Hybrid Deep Learning Framework for EEG-Based Anxiety Detection and Classification in Brain–Computer Interfaces

Abstract

 Anxiety disorders, as depicted with panic disorder, social anxiety disorder (SAD), and generalized anxiety disorder (GAD), are some of the most widespread form of mental health illness, and the challenges of their diagnosis is just as profound as their prevalence. Even though many studies in the field utilize EEG and Deep Learning (DL) to detect anxiety, this work is the first to build a Brain-Computer Interaction (BCI) ecosystem utilizing Electroencephalography (EEG) signals to detect anxiety. The proposed BCI uses a hybrid architecture that integrates CNN, LSTM, and Transformer modules to detect the spatial, the temporal, and the contextual features of the EEG. The data underwent Independent Component Analysis (ICA), outlier removal, normalization, and SMOTE to increase the quality of the outcomes. The best performing individual models were CNN, LSTM, and Transformer, which achieved 85.90%, 91.54%, and 91.12% of accuracy, respectively. The hybrid model, with 97.45% of accuracy, soared ahead of the rest with 7% to 12% presented by the other Deep learning (DL) approaches based on Brain-Computer Interface (BCI). This holds promise and generalizes the model to serve as a highly effective clinical tool that is scalable and non- invasive to conduct clinical anxiety diagnosis. Further work is promised for this to include more physiological signals such as the electrocardiography (ECG) and functional near-infrared spectroscopy (fNIRS), to increase the robustness of the model.