A Novel Pixel-Based Spatial Modeling Framework for Mapping At-Risk Retina in Geographic Atrophy Progression

To develop a Pixel-based At-risk Retina Mapping (PARM) framework quantifying local geographic atrophy (GA) progression in nonexudative age-related macular degeneration and evaluate its predictive performance and statistical efficiency relative to conventional area-based measures.

Secondary analysis of a multicenter randomized controlled trial.

Eyes with GA from the Age-Related Eye Disease Study.

Annual color fundus photographs were manually segmented, registered, and converted into 10-μm pixel grids. The perilesional "at-risk" zone was empirically defined as pixels with ≥5% 1-year GA conversion probability. Within this zone, 1-year GA pixel involvement was modeled using a generalized additive model with random effects for eye and participant. Predictors included pixel-level measures (distances to GA border and foveal center) and 9 eye-level factors. Prediction accuracy was evaluated using the area under the receiver operating characteristic curve (AUC). Simulation-based power analyses compared sample size requirements between the PARM framework and conventional area-based measures for detecting a 30% treatment effect (odds ratio 0.70 for pixel conversion) at 80% power (2-sided α = 0.05).

One-year GA pixel involvement.

We included 239 eyes from 161 participants. Pixels within 500 μm of the GA border had ≥5% 1-year conversion risk, defining the "at-risk" region. Shorter distance to the GA border and greater distance from the foveal center were independently associated with a higher 1-year risk of GA pixel involvement (P < 0.001); 9 eye-level factors did not improve model fit. The pixel-based model achieved an AUC of 0.875 (95% confidence interval [CI], 0.865-0.887), outperforming the eye-level model (AUC = 0.750, 95% CI, 0.740-0.768). Detecting a 30% treatment effect required 30 eyes per arm using the PARM framework, compared with 140 eyes for square-root-transformed area and 310 eyes for total GA area.

Pixel-based spatial modeling of GA progression within a 500-μm perilesional at-risk zone showed stronger predictive performance than conventional eye-level modeling and improved statistical efficiency relative to area-based measures in internal simulations by leveraging pixel-level distances to the GA border and foveal center. These findings support further evaluation of PARM framework for studying GA progression, with external validation required before clinical-trial implementation.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.