Objective: To evaluate the discriminative power of a novel saccade waveform reconstruction approach compared to conventional saccadic metrics in differentiating Parkinson’s disease (PD), progressive supranuclear palsy (PSP), healthy controls, and PD subgroups based on disease duration and medication status.

Background: Differentiating PD from PSP is challenging due to symptom overlap, particularly in early stages. Saccades, rapid eye movements between fixation points, are affected by both diseases. Conventional saccadic metrics capture only a fraction of the information present in saccadic waveforms.

Method: We analysed 13,309 saccadic eye movements from 127 participants, including 11 early unmedicated PD patients, 51 early medicated PD patients, 18 advanced PD patients, 12 PSP patients, and 35 age-matched healthy controls. Conventional metrics were compared to a novel waveform reconstruction approach using linear and deep autoencoders to reconstruct individual saccade trajectories, then training a second model to predict a participant’s disease status from the distributions of reconstruction errors of all their saccades. We evaluated six different hierarchical methods, using quicksort leave-pair-out cross-validation to assess classification of saccadic waveforms alone.

Results: Hierarchical models on conventional saccade metrics yielded AUC scores ranging from 0.35 to 0.76, with the highest accuracy in differentiating PSP from healthy controls (AUC: 0.76) and lower performance within PD subgroups and between PSP and advanced PD. Waveform reconstruction approaches significantly outperformed conventional metrics, demonstrating high discriminative power (AUC up to 0.98) across most group comparisons. The models effectively distinguished between participants with different PD stages and medication status, indicating effects of dopaminergic therapy and disease progression on saccade trajectories. A post-hoc analysis on model generalisation  

identified late-stage PD patients when trained on early-stage medicated PD and PSP with 94% accuracy.

Conclusion: The study underscores the precise discriminative utility of waveform reconstruction methods over conventional saccadic metrics to classify between Parkinsonian disorders. This highlights the potential of saccades, an inexpensive, non-invasive, portable biomarker, for precise classification between PD, PD subgroups and PSP.

« Back to 2024 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/hierarchical-machine-learning-classification-of-parkinsonian-disorders-using-saccades/