DSCNN-CAU: Deep-Learning-Based Mental Activity Classification for IoT Implementation Toward Portable BCI

Mental activity classification (MAC) based on electroencephalogram (EEG) is used in the brain-computer interface (BCI) and neurofeedback applications. For this purpose, machine learning and deep learning (DL) techniques are utilized in the previous studies. However, in real time, there is a need to deploy these techniques on Internet of Things (IoT)-enabled mobile devices for portability. Toward this aspect, we propose DL-based MAC using a depthwise separable convolutional neural network with a custom attention unit (DSCNN-CAU) and IoT implementation using a smartphone for portable BCI applications. The performance assessment on EEG signals from two public and two self-acquired databases demonstrates that the overall accuracies of 99.25%, 95.00%, and 91.50%, 89.25% are obtained, respectively. Evaluation results on smartphone depict that most of the real-time recorded EEG signals in self-acquired databases are classified correctly. Further comparative analysis demonstrates that the proposed model and the IoT implementation outperform the existing techniques in terms of accuracy, robustness against artifacts, latency, and battery current dissipation. A total current of 21 mAh is dissipated in EEG signal recording, processing, and event-based transmission to the server, and low latency of 71 ms is achieved in MAC. Further analysis of cross-database performance enables the selection of the $F_{P1}$ channel for real-time MAC in IoT-based scenarios.

Automated summary

In this study, a deep learning-based method for mental activity classification using a depthwise separable convolutional neural network with a custom attention unit (DSCNN-CAU) and an IoT implementation using a smartphone for portable brain-computer interface (BCI) applications is proposed. The performance assessment on EEG signals demonstrates high overall accuracies and the evaluation results on smartphone show correct classification of real-time recorded EEG signals. The proposed model and IoT implementation outperform existing techniques in terms of accuracy, robustness against artifacts, latency, and battery current dissipation. Furthermore, cross-database analysis enables the selection of the $F_{P1}$ channel for real-time mental activity classification in IoT-based scenarios.