Generalizing AI for eye diagnosis

Abstract

The generalizability of healthcare algorithms to different healthcare settings is challenging for two reasons. First, different medical devices used for clinical diagnosis require algorithms to consider possible systematic differences due to hardware settings – signal idiosyncrasies. Second, epidemiological differences of patients across different regions further lead to differences between model-building and actual clinical conditions – patient heterogeneity. In this study, we propose a methodology and develop a convolutional neural network (CNN)-based algorithm to read optical coherence tomography eye scans to diagnose if a patient suffers from diabetic macular edema (DME) or age-related macular degeneration (AMD). We utilized clinical data from different sources and are in the process of testing the accuracy of this diagnosis algorithm against a heterogenous set of patients. Our preliminary results suggest an Area under the ROC Curve (AUC) exceeding 97.98% with a performance equivalent to a human physician assessment.