Abstract
About this talk
Purpose: To identify magnification-corrected OCTA-derived microvascular phenotypes in pediatric and adolescent myopia using participant-aware unsupervised machine learning, and to compare corrected FAZ and vessel-density metrics as severity biomarkers.
Methods: Only baseline eyes clinically marked IN were included. Zeiss/Cirrus OCTA vessel-length density, FAZ area and FAZ perimeter were corrected using the Littmann–Bennett scaling factor derived from eye-specific axial length. Six corrected OCTA variables were standardized in a participantlevel discovery set, reduced by PCA and clustered by K-means. IA (machine learning) was assessed using discovery/validation silhouette and participant-subsampling ARI. Fellow eyes were kept together in all splits. Corrected FAZ, corrected vessel density, all OCTA, AL and AL+OCTA were compared for SE ≤ −3 D using participant-held-out ROC analysis and repeated stratified group cross-validation.
Results: The cohort comprised 149 eyes from 75 participants; 137 eyes from 71 participants had all six corrected OCTA features. A stable three-phenotype solution separated a high-density/compact-FAZ profile, an FAZ-expanded/lower-central-VD profile, and a diffuse low-density profile with comparatively compact FAZ. Bilateral phenotype agreement was 78.8%. Corrected inner and full VD, but not corrected FAZ area, were associated with SE in patient-correlated models. In held-out participants, AUCs were 0.320 for corrected FAZ area, 0.899 for corrected VD, 0.924 for all OCTA, and 0.964 for AL. Repeated grouped cross-validation confirmed weak FAZ-alone performance.
Conclusions: Pediatric myopia shows multivariable OCTA heterogeneity not captured by FAZ alone. Magnification correction materially affects both OCTA values and unsupervised phenotype assignment. These data support microvascular phenotyping as a hypothesis-generating framework requiring longitudinal and external validation.
Speaker
About the speaker
