The Use of Machine Learning in the Evaluation of VA and Anti-VEGF Efficacy During the First Year of Treatment in nAMD

Abstract

Introduction: Neovascular age related macular degeneration (nAMD) is a sight threatening, ocular condition that can be managed with varying doses of anti-vascular endothelial growth factor (anti-VEGF) drugs and is routinely monitored with optical coherence tomography (OCT) retinal scans. Artificial intelligence (AI) based technologies also now offer automated analysis of such scans making available additional information on features within the scanned area. Purpose: This study aims to use OCT determined information to predict anti-VEGF treatment frequency and visual prognosis in nAMD, potential influence on treatment regimen and the role AI might play in managing nAMD in the future. Methods: This was a retrospective, non-interventional, observational study of patients aged 50 and over diagnosed with nAMD between May 2016 and March 2020. From electronic medical records, measures of visual acuity (VA), demographic information and anti-VEGF dosing for the duration of the management were included. OCT characteristics from the baseline visit and the post loading visits were extracted by automated segmentation and AI-enabled retinal segmentation. These were analysed using AI driven technology to predict outcomes. Results: 327 eyes of 308 individuals were enrolled within the study. It was found that classification modelling differentiating between eyes that required 3 or >3 injections could predict between the classes to an area under the receiver operating characteristic curve (AUC) of 0.63 with ganglion cell layer and drusenoid PED found to be the most informative features. In attempting to sort between eyes that lost or gained VA over 12 months, classification accuracy of AUC 0.88 was achieved with baseline VA deemed the most informative feature. Conclusion: This study evaluated the application of AI based technologies in investigating anti-VEGF dosing and visual outcomes. The results determined the presence of relationships in predicting injection numbers and VA and perhaps gave some further insights into the role AI may play in the future nAMD management.