Integrating Machine Learning and Traditional Survival Analysis to Identify Key Predictors of Foveal Involvement in Geographic Atrophy.

Abstract

The purpose of this study was to investigate the incidence of foveal involvement in geographic atrophy (GA) secondary to age-related macular degeneration (AMD), using machine learning to assess the importance of risk factors. Retrospective, longitudinal cohort study. Patients diagnosed with foveal-sparing GA, having GA size ≥ 0.049mm² and follow-up ≥ 6months, were included. Baseline GA area, distance from the fovea, and perilesional patterns were measured using fundus autofluorescence. Optical coherence tomography assessed foveal involvement, structural biomarkers, and outer retinal layers thickness. Onset of foveal involvement was recorded. Foveal survival rates were estimated using Kaplan-Meier curves. Hazard ratios (HRs) were assessed with mixed model Cox regression. Variable Importance (VIMP) was ranked with Random Survival Forests (RSF), with higher scores indicating greater predictive significance. One hundred sixty-seven eyes (115 patients, average age = 75.8 ± 9.47years) with mean follow-up of 50 ± 29months, were included in this study. Median foveal survival time was 45months (95% confidence interval [CI] = 38-55). Incidences of foveal involvement were 26% at 24months and 67% at 60months. Risk factors were GA proximity to the fovea (HR = 0.97 per 10-µm increase, 95% CI = 0.96-0.98), worse baseline visual acuity (HR = 1.37 per 0.1 LogMAR increase, 95% CI = 1.21-1.53), and thinner outer nuclear layer (HR = 0.59 per 10-µm increase, 95% CI = 0.46-0.74). RSF analysis confirmed these as main predictors (VIMP = 16.7, P = 0.002; VIMP = 6.2, P = 0.003; and VIMP = 3.4, P = 0.01). Lesser baseline GA area (HR = 1.09 per 1-mm2 increase, 95% CI = 1.01-1.16) and presence of a double layer sign (HR = 0.42, 95% CI = 0.20-0.88) were protective but less influential. This study identifies anatomic and functional factors impacting the risk of foveal involvement in GA. These findings may help identify at-risk patients, enabling tailored preventive strategies.