Deep learning-based normative database of anterior chamber dimensions for angle closure assessment: the Singapore Chinese Eye Study.

Abstract

The lack of context for anterior segment optical coherence tomography (ASOCT) measurements impedes its clinical utility. We established the normative distribution of anterior chamber depth (ACD), area (ACA) and width (ACW) and lens vault (LV), and applied percentile cut-offs to detect primary angle closure disease (PACD; ≥180° posterior trabecular meshwork occluded). We included subjects from the Singapore Chinese Eye Study with ASOCT scans. Eyes with ocular surgery or laser procedures, and ocular trauma were excluded. A deep-learning algorithm was used to obtain Visante ASOCT (Carl Zeiss Meditec, USA) measurements. Normative distribution was established using 80% of eyes with open angles. Multivariable logistic regression was performed on 80% open and 80% angle closure eyes. Diagnostic performance was evaluated using 20% open and 20% angle closure eyes. We included 2157 eyes (1853 open angles; 304 angle closure) for analysis. ACD, ACA and ACW decreased with age and were smaller in females, and vice versa for LV (all p<0.022). ACD 20th percentile and LV 85th percentile had a balanced accuracy of 84.4% and 84.2% in detecting PACD, respectively. When combined, ACD 20th and LV 85th percentile had 88.68% sensitivity and 88.85% specificity in detecting PACD as compared with a multivariable regression model (ACA, angle opening distance, LV, iris area) with 88.33% sensitivity and 83.75% specificity. Anterior chamber parameters varied with age and gender. The ACD 20th and LV 85th percentile values may be used in silos or in combination to detect PACD in the absence of more sophisticated classification algorithms.