A convolutional oculomotor representation to model parkinsonian fixational patterns from magnified videos

Abstract

Oculomotor alterations are a promising biomarker to detect and characterize Parkinson’s disease (PD), even in prodromal stages. Nowadays, however, only global and simplified gaze trajectories are used to approximate the complex interactions between neuromotor commands and ocular muscles. Besides, the acquisition of such signals often requires sophisticated calibration and invasive settings. This work presents a novel imaging biomarker for PD assessment that models ocular fixational movements, recorded with conventional cameras. Firstly, a video acceleration magnification is performed to enhance small relevant fixation patterns on standard gaze video recordings. Hence, from each video are extracted a set of spatio-temporal slices, which thereafter are represented as convolutional feature maps, recovered as the first-layer responses of pre-trained CNN architectures. The feature maps are then efficiently encoded by means of covariance matrices to train a support vector machine and perform the disease classification. From a set of 130 recordings of 13 PD patients and 13 age-matched controls, the proposed approach achieved an average accuracy of 95.4% and an AUC of 0.984, following a leave-one-patient-out cross-validation scheme. The proposed imaging-based descriptor properly captures known disease tremor patterns, since PD classification performance is outstanding when augmented motion frequencies were fixed within tremor-related ranges. These results suggest a successful PD characterization from fixational eye motion patterns using ordinary videos.