Objective: To establish the extent to which high temporal resolution is required to differentiate between parkinsonian disease states from eye movements.

Background: Previous studies have demonstrated that high-resolution eye movement recordings can be used in differential diagnosis of Parkinson’s disease and progressive supranuclear palsy. The range of eye tracking hardware has expanded to include consumer devices, but spatial and temporal resolution of these varies.

Method: Retrospective analysis of saccadic eye movement recordings obtained from people with idiopathic Parkinson’s disease (n=11 medication-naïve; n=69 medicated), PSP (n=10), and healthy controls (n=35). 

We used PREDA models [1] to distinguish unmedicated PD patients from other groups. Models were trained using either original (200 Hz) or downsampled (40-167 Hz) data, and tested using downsampled data. To adapt models trained on high-resolution input to handle low-resolution input we compared PCHIP polynomial interpolation with machine learning models (random forest, RF; multilayer perceptron, MLP) trained to interpolate saccadic data. Performance was evaluated by 4-fold cross-validation and the mean area under the receiver operating characteristic (AUROC) calculated.

Results: Upsampling methods performed differently at different sampling rates (Friedman, p < 0.05). RF and MLP performed similarly for eye position (Games-Howell, p > 0.05 at all rates) but RF outperformed MLP for eye velocity above 100 Hz (p < 0.05). RF and MLP outperformed PCHIP at 30 Hz for both position and velocity (p<0.0001). Based on these results, we compared RF and PCHIP with models trained on downsampled data.

Nearly all models performed worse at lower sampling rates, and poorly at 40 Hz (medication-naïve vs. medicated PD: highest AUROC 0.64, vs. HC: highest 0.57) except when differentiating between patients with PD and PSP (highest 0.94 at 40 Hz). Models trained directly on lower-resolution data outperformed those adapted from high-resolution data in all tasks. Despite good reconstruction, RF-adapted models performed poorly at higher sampling rates (PD vs. HC: RF 0.67, original 0.87; naïve vs. medicated PD: RF 0.70, original 0.91 at 200 Hz).

Conclusion: This study shows that the high performance of PREDA models depends on input data with a high temporal resolution. Further studies should assess the potential for performance gains with input at higher sampling frequencies.

References: 1. Patel SB, Bredemeyer OB, FitzGerald JJ, Antoniades CA. Preprint: Hierarchical Machine Learning Classification of Parkinsonian Disorders using Saccadic Eye Movements: A Development and Validation Study. (2024) doi: 10.48550/arXiv.2407.16063

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/resolution-dependence-of-machine-learning-for-differential-diagnosis-of-parkinsonian-eye-movements/