Neuroimaging investigations have proven that social anxiety disorder (SAD) is associated with aberrations in the connectivity of human brain functions. The assessment of the effective connectivity (EC) of the brain and its impact on the detection and medication of neurodegenerative pathophysiology is hence a crucial concern that needs to be addressed. Nevertheless, there are no clinically certain diagnostic biomarkers that can be linked to SAD. Therefore, investigating neural connectivity biomarkers of SAD based on deep learning models (DL) has a promising approach with its recent underlined potential results. In this study, an electroencephalography (EEG)-based detection model for SAD is constructed through directed causal influences combined with a deep convolutional neural network (CNN) and the long short-term memory (LSTM). The EEG data were classified by applying three different DL models, namely, CNN, LSTM, and CNN + LSTM to discriminate the severity of SAD (severe, moderate, mild) and healthy controls (HC) at different frequency bands (delta, theta, alpha, low beta, and high beta) in the default mode network (DMN) under resting-state condition. The DL model uses the EC features as input, which are derived from the cortical correlation within different EEG rhythms for certain cortical areas that are more susceptible to SAD. Experimental results revealed that the proposed model (CNN + LSTM) outperforms the other models in SAD recognition. For our dataset, the highest recognition accuracies of 92.86, 92.86, 96.43, and 89.29, specificities of 95.24, 95.24, 100, and 90.91, and sensitivities of 85.71, 85.71, 87.50, and 83.33 were achieved by using CNN + LSTM model for severe, moderate, mild, and HC, respectively. The fundamental contribution of this analysis is the characterization of neural brain features using different DL models to categorize the severity of SAD, which can represent a potential biomarker for SAD. © 2013 IEEE. © 2021 Institute of Electrical and Electronics Engineers Inc.. All rights reserved.