Sabotage Detection Using DL Models on EEG Data From a Cognitive-Motor Integration Task

Mahima Chaudhary; Meaghan S Adams; Sumona Mukhopadhyay; Marin Litoiu; Lauren E Sergio

doi:10.3389/fnhum.2021.662875

Sabotage Detection Using DL Models on EEG Data From a Cognitive-Motor Integration Task

Front Hum Neurosci. 2021 Oct 8:15:662875. doi: 10.3389/fnhum.2021.662875. eCollection 2021.

Authors

Mahima Chaudhary¹, Meaghan S Adams^{2

3}, Sumona Mukhopadhyay¹, Marin Litoiu¹, Lauren E Sergio²

Affiliations

¹ Lassonde School of Engineering, York University, Toronto, ON, Canada.
² Faculty of Health, York University, Toronto, ON, Canada.
³ KITE - Toronto Rehabilitation Institute, University Health Network, Toronto, ON, Canada.

Abstract

Objective clinical tools, including cognitive-motor integration (CMI) tasks, have the potential to improve concussion rehabilitation by helping to determine whether or not a concussion has occurred. In order to be useful, however, an individual must put forth their best effort. In this study, we have proposed a novel method to detect the difference in cortical activity between best effort (no-sabotage) and willful under-performance (sabotage) using a deep learning (DL) approach on the electroencephalogram (EEG) signals. The EEG signals from a wearable four-channel headband were acquired during a CMI task. Each participant completed sabotage and no-sabotage conditions in random order. A multi-channel convolutional neural network with long short term memory (CNN-LSTM) model with self-attention has been used to perform the time-series classification into sabotage and no-sabotage, by transforming the time-series into two-dimensional (2D) image-based scalogram representations. This approach allows the inspection of frequency-based, and temporal features of EEG, and the use of a multi-channel model facilitates in capturing correlation and causality between different EEG channels. By treating the 2D scalogram as an image, we show that the trained CNN-LSTM classifier based on automated visual analysis can achieve high levels of discrimination and an overall accuracy of 98.71% in case of intra-subject classification, as well as low false-positive rates. The average intra-subject accuracy obtained was 92.8%, and the average inter-subject accuracy was 86.15%. These results indicate that our proposed model performed well on the data of all subjects. We also compare the scalogram-based results with the results that we obtained by using raw time-series, showing that scalogram-based gave better performance. Our method can be applied in clinical applications such as baseline testing, assessing the current state of injury and recovery tracking and industrial applications like monitoring performance deterioration in workplaces.

Keywords: CNN; LSTM; baseline testing; cognitive-motor integration; deep learning; wearable devices.