Artificial intelligence applications in Parkinson's disease via retinal imaging

TL;DR

This paper reviews 35 years of research on using artificial intelligence to analyze retinal images for early detection and risk assessment of Parkinson's disease, highlighting promising models and their performance.

Contribution

It systematically evaluates AI applications in retinal imaging for Parkinson's, identifying key diagnostic tasks and summarizing the best-performing models and their accuracy.

Findings

AI models achieve high accuracy in disease classification

Retinal vessel segmentation models reach over 99% accuracy

AI-based risk prediction shows promising AUC scores

Abstract

Parkinson's disease (PD) is projected to increase substantially due to population aging, making early diagnosis increasingly important, as timely detection may delay progression and reduce long-term complications. Retinal microvasculature has emerged as a promising anatomical biomarker of neurodegeneration, and when combined with artificial intelligence AI, retinal imaging may provide an advanced, noninvasive, and cost-effective screening strategy for PD. This study evaluated the evidence from the past 35 years regarding the capability of AI to detect early PD-related changes in retinal vascular structure. Five electronic databases including PubMed, Web of Science, Scopus, ScienceDirect, and ProQuest were systematically searched from January 1990 to January 2025. In addition, Annals of Neurology and Frontiers in Neuroscience were hand-searched, and the reference lists of included studies were screened for additional eligible publications. Nineteen studies met the inclusion criteria. Three principal diagnostic AI tasks were identified, including disease classification, retinal vessel segmentation, and PD risk stratification. The best-performing models were ShAMBi-LSTM on the Drishti dataset with 97.2 percent accuracy, 99.5 percent precision, 96.9 percent sensitivity, and an F1 score of 0.981 for classification, nnU-Net with 99.7 percent accuracy, 98.7 percent precision, 98.9 percent sensitivity, 99.8 percent specificity, and a Dice score of 98.9 percent for segmentation, and AlexNet for risk prediction with area under the curve values of 0.77, 0.68, and 0.73 across datasets. Overall, application of AI algorithms to retinal vasculature for detecting early signs of PD and predicting disease severity suggests that integration of AI with retinal biomarkers holds substantial potential for earlier and more accurate detection compared with traditional clinical evaluation alone.