Real-World External Evaluation of an OCT Image Analysis System (AI-CO) in Relation to Humphrey Visual Field Mean Deviation

Purpose: To externally evaluate the real-world performance of a commercially implemented artificial intelligence system that analyzes OCT image data and provides numerical and grayscale reference outputs, by characterizing the association and numerical relationship between AI-CO output and Humphrey visual field mean deviation (MD) in routine clinical practice. Study Design: Retrospective, single-center observational study. Methods: This study included 819 eyes of 424 patients who underwent AI-CO analysis and Humphrey 24-2F or 30-2 visual field testing within 1 year. The 24-2F and 30-2 groups comprised 343 and 476 eyes, respectively. Because the AI-CO all value, corresponding to the index labeled "peripheral" in the Japanese user interface, is a unitless index, its numerical relationship with Humphrey MD was explored using Pearson correlation, Bland-Altman analysis, and Passing-Bablok regression. Linearity was assessed using the original cumulative sum (CUSUM) procedure described by Passing and Bablok. Results: The AI-CO all value was strongly correlated with Humphrey MD in both the 24-2F (r = 0.836) and 30-2 (r = 0.854) groups (both P < 0.001). In Bland-Altman analysis, the mean difference (AI-CO all minus Humphrey MD) and 95% limits of agreement were -1.23 (-7.23 to 4.77) in the 24-2F group and -1.44 (-7.96 to 5.08) in the 30-2 group. In both groups, the 95% confidence interval for the Passing-Bablok slope included 1, whereas that for the intercept excluded 0. The original CUSUM procedure did not reject linearity in the overall cohort, the 24-2F group, or the 30-2 group (P = 0.086, 0.091, and 0.145, respectively). Conclusions: In this independent real-world external evaluation of a clinically implemented system, AI-CO all was strongly correlated with Humphrey MD, and linearity of their numerical relationship was not rejected. However, its unitless output was lower on average and showed substantial eye-level variability. AI-CO may therefore provide complementary structural-functional information in routine care but should not be interpreted as interchangeable with standard automated perimetry.