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

Area of Science:

• Neuroscience
• Psychiatry
• Computational Neuroscience

Background:

• Generalized Anxiety Disorder (GAD) assessment relies on subjective methods, lacking objective neurobiological markers.
• There is a need for reliable, objective tools to quantify GAD severity for effective treatment planning.

Purpose of the Study:

• To develop a deep learning (DL) model using electroencephalography (EEG) functional connectivity (FC) for objective GAD severity quantification.
• To assess the model's performance in predicting the Hamilton Anxiety Rating Scale (HAM-A) scores.

Main Methods:

• Resting-state EEG data were collected from 80 GAD patients and 39 healthy controls.
• Band-limited FC features were extracted from EEG segments and used to train a convolutional gated multilayer perceptron (Conv_gMLP) model.
• The model continuously predicted HAM-A total scores.

Main Results:

• The Conv_gMLP model achieved a mean absolute error of 0.32 ± 0.07 in predicting HAM-A scores, outperforming other machine learning and DL models.
• Feature attribution highlighted the importance of frontal-temporal connectivity in the beta frequency range for GAD severity prediction.
• EEG FC and beta rhythms contain clinically relevant information about GAD severity.

Conclusions:

• EEG-based DL models, particularly Conv_gMLP, show promise for objective and efficient GAD severity assessment.
• These findings support the use of neurobiological markers for individualized anxiety disorder treatment planning.