Identifying de novo Parkinson's disease with optical coherence tomography of the retina : A machine learning classification approach

Objective To identify and classify de novo Parkinson’s disease patients compared to healthy controls (HC). Background PD patients experience visual symptoms and retinal degeneration. Studies using spectral-domain optical coherence tomography (SD-OCT) have shown retinal thinning in PD, even at the beginning of disease. This study investigated the utility of these retinal changes in de novo Ldopa-naive PD patients, to evaluate if the profile of retinal thinning could serve as a classification biomarker. Methods SD-OCT data were collected in de novo Ldopa-naive PD patients at the University Medical Center Groningen. 5x5 mm macular scans of right and left eyes were made. These were compared to age-matched HC scans. Good quality scans (≥4) were segmented by Iowa Reference Algorithms [1]; each retina was segmented into 10 individual cell layers. Results 121 PD, 110 HC were included. A random forest classification of all cell layers, across both eyes was run. Data was split into 102 training, 26 validation and 31 testing. Total test accuracy was 0.74, Out of the box accuracy 0.64. True positive rate: area under the curve receiver operating characteristic (AUROC) of 0.82, to classify PD compared to HC. Conclusions Retinal cell layer changes could play an important role in a model of classifying PD; presenting with significant differences in medication naive, newly diagnosed patients, being able to provide a high true positive classification, with automatically segmented retinal cell layer data, from straightforward SD-OCT retinal imaging.