Neurologic Dysfunction Assessment in Parkinson Disease Based on Fundus Photographs Using Deep Learning

Area of Science:

• Ophthalmology
• Neurology
• Artificial Intelligence

Background:

• Assessing neurologic dysfunction in Parkinson disease (PD) typically requires complex and time-consuming neurologic tests.
• There is a need for accessible, reliable, noninvasive indicators of neurologic dysfunction in PD patients.

Purpose of the Study:

• To investigate the utility of fundus photography as a noninvasive tool for evaluating neurologic dysfunction in PD.
• To develop a deep learning algorithm capable of predicting Hoehn and Yahr (H-Y) scale and Unified Parkinson's Disease Rating Scale part III (UPDRS-III) scores using fundus images.

Main Methods:

• A prospective decision analytical model was employed at a single tertiary-care hospital.
• Fundus photographs from 615 participants (266 with PD, 349 with non-PD motor abnormalities) were analyzed.
• A convolutional neural network was developed to predict H-Y and UPDRS-III scores based on fundus photography and demographic data.

Main Results:

• Internal validation showed sensitivities of 83.23% (H-Y scale) and 82.61% (UPDRS-III), with specificities of 66.81% (H-Y scale) and 65.75% (UPDRS-III).
• External validation yielded sensitivity and specificity of 70.73% and 66.66%, respectively.
• The overall accuracy was 70.45% with an AUROC of 0.67.

Conclusions:

• Deep learning analysis of fundus photography provides insights into neurologic dysfunction in PD patients.
• This method demonstrates a potential association between retinal and brain pathology in PD.
• Further research is necessary to broaden the clinical applicability of this algorithm for Parkinson disease assessment.