MEG-Based Detection of Voluntary Eye Fixations Used to Control a Computer.

Anastasia O Ovchinnikova,Anatoly N Vasilyev,Sergei L Shishkin,Ivan P Zubarev,Bogdan L Kozyrskiy

doi:10.3389/fnins.2021.619591

Abstract

Gaze-based input is an efficient way of hand-free human-computer interaction. However, it suffers from the inability of gaze-based interfaces to discriminate voluntary and spontaneous gaze behaviors, which are overtly similar. Here, we demonstrate that voluntary eye fixations can be discriminated from spontaneous ones using short segments of magnetoencephalography (MEG) data measured immediately after the fixation onset. Recently proposed convolutional neural networks (CNNs), linear finite impulse response filters CNN (LF-CNN) and vector autoregressive CNN (VAR-CNN), were applied for binary classification of the MEG signals related to spontaneous and voluntary eye fixations collected in healthy participants (n = 25) who performed a game-like task by fixating on targets voluntarily for 500 ms or longer. Voluntary fixations were identified as those followed by a fixation in a special confirmatory area. Spontaneous vs. voluntary fixation-related single-trial 700 ms MEG segments were non-randomly classified in the majority of participants, with the group average cross-validated ROC AUC of 0.66 ± 0.07 for LF-CNN and 0.67 ± 0.07 for VAR-CNN (M ± SD). When the time interval, from which the MEG data were taken, was extended beyond the onset of the visual feedback, the group average classification performance increased up to 0.91. Analysis of spatial patterns contributing to classification did not reveal signs of significant eye movement impact on the classification results. We conclude that the classification of MEG signals has a certain potential to support gaze-based interfaces by avoiding false responses to spontaneous eye fixations on a single-trial basis. Current results for intention detection prior to gaze-based interface’s feedback, however, are not sufficient for online single-trial eye fixation classification using MEG data alone, and further work is needed to find out if it could be used in practical applications.

Highlights

Brain-computer interfaces (BCIs) are a promising tool that could augment human-computer interaction for patients with motor disabilities and even for healthy users (Allison et al, 2007; Nijholt et al, 2008; Blankertz et al, 2016; Cinel et al, 2019)
We attempted to determine if short single-trial segments of MEG data related to voluntary and spontaneous eye fixations can help to distinguish between such types of fixations
The adaptive convolutional neural networks (CNNs), linear finite impulse response filters CNN (LF-CNN) and vector autoregressive CNN (VAR-CNN), developed recently for MEG data classification (Zubarev et al, 2019), were applied for binary classification of the MEG signals corresponding to spontaneous and voluntary eye fixations collected in participants who used voluntary fixations with 500 ms dwell time threshold to play a game

Summary

Introduction

Brain-computer interfaces (BCIs) are a promising tool that could augment human-computer interaction for patients with motor disabilities and even for healthy users (Allison et al, 2007; Nijholt et al, 2008; Blankertz et al, 2016; Cinel et al, 2019). There is a trade-off between fluent interaction and error rate: while long dwell time thresholds make the interaction tiresome, short thresholds provide a notably effortless interaction and lead to frequent misclassification of spontaneous dwells as intended ones (Jacob, 1990) Such false positives are remarkably difficult to avoid because eye movements serve primarily for vision and escape conscious control (Jacob, 1990). To solve this problem, Ihme and Zander (2011) and Protzak et al (2013) proposed to use a passive BCI (Zander and Kothe, 2011), which could detect the expectationrelated brain activity measured by the EEG in eye fixations. In our previous work, using EEG, we applied this approach to a realistic gaze interaction model implemented as a gaze-controlled game (Shishkin et al, 2016; Nuzhdin et al, 2017)

Objectives

Methods

Results

Discussion

Conclusion