A novel deep learning model for objective quantification of generalized anxiety disorder severity using EEG functional connectivity

Generalized anxiety disorder (GAD) is a prevalent and disabling psychiatric condition, yet its severity is still assessed mainly through clinical interviews and self-report scales, which lack objective neurobiological markers. This study aimed to develop an electroencephalography (EEG)-based deep learning (DL) model for objective quantification of GAD severity based on functional connectivity (FC) features. Resting-state EEG was recorded for 10 min from 80 patients with GAD and 39 healthy controls (HC). EEG segments with window lengths between 2 and 10 s were used to compute band-limited FC features, which were then used as input to a convolutional gated multilayer perceptron (Conv_gMLP) network for continuous prediction of the Hamilton Anxiety Rating Scale (HAM-A) total scores. The Conv_gMLP model achieved a mean absolute error (MAE) of 0.32 ± 0.07 in predicting the HAM-A total score (range: 0-56), outperforming conventional machine learning (ML) models and other DL architectures. Feature attribution analyses indicated that connectivity between frontal and temporal regions, particularly in the beta frequency range, contributed most strongly to the prediction of GAD severity. These findings suggest that EEG FC and beta rhythms encode clinically meaningful information about GAD severity, and that Conv_gMLP-based models may provide a promising tool for objective, time-efficient assessment to support individualized treatment planning.