A Deep Learning Technique for Real-Time Detection of Cognitive Load Using Optimal Number of EEG Electrodes

Abstract

Cognitive load analysis has the potential to significantly enhance brain–computer interfaces (BCIs) by enabling adaptive assistance based on the cognitive state of individuals. This article presents a real-time approach for detecting cognitive load through electroencephalogram (EEG) signals, with a focus on optimizing computational resources, such as CPU time, memory, and the number and positioning of EEG electrodes. The study investigates various brain regions, including the prefrontal, frontal, parietal, temporal, and occipital areas, which are critical for identifying cognitive shifts. By leveraging established knowledge of EEG frequency band changes, the research constructs a 2-D brain state image using the Lambert cylindrical equal-area projection and an appropriate interpolation method to represent active brain regions. These 2-D images are then processed by a lightweight convolutional neural network (CNN) designed to distinguish between cognitive and resting states. To validate the proposed model, three EEG datasets were employed: one prepared by the authors through experiments involving 15 healthy subjects and two publicly available datasets. The model achieved an overall accuracy of 95.81% for previously seen subjects and 92.73% for entirely new subjects, utilizing only five electrodes (one prefrontal and four frontal). Furthermore, the model is demonstrated to be suitable for implementation in digital systems with limited computational resources, while maintaining performance and meeting real-time system requirements.