Diagnostic Accuracy of a Federated Learning Algorithm for Proliferative Diabetic Retinopathy Detection: A Multicenter Indonesian Study

Abstract

Introduction: Proliferative diabetic retinopathy (PDR) remains a leading cause of preventable blindness in resource-limited settings. Federated learning (FL) enables collaborative artificial intelligence (AI) model training without sharing patient data. This study evaluated the diagnostic accuracy of an FL-based algorithm for PDR detection across three Indonesian ophthalmology centers.

Methods: This multicenter, prospective, diagnostic accuracy study enrolled 512 eyes from 289 patients with type 2 diabetes at three private hospital ophthalmology clinics in Palembang, Jakarta, and Surabaya, Indonesia (January 2023–December 2024). All eyes underwent standardized fundus photography, spectral-domain optical coherence tomography, and comprehensive ophthalmic examination. The FL-based deep learning algorithm was evaluated against two independent retinal specialists using the International Clinical Diabetic Retinopathy classification. Primary outcomes were sensitivity, specificity, and area under the receiver operating characteristic curve (AUC).

Results: The FL-based AI achieved a sensitivity of 94.6% (95% CI 81.8–99.3), specificity of 91.2% (95% CI 88.2–93.6), and AUC of 0.962 (95% CI 0.943–0.981) for PDR detection. Agreement with retinal specialists was substantial (κ = 0.87). Performance was consistent across centers (AUC 0.955–0.968; p = 0.841). Media opacity was the strongest predictor of misclassification (OR 3.42; 95% CI 1.87–6.25; p < 0.001).

Conclusion: The FL-based AI demonstrated high diagnostic accuracy for PDR detection comparable to retinal specialists across multiple Indonesian centers. This privacy-preserving approach may facilitate scalable diabetic retinopathy screening in resource-limited ophthalmology settings.