Abstract
Purpose
To compare AI-based biological-sex classification in control eyes and clinically unaffected fellow eyes of patients with central serous chorioretinopathy (CSC) or polypoidal choroidal vasculopathy (PCV).Methods
This retrospective exploratory study included one eye per participant (82 controls, 48 CSC fellow eyes, and 61 PCV fellow eyes). A publicly released Japan Ocular Imaging Registry EfficientNet-B7 model classified sex from JPEG color fundus photographs using a prespecified female-probability threshold of 0.5. Accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and Cohen's kappa were estimated. Logistic regression was adjusted for age and biological sex and was kept parsimonious because only 32 misclassification events occurred; Firth penalized logistic regression was performed as a sensitivity analysis.Results
Accuracy was 89.0% (95% confidence interval [CI], 80.4-94.1%) in controls, 87.5% (75.3-94.1%) in CSC fellow eyes, and 72.1% (59.8-81.8%) in PCV fellow eyes (overall p = 0.018). Accuracy was lower for PCV fellow eyes than for controls (Bonferroni-adjusted p = 0.045); the CSC-PCV comparison was not significant (adjusted p = 0.183). AUCs were 0.935, 0.879, and 0.715, respectively. PCV was associated with misclassification after adjustment for age and biological sex (odds ratio [OR], 3.13; 95% CI, 1.23-8.00; p = 0.017), with a similar Firth estimate (OR, 2.97; 95% CI, 1.22-7.69; p = 0.017).Conclusion
AI-based sex classification was less accurate in clinically unaffected PCV fellow eyes than in controls, whereas performance in CSC fellow eyes was comparable to controls. Because the groups were unmatched, only 32 misclassification events occurred, and imaging-related confounding could not be fully assessed, these findings are hypothesis-generating.