Real-World External Validation of a Clinically Implemented Artificial Intelligence Index for Visual Field Severity from Optical Coherence Tomography
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Purpose
To externally evaluate the unitless optical coherence tomography–derived AI-CO index for association with Humphrey mean deviation (MD) and discrimination of MD-defined visual field severity.
Methods
This retrospective study included 819 eyes of 424 patients receiving glaucoma care, with AI-CO analysis and Humphrey 24-2 SITA Faster or 30-2 SITA Fast testing within 365 days. The unitless output labeled “peripheral” was termed AI-CO all. Primary analyses assessed Spearman rank correlation and discrimination at MD thresholds of −2, −6, and −12 dB. Additional analyses assessed adjusted associations and one-eye-per-patient and program-specific robustness. Exploratory analyses assessed nonlinearity and cutoffs with patient-level cross-validation.
Results
AI-CO all correlated with Humphrey MD (Spearman ρ=0.785; patient-cluster 95% confidence interval [CI], 0.750–0.818). Median AI-CO all values decreased from −2.10 in the normal range to −4.40, −10.15, and −17.35 in mild, moderate, and severe categories, respectively (ordered trend P<0.001). Areas under the receiver operating characteristic curve were 0.870 (95% CI, 0.842–0.895), 0.919 (0.894–0.939), and 0.943 (0.920–0.963) at the three thresholds. The adjusted spline detected modest nonlinearity with little gain in explained variance over a linear model (R 2 , 0.735 to 0.738). Associations remained after covariate adjustment, and findings were similar in the one-eye-per-patient analysis and across visual field programs.
Conclusions
AI-CO all showed a rank correlation with Humphrey MD and differentiated eyes across specified MD thresholds. It should be interpreted as a severity-related index that complements, rather than replaces, measured perimetry. Exploratory AI-CO all cutoffs require independent validation.
Translational Relevance
In routine glaucoma care, AI-CO all showed a rank correlation with Humphrey MD and discriminated MD-defined severity.
