Predictive Identification of the Fastest Progressing Geographic Atrophy Lesions based on Deep Learning in the Phase 2 FILLY clinical trial of Pegcetacoplan

Abstract

Purpose : In patients with geographic atrophy (GA) secondary to age-related macular degeneration (AMD), the speed of progression is highly variable. To evaluate the potential of artificial intelligence (AI) for therapeutic dosing frequency, we trained a deep learning-based algorithm to predict the growth of GA lesions and to identify patients with the fastest growing lesions (“fast progressors”).

Methods : OCT scans (512x49x496 voxels, Heidleberg Engineering) of study eyes of patients with GA; that were enrolled in FILLY phase II clinical trial to study the effect of pegcetacoplan (APL-2), and investigational therapy targeting complement C3. GA lesion size at baseline and year 1 was manually annotated on fundus autofluorescence (FAF) by a reading center using RegionFinder and on OCT by an expert reader. Fast progressors were defined as the top 20% of the lesion size growth distribution. A deep learning method was then trained to predict the one year GA growth from a baseline OCT scan. Predictive performance was evaluated using a five-fold cross-validation.

Results : 155 study eyes were included in the analysis that completed the one-year follow-up and had a sufficient image quality to allow for the training and evaluation. Patients from each study arm were considered and fast progressors were identified independent of treatment. Of the top 20% of patients (n =31), 15 were in the sham treatment arm, 11 were treated with APL-2 every other month, and 5 were treated with APL-2 monthly. Atrophy measured on FAF and on OCT showed a concordance of R²=0.9. The evaluation showed that the growth was predicted with a correlation of R=0.57, R²=0.32, and the detection of fast progressors achieved an area under the curve AUC of 0.83 (CI: 0.75-0.90) corresponding to a sensitivity of 0.78 and specificity of 0.70.

Conclusions : Fully automated prediction of GA lesion growth from OCT scans is possible, and it allows identification of fast progressors, who may need more intensive treatment. FAF atrophy measurements were consistent with the OCT-based ones. The results of this pilot study are a promising step toward clinical decision support tools for therapeutic dosing and guidance of patient management on a large scale, once such treatments become available.

This is a 2021 ARVO Annual Meeting abstract.