Post Hoc Analysis of FILLY Trial Using AI

July 2023

Edward Korot, MD
Retina Specialists of Michigan, Grand Rapids, MI
Adjunct Clinical Assistant Professor, Stanford University, Palo Alto, CA


Fu DJ, Glinton S, Lipkova V, et al. Deep-learning automated quantification of longitudinal OCT
scans demonstrates reduced RPE loss rate, preservation of intact macular area and predictive
value of isolated photoreceptor degeneration in geographic atrophy patients receiving C3
inhibition treatment. Br J Ophthalmol. Published online April 24, 2023:bjo-2022-322672.

Geographic atrophy (GA) is a devastating form of age-related macular degeneration (AMD) that
can slowly lead to irreversible vision loss. There is currently one approved treatment for GA,
pegcetacoplan, with several others in the pipeline and many trials ongoing. Pegcetacoplan
targets a component of the complement system (C3), which is thought to cause inflammation
and tissue damage downstream. Fu et al. conducted a post hoc analysis of the FILLY trial to
assess the efficacy of pegcetacoplan for GA using a fully automated deep-learning approach for
analyzing optical coherence tomography (OCT) scans. The researchers used this technique on
OCT images of patients with GA who received pegcetacoplan or a placebo. They extracted
structural data and used the AI to segment images according to the constituent features of RPE
and outer retinal atrophy (RORA) including hypertransmission, photoreceptor degeneration
(PRD), and RPE loss (which were all secondary endpoints). The primary endpoint was the
square root change in GA area defined as complete RORA (cRORA).

In total, 197 eyes were included in the analysis, with 71 in the pegcetacoplan monthly group, 61
in the every other month group and 65 in the sham group. Monthly treated eyes demonstrated
significantly slower progression of cRORA and RPE loss. Every other month treated eyes
showed only significantly slower RPE loss, while not demonstrating significant differences in
other endpoints. As compared to sham, the monthly group had a 45% reduction in cRORA
growth rate in contrast with 27% reduction for the every other month group. Intact macular area
was preserved in monthly treatment compared with sham; however, this was not the case in those receiving treatment every other month.

Using a regression model, the researchers identified two predictive markers of reduced GA
growth at 12 months, PRD in isolation (without the other two constituents of RORA) (coefficient
0.0195, p=0.01), and intact macular area (coefficient 0.00752, p=0.02). These may be predictors
of GA progression, and may serve as clinical endpoints in future studies.
Overall, the results of this study suggest that pegcetacoplan, especially when dosed monthly, is
effective in reducing GA lesion size and preserving macular integrity in patients with GA. The
deep learning algorithm provided accurate and efficient analysis of GA lesion subtypes, which
could be useful in assessing the efficacy of other emerging therapeutics for GA.

One of the strengths of this study is its use of a fully automated deep-learning approach for
analyzing SD-OCT images. This approach allows for efficient and accurate extraction of
structural data from retinal imaging, which is essential for assessing the efficacy of emerging
therapeutics for GA. Given that OCT-based analysis captures the morphology of various retinal
layers, this modality is preferred to fundus autofluorescence-based quantification of GA.
Without deep learning algorithms like these, manual segmentation would be slow, and suffer
from inter-grader variability.

The researchers also incorporated topographical and volumetric analysis, which provided
greater insight into the effect of treatment beyond binary presence or absence of features.

One limitation of this study is its small sample size (197 eyes). This post hoc analysis included
only a subset of patients (those with Imaging on Heidelberg OCT) from the original clinical trial,
which may limit the generalizability of the findings. Additionally, the study did not assess the
long-term safety and efficacy of pegcetacoplan, which will be important to evaluate in future

Conclusion and takeaway points:
Overall, this study provides further evidence of the efficacy of pegcetacoplan for GA and
demonstrates the value of a fully automated deep-learning approach for rapid and repeatable
analysis of SD-OCT images. The identification of two novel clinical endpoints, PRD in isolation
and intact macula, may facilitate more efficient assessment of emerging therapeutics for GA. If
used in clinical trials, AI algorithms such as these may provide improved quantification as
compared to human graders. If used in clinic, algorithms like these may actually provide
quantitative tracking of treatment results, an area in which our current capacities are lacking.
This study brings up several areas of concern for routine clinical use of pegcetacoplan. No
statistically significant differences were found for every other month treatment groups with
regards to intact macula, photoreceptor degeneration, or GA growth rate. This is worrisome, as
most clinicians plan to use the treatment every other month in order to limit treatment burden,
while balancing similar efficacy to monthly treatment as was observed in the pivotal trials.
Furthermore, the novel analysis by ETDRS region demonstrated that the effect of growth rate
reduction was greatest further from the fovea, although this analysis may have been affected by
a baseline foveal involvement rate of 59.7%. This is concerning for real-world clinical practice
treatment patterns, as clinicians are more likely to treat patients with fovea-threatening GA,
where it appears this medication has least effect. The researchers did not find a protective
effect of the drug on the foveal region.
Further research is needed to confirm these findings and to evaluate the long-term safety and
efficacy of pegcetacoplan, especially in routine clinical use, where it will undoubtedly be injected
less than monthly, and may be used primarily for fovea-threatening disease.

CME Link: