Eye Discovery · Jul 29, 2026
Transportability of DeepSeeNet for age-related macular degeneration classification in a population-based cohort: the Canadian longitudinal study on aging (CLSA)
By Aljied R, Chaudhary V, Srinivasan S et al.
Objective To adapt and validate DeepSeeNet for binary age-related macular degeneration (AMD) classification from colour fundus photographs in a population-based cohort and to evaluate how residual classification error may influence downstream epidemiological inferences. Methods We conducted a cross-sectional diagnostic accuracy study using baseline retinal images from the Canadian Longitudinal Study on Aging collected between 2011 and 2015. DeepSeeNet was first updated using cohort-specific data from 300 participants and subsequently fine-tuned with an additional 172 participants. The reference standard comprised independent image grading followed by adjudicated reading centre consensus. After model development was completed, the final model was locked and evaluated in a participant-disjoint held-out test set consisting of 200 images from 100 participants. Model performance was assessed against both adjudicated reading centre consensus grades and individual grader assessments. We also performed quantitative bias analysis to evaluate the potential epidemiological consequences of residual outcome misclassification. Results In the held-out test set, the updated DeepSeeNet achieved an accuracy of 0.83 (95% CI, 0.78 to 0.88), a κ value of 0.67 (95% CI, 0.59 to 0.75), a sensitivity of 0.75 (95% CI, 0.68 to 0.82), a specificity of 0.88 (95% CI, 0.83 to 0.93), and an area under the receiver operating characteristic curve of 0.82 (95% CI, 0.77 to 0.87). Individual trained graders achieved accuracies ranging from 0.76 to 0.78, with κ values ranging from 0.58 to 0.60. Quantitative bias analysis demonstrated that non-differential outcome misclassification had little influence on the estimated prevalence of AMD (observed prevalence, 0.34; median bias-adjusted prevalence, 0.33), although it substantially increased the associated uncertainty (95% simulation interval, 0.23 to 0.44). Overall, cohort-specific updating enabled good agreement between DeepSeeNet and the adjudicated reference standard in this community-based population. Conclusions Following cohort-specific updating, DeepSeeNet demonstrated good classification performance in a strictly participant-disjoint evaluation against an adjudicated reading centre reference standard. Although residual misclassification had minimal impact on the estimated prevalence of AMD, it increased uncertainty around the estimate. These findings suggest that epidemiological applications of AI-derived AMD phenotypes should incorporate targeted grader adjudication together with analytical approaches that explicitly account for outcome misclassification.