Education on Machine Learning-Derived OCT Biomarkers Improves Clinician Performance for Prediction of Geographic Atrophy Progression

Abstract

Purpose : As therapies for geographic atrophy (GA) become available, ophthalmologists will be challenged to predict which eyes are at greatest risk for progression and identify those patients that would benefit most from therapy. Machine learning (ML) techniques have recently proven valuable for identifying specific OCT features that predict GA progression. These technologies are not widely available, and improving a clinician’s ability to predict future GA progression would be highly valuable. In this analysis, we evaluate the utility training a clinician with ML features as a means of improving clinician identification of patients who will progress to subfoveal GA.

Methods : In this analysis, a grader who was a senior ophthalmology resident reviewed OCT macular cubes of 117 eyes of 117 patients in an IRB-approved retrospective image analysis study with dry AMD. The grader predicted whether each eye progressed to subfoveal GA at 3 and 5 years. The grader received no initial feedback regarding performance. After the initial review, the grader underwent training on ML-derived OCT features that have been identified to portend progression to subfoveal GA using representative images and ML outputs. The grader then repeated the review of the same image set without any additional data. Performance for the 2 grading sessions were compared.

Results : In the initial review of the images, the grader’s sensitivity for identification of eyes progressing to GA at 3 and 5 years was 62.5% and 77.8%, respectively. Following training, the grader’s sensitivity improved dramatically to 90.9% and 94.4% at 3 and 5 years, respectively. Specificity improved, though to a lesser extent, from pre-training levels of 92.08% and 71.72% to 94.23% and 75.75% at 3 and 5 years respectively. Following training, negative predictive value improved to 99% and 98.7% at 3 and 5 years, respectively.

Conclusions : These preliminary results demonstrate the potential utility of using advanced technology-based imaging feature analysis as a basis for training clinicians to become more discriminating in their predictive ability to identify eyes at greatest risk for the development of GA. Future work will focus on validation and refinement of the training system.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.