Learning a geometric deep representation to classify Parkinson smooth pursuit patterns

Parkinson’s disease (PD) is characterized by the degeneration of dopaminergic neurotransmitters, leading to motor disturbances that are subtle in the prodromal stages but become more pronounced as the disease progresses. These disturbances report variations regarding manifestation scale and patient phenotyping. Currently, Smooth Pursuit Eye Movement (SPEM) analysis has been suggested to be a potential biomarker for PD. However, conventional recording SPEM methods involve intrusive procedures, specialized protocols, and mainly provide information based on a single global displacement trajectory. We hypothesize that SPEM patterns encompass a diverse range of movement, characterized by intricate spatio-temporal relationships, which may be related to PD, even at early stages. This work introduces a novel end-to-end deep learning representation model that encodes spatio-temporal SPEM patterns and captures geometric second-order relationships to differentiate between PD and control subjects. The geometric learning scheme considers a Riemannian manifold structure from the spatio-temporal deep activations resulting from 3D volumetric convolutions of a set of video recordings. Following a non-intrusive video-based recording protocol, the proposed approach achieved an excellent AUC-ROC score across several SPEM task configurations, with a total of 22 subjects (11 control and 11 PD patients) participating in the study. The geometrical learning encodes spatio-temporal SPEM relationships to support the classification between PD patients and control subjects.