Human activity recognition using deep electroencephalography learning

• Published in: Biomedical signal processing and control 2020-09, Vol.62, p.102094, Article 102094
• Main Authors: Salehzadeh, Amirsaleh, Calitz, Andre P., Greyling, Jean
• Format: Article
• Language: English
• Subjects: Deep learning; EEG artifact classification; EEG wearable device; Human activity recognition
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2020.102094

•A deep learning-based framework for classifying EEG artifacts is proposed and evaluated, using a stack of an LSTM Network against a CNN.•Evaluating the state-of-the-art works in classifying raw EEG signals and HAR shows the benefits of using the proposed framework.•The human cognition and emotion data extracted by classifying EEG signals captured by a consumer-grade EEG wearable device is emphasized.•Using a consumer-grade EEG wearable device to collect data for HAR provides advantages over the common sensory technologies used in HAR. Electroencephalography (EEG) signals can be contaminated by the noise raised from a non-cerebral artifact and vary in magnitude due to physiological differences between individuals. Hence, EEG has not been extensively applied to the design of Human Activity Recognition (HAR) systems. HAR involves classifying the activities of an individual that are captured by sensory technologies. To address this issue, this paper introduces a deep learning-based framework for classifying EEG artifacts (FCEA), based on a person’s physiological activity. The FCEA proposes an approach for designing a processing pipeline involving a Convolutional Neural Network and a Long Short-Term Memory Recurrent Neural Network, in order to classify raw EEG signals based on the EEG artifacts. To demonstrate the performance of the FCEA, a 3-class dataset of EEG activities was created. A consumer-grade EEG wearable device was used to collect the data from two prefrontal EEG channels, from 8 participants whilst speaking loudly, reading printed text and watching a TV program. These activities include jaw-clenching and head and eye movements, that are part of a wide range of daily human activities. Moreover, these activities cannot easily be detected by using the sensory technologies that have commonly been used in HAR. The comparative performance analysis results demonstrate that the 1- and 2-channel models trained by the FCEA outperform competitive state-of-the-art deep-learning models for HAR and raw EEG data. The deep-learning approach proposed by the FCEA improves raw data processing to obtain a better generalization performance.