Choroidal neovascularization classification system based on machine learning to distinguish pachychoroid neovasculopathy from age-related macular degeneration.

Abstract

Purpose : To automatically categorize choroidal neovascularization (CNV) patients based on their clinical presentation, and to construct a classification model.

Methods : A total of treatment naïve 537 patients with unilateral CNV were recruited in this prospective study. Sixty-one baseline characteristics were converted to 60 principal components after standardization. By applying unsupervised machine learning algorithm, k-means method, to all 60 principal components, patients were categorized into the optimal number of clusters determined by gap statistics analysis. Based on the result of clustering, CNV classification model was constructed using 13 parameters. For this, patients were randomly divided into training (n = 402) and validation (n = 135) data sets. We developed CNV scoring system based on the model, and assessed the association between the score and visual outcome after fixed regimen intravitreal aflibercept injections.

Results : K-means clustering method divided CNV patients into 2 clusters (n =248 and 289, respectively), which was the optimal number of clusters according to gap statistics analysis. Patients in cluster 1 (n=248, 46.2%) showed significantly thicker choroid, higher prevalence of choroidal vascular hyperpermeability, dilated choroidal vessel and reduced fundus tessellation, and less soft confluent drusen than patients in cluster 2 (P < 0.001, respectively), so that we speculate CNV in cluster 1 correspond to pachychoroid neovascularization (PNV), and those in cluster 2 correspond to conventional age-related macular degeneration (AMD). CNV classification model showed high performance (area under curve = 0.98) to distinguish PNV from conventional AMD. The developed CNV score, which represent the likelihood of PNV, was significantly associated with the improvement of visual acuity at 3 months (P = 0.0033), but not at 12 months (P = 0.12).

Conclusions : Based on the clinical characteristics, CNV were automatically categorized to 2 clusters, which was optimal. PNV would represent 46.2% of Japanese AMD. Our scoring system would be beneficial for clinicians to differentiate PNV from conventional AMD, and will lead to personalized treatment of patients with CNV. After aflibercept injections, PNV may show rapid improvement in visual acuity than conventional AMD.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.