Artificial Intelligence-Assisted Community Eye Screening in Primary Healthcare: A Prospective Multicentre Diagnostic Accuracy and Implementation Study

Background: Artificial intelligence (AI)-assisted retinal screening may extend access to eye care in primary healthcare, but prospective evidence on diagnostic performance and implementation under routine community conditions remains limited. Methods We conducted a prospective multicentre diagnostic accuracy and implementation study across 12 urban and rural community eye-screening centres in India from 1 January to 30 June 2025. Adults aged ≥18 years underwent AI-assisted retinal-image analysis followed by masked comprehensive ophthalmic examination. The primary outcome was participant-level diagnostic accuracy of the AI-generated referral classification for the composite reference-standard outcome of any referable ocular disease. Implementation outcomes included image-acquisition success, workflow completion, referral compliance, screening time, and questionnaire-based acceptance and satisfaction. Results Among 1,732 participants, 610 (35.2%) had referable ocular disease. The AI system produced 566 true-positive, 1,017 true-negative, 105 false-positive, and 44 false-negative classifications. Sensitivity was 92.8% (95% CI 90.4%–94.7%), specificity 90.6% (88.8%–92.3%), positive predictive value 84.4% (81.4%–87.0%), negative predictive value 95.9% (94.5%–97.0%), and overall accuracy 91.4% (90.0%–92.7%). The F1 score was 0.884, Cohen’s kappa was 0.816, and the AUC based on the three-level AI risk classification was 0.925 (bootstrap 95% CI 0.912–0.937). Image acquisition succeeded in 1,668 participants (96.3%), workflow completion was 98.7%, and referral compliance was 550/671 (82.0%). Mean community-acceptance, healthcare-provider-satisfaction, and participant-satisfaction scores were 4.56, 4.44, and 4.49, respectively. Conclusions AI-assisted community eye screening showed high sensitivity and good overall diagnostic performance with strong operational feasibility. The false-positive burden, particularly among participants with diabetes, supports continued clinical oversight, image-quality assurance, and subgroup-specific validation before wider health-system adoption.