A digital eye-fixation biomarker using a deep anomaly scheme to classify Parkisonian patterns

TL;DR

The paper addresses the challenge of detecting Parkinson's disease via oculomotor biomarkers by moving from discriminative, data-hungry classifiers to a one-class anomaly detection framework. It proposes a GANomaly-based pipeline that learns Parkinsonian eye-fixation patterns from horizontal/vertical video slices and flags non-Parkinsonian samples as anomalies using a defined anomaly score. On a dataset of 13 PD patients and 13 controls, the method achieves an AUC-ROC of approximately 0.95 and very high sensitivity (≈0.97), with modest specificity (≈0.63), demonstrating strong discrimination with limited data. This approach reduces data requirements and could enable practical, automated PD screening and progression monitoring using video data, potentially accessible with consumer cameras.

Abstract

Oculomotor alterations constitute a promising biomarker to detect and characterize Parkinson's disease (PD), even in prodromal stages. Currently, only global and simplified eye movement trajectories are employed to approximate the complex and hidden kinematic relationships of the oculomotor function. Recent advances on machine learning and video analysis have encouraged novel characterizations of eye movement patterns to quantify PD. These schemes enable the identification of spatiotemporal segments primarily associated with PD. However, they rely on discriminative models that require large training datasets and depend on balanced class distributions. This work introduces a novel video analysis scheme to quantify Parkinsonian eye fixation patterns with an anomaly detection framework. Contrary to classical deep discriminative schemes that learn differences among labeled classes, the proposed approach is focused on one-class learning, avoiding the necessity of a significant amount of data. The proposed approach focuses only on Parkinson's representation, considering any other class sample as an anomaly of the distribution. This approach was evaluated for an ocular fixation task, in a total of 13 control subjects and 13 patients on different stages of the disease. The proposed digital biomarker achieved an average sensitivity and specificity of 0.97 and 0.63, respectively, yielding an AUC-ROC of 0.95. A statistical test shows significant differences (p < 0.05) among predicted classes, evidencing a discrimination between patients and control subjects.

Figures (4)

• Figure 1: Overview of the proposed approach. (a) Transformation of video recordings into horizontal and vertical slices for feature extraction. (b) Anomaly detection framework based on GANomaly, where a generative autoencoder learns Parkinsonian oculomotor fixation patterns, and deviations from this learned representation are classified as outliers. The dimensions $H$, $W$, and $D$ represent the height, width, and number of feature maps at different computational levels.
• Figure 2: ROC curves for the different cross-validation folds of the GANomaly approach. The solid lines correspond to individual folds, while the bold blue line represents the mean ROC curve. The shaded region denotes the variability across folds, and the dashed red line indicates the performance of a random classifier.
• Figure 3: Boxplots comparing normal and anomalous populations across the Encoder, Contextual, and Adversarial subnetworks of the GANomaly model. The significant differences, confirmed by one-way ANOVA, highlight the model's effectiveness in distinguishing between control and PD populations.
• Figure 4: Subtraction of some images before and after entering the network (real and synthetic images respectively). The differences are shown in the domain of frequencies, with the relevant alterations in frequency represented by high contrast.