Abstract
Purpose :
To investigate the impact of photoreceptor (PR) and retinal epithelium (RPE) loss as imaging biomarkers in geographic atrophy (GA) on individual predictions of fast progressors using AI in a challenging real-world cohort.
Methods :
Real-world data was obtained from patients with GA visiting the out-patient service at the Vienna General Hospital from 2007 to 2018. Spectralis OCT (Heidelberg Engineering, Heidelberg, Germany) scans of patients with more than 2 visits and observation of 6 months after the first visit were included, resulting in 60 eyes with 793 volumes. PR thickness and RPE loss were automatically segmented using an MDR-approved tool (GA monitor, RetInSight, Austria). To identify fast progressors, we fit a linear regression on the measured square root RPE-loss (SQRT-RPEL) area for each eye using a longitudinal mixed effects model with nested eye and patient random factors. The slope of the fitted line reflects the growth of the lesion. To distinguish between fast and slow progressors the median slope was used as a threshold. The total RPEL and PR loss area (PRL-area), the ratio of PRL/RPEL area (Ratio), the ratio of its SQRT areas (SQRT-Ratio), and the mean thickness in the junctional zone of the RPEL border up to 800 µm distance (PRThick) were computed for baseline, month 6 and the difference between these two visits.
Random forests classifiers were trained on subsets of the features in a leave-one-out cross-validation setting. Classification performance measures were computed to determine the importance of feature combinations to identify fast and slow progressors.
Results :
Adding PRL–area and PRThick features on baseline scans only improves sensitivity/specificity in identifying fast progressors from 0.50/0.57 to 0.77/0.63. Including the 6-month follow-up visit, fast progressors could be identified using RPEL-area with a sensitivity/specificity of 0.67/0.70. When adding PRL-area and SQRT-Ratio biomarkers levels of 0.80/0.70 were achieved.
Conclusions :
AI-derived PR measurements allow to identify patients at risk of fast disease progression from the first visit onwards. This is of relevance in patient selection for forthcoming clinical trials as well as in clinical routine, supporting clinicians in treatment decision and prognosis assessment.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.