Defining disease endophenotypes in neovascular AMD by unsupervised machine learning of large-scale OCT data

Abstract

Purpose : The morphologic features of choroidal neovasculariziation (CNV) in patients with neovascular AMD show substantial inter-individual variability. However, existing systems to sub-classify CNV rely on predefined hypotheses, which may represent a bias and may hinder the discovery of novel biomarkers. Our aim was to use an unbiased representation of spectral-domain OCT images to sub-classify AMD and to identify endophenotypes purely based on image information.

Methods : As a representative sample of treatment-naive neovascular AMD eyes, we used baseline data of patients enrolled in the HARBOR trial (n=1096). We propose to learn OCT volume representations in a completely unsupervised manner. A multi-level embedding is used to derive a patient level embedding despite the disparate proportion of very high-dimensional data and few patients in the cohort. In the first level, we use an auto-encoder to obtain low dimensional embeddings of A-scans. In the second level, the resulting coefficients for each volume are embedded with a second auto-encoder model to obtain volume level embedding coefficients. Finally, these low dimensional volume embeddings are clustered using K-means, where the number of clusters is selected automatically using the Davies-Bouldin (DB) index.

Results : The A-scan embedding captured several known disease characteristics such as intraretinal and subretinal fluid, pigment epithelial detachment and subretinal hyperreflective material, in addition to other hitherto unknown biomarkers represented in en-face maps (Fig 1). The volume embedding enabled a compact and well separated clustering of patients indicated by a low DB-index. Furthermore, clusters were visually plausible (Fig 1). The detected endophenotypes showed clinically meaningful separation, e.g. in the distribution of lesion types (Tab 1). Of 34 tested SNPs, 1 gene showed a statistically significant association with the clusters.

Conclusions : Unsupervised machine learning based purely on image data enabled a low-dimensional representation of OCT volumes. A subsequent clustering allowed identification of clinically meaningful novel disease endophenotypes in neovascular AMD.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.