power of prediction. The presence of certain high-risk single-nucleotide polymorphisms (SNPs) may indicate eyes at risk of progression and, therefore, patients who require more frequent monitoring.
A SEA OF DATA
SD-OCT scans are safe, noncontact, and fast. It is even possible to obtain a high-quality image through an undilated pupil. Given that retina specialists typically perform SD-OCT scans during most encounters with patients who have AMD, there is a lot of potential data available for review. Each of these scans has more data in the cube than can be manually analyzed during the time of an average clinical encounter. Specifically, a typical SD-OCT scan contains more than 67 million points of data. (To construct a cube scan, the typical OCT platform performs about 512 A-scans per B-scan, with each individual A-scan collecting about 1024 data points; 128 B-scans make the typical cube.) Unfortunately, less than 1% of the data from these scans is used during the visit.
Ideally, we could analyze all of the data in each OCT scan in an automated fashion and use the output from that analysis to make predictions about current and future disease activity. Moreover, the method for doing this type of automatic analysis should be device-agnostic so that we could input OCT data from any manufacturer’s device to get the analysis. Finally, this ideal method for OCT analysis should be able to process OCT scans of poor quality and analyze scans that have intermediate to advanced macular disease.
STARTING FROM SCRATCH
Interestingly, the software of most commercial OCT devices does not take full advantage of all the data that are derived from each OCT scan. The interpretation algorithms used by most platforms measure a few parameters but cannot perform in-depth measurements of multiple retinal imaging features. Thus, if a retina specialist wants to measure more than three or four biomarkers in an OCT scan, then he or she would have to create an algorithm using the raw imaging data.
To account for all scan types and quality and to make “noisy” scans usable for automated analysis, my colleagues and I have designed a unique “denoising” algorithm for SD-OCT images to prepare them for analysis.3 Typically, drusen must be identified after images are cleaned up. Because we were dissatisfied with the commercially available software for identifying drusen, we created an automated segmentation algorithm to extract drusen features from denoised images.4 We also designed the algorithms to work with images from any vendor and have successfully segmented images from devices made by Zeiss, Heidelberg, and Bioptigen.
Although it is assumed that a larger total drusen area and volume in the macula would increase risk for AMD progression, those are only two possible drusen biomarkers that one could consider in a statistical model to predict which eyes with AMD are at risk for progression. Our research team created a model that incorporates 11 separate drusen biomarkers identified on imaging in an AMD prediction model.5 When combined with the change in those biomarkers over time as well as with demographic information, our predictive model assessed 26 total features (Table).
We tested our model on a retrospective dataset from our patient population at the Byers Eye Institute at Stanford University. In all, 330 eyes from 244 patients over a 5-year period were included in the dataset. A total of 2146 SD-OCT scans were denoised and segmented and had imaging biomarkers extracted. We knew that 36 eyes in this dataset progressed to wet AMD over the time period, so we compared the 26 features in a statistical model between the eyes with progression and those without progression and found that the area, volume, height, and reflectivity features of drusen were most important in distinguishing between progressing and nonprogressing cases. The patient’s age and dry/wet status of the fellow eye were also significant. After the most relevant features were identified, a predictive model for AMD progression was created.
One unique characteristic of this model was that it could be used to calculate progression risk for any timespan. Unlike the AREDS and genetic risk models that supply risk predictions over multiyear timespans, our model could predict the risk of progression within 3, 6, or 12 months—which is what retina specialists and patients are probably most interested in knowing. If a retina specialist knew there was a high risk of an intermediate dry AMD eye progressing in 6 months, would he or she want to closely monitor that patient? With our algorithm, we hope to bring this sort of real-time decision support to clinics.
Although the prospect of predicting which eyes with AMD are at risk for progression and using that information to guide patients’ education and create a clinical follow-up plan is exciting, our work is preliminary at this point. Multiple limitations exist, such as the single patient population used to create the model, the retrospective nature of the data, the inconsistent time spacing between OCT scans in our dataset (it was based on a real-world practice dataset, and every patient had different follow-up intervals), and the lack of genetic features in the statistical model.
To address these issues, we are in the process of validating our model by using other retrospective AMD OCT datasets to verify that our predictive features are sound. We are also beginning to incorporate genetic risk factors into our statistical model, which we expect will boost our predictive accuracy. Additionally, we have begun a prospective clinical trial (NCT02422160) at the Byers Eye Institute that will follow eyes with dry AMD longitudinally with regular SD-OCT scans, incorporate demographic and genetic data, and provide real-time risk assessment for patient management. Each time an eye gets an OCT scan, imaging biomarker features will be extracted and integrated into the risk assessment model for that patient, giving progression risk information that can be used at the point of care to decide which patients to follow closely. This is the last step in proving that this type of disease prediction system is possible. n
Theodore Leng, MD, MS, is director of ophthalmic diagnostics at the Byers Eye Institute at Stanford University and is a clinical assistant professor of ophthalmology at the Stanford University School of Medicine, Palo Alto, California. He has no relevant financial relationships to
the material described in this article. Dr. Leng may be
followed at @tedleng and reached at +1-650-498-4264; fax: +1-888-565-2640; or firstname.lastname@example.org.
1. Eldaly MA, Styles C. First versus second eye intravitreal ranibizumab therapy for wet AMD. Retina. 2009;29(3):325-328.
2. Age-Related Eye Disease Study Research Group. A randomized, placebo-controlled, clinical trial of high-dose supplementation with vitamins C, and E, beta carotene, and zinc for age-related macular degeneration and vision loss: AREDS Report No. 8. Arch Ophthalmol. 2001;119(10):1417-1436.
3. Chen Q, de Sisternes L, Leng T, Rubin DL. Application of improved homogeneity similarity based denoising in optical coherence tomography retinal images. [published online ahead of print November 18, 2014] J Digital Imaging. doi:10.1007/s10278-014-9742-8.
4. Chen Q, Leng T, Zheng L, et al. Automated drusen segmentation and quantification in SD-OCT images. Med Image Anal. 2013;17(8):1058-1072.
5. De Sisternes L, Simon N, Tibshirani R, et al. Quantitative SD-OCT imaging biomarkers as indicators of age-related macular degeneration progression. Invest Ophthalmol Vis Sci. 2014;55(11):7093-7109.