Abstract
The COVID-19 epidemic has created highly unprocessed emotions that trigger stress, anxiety, or panic attacks. These attacks exhibit physical symptoms that may easily lead to misdiagnosis. Deep-learning (DL)-based classification approaches for emotion detection based on electroencephalography (EEG) signals are computationally costly. Nowadays, limiting memory potency, considerable training, and hyperparameter optimization are always needed for DL models. As a result, they are inappropriate for real-time applications, which require large computational resources to detect anxiety and stress through EEG signals. However, a two-dimensional residual separable convolution network (RCN) architecture can considerably enhance the efficiency of parameter use and calculation time. The primary aim of this study was to detect emotions in undergraduate students who had recently experienced COVID-19 by analyzing EEG signals. A novel separable convolution model that combines residual connection (RCN-L) and light gradient boosting machine (LightGBM) techniques was developed. To evaluate the performance, this paper used different statistical metrics. The RCN-L achieved an accuracy (ACC) of 0.9263, a sensitivity (SE) of 0.9246, a specificity (SP) of 0.9282, an F1-score of 0.9264, and an area under the curve (AUC) of 0.9263 when compared to other approaches. In the proposed RCN-L system, the network avoids the tedious detection and classification process for post-COVID-19 emotions while still achieving impressive network training performance and a significant reduction in learnable parameters. This paper also concludes that the emotions of students are highly impacted by COVID-19 scenarios.