New Method of Early RRMS Diagnosis Using OCT-Assessed Structural Retinal Data and Explainable Artificial Intelligence.

Abstract

Purpose: The purpose of this study was to provide the development of a method to classify optical coherence tomography (OCT)-assessed retinal data in the context of automatic diagnosis of early-stage multiple sclerosis (MS) with decision explanation.

Methods: The database used contains recordings from 79 patients with recently diagnosed relapsing-remitting multiple sclerosis (RRMS) and no history of optic neuritis and from 69 age-matched healthy control subjects. Analysis was performed on the thicknesses (average right and left eye value and inter-eye difference) of the macular retinal nerve fiber layer (mRNFL), macular ganglion cell layer (mGCL), macular inner plexiform layer (mIPL), and macular inner retinal complex layer (mIRL), dividing the macular area into six analysis zones. Recursive feature extraction (RFE) and Shapley additive explanations (SHAP) are combined to rank relevant features and select the subset that maximizes classifier (support vector machine [SVM]) performance.

Results: Of the initial 48 features, 20 were identified as maximizing classifier accuracy (n = 0.9257). The SHAP values indicate that average thickness has greater relevance than inter-eye difference, that the mGCL and mRNFL are the most influential structures, and that the peripheral papillomacular bundle and the supero-temporal quadrant are the zones most affected.

Conclusions: This approach improves the success rate of automatic diagnosis of early-stage RRMS and enhances clinical decision making transparency.

Translational relevance: Retinal structure assessment using OCT data could constitute a noninvasive means of diagnosing early-stage MS. This new high-accuracy and high-explainability method of analysis can be implemented in most hospitals and healthcare centers.