Automatic recognition of schizophrenia from brain-network features using graph convolutional neural network.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Psychiatry

Background:

• Schizophrenia diagnosis relies on subjective clinical experience and lacks objective methods.
• Graph convolutional neural networks (GCN) show promise for analyzing complex spatial-association information in psychiatric disorders.

Purpose of the Study:

• To develop and validate an automatic schizophrenia recognition model using GCN and resting-state electroencephalography (EEG) data.
• To identify key brain regions and their network features crucial for schizophrenia diagnosis.

Main Methods:

• Utilized resting-state EEG data from 103 first-episode schizophrenia patients and 92 normal controls.
• Trained a GCN model using nodal features from time-domain, frequency-domain, and brain network analyses.
• Employed 10-fold cross-validation to evaluate model performance, focusing on the theta frequency band and phase-locked values.

Main Results:

• The GCN model achieved a high recognition accuracy of 90.01% for schizophrenia.
• The parietal lobe was identified as the most significant region for distinguishing schizophrenia patients from controls.
• Explored correlations between node topological features in significant regions and clinical metrics.

Conclusions:

• The proposed GCN-based method demonstrates high applicability and accuracy for the objective identification and diagnosis of schizophrenia.
• This approach offers a novel, data-driven tool to complement existing diagnostic practices for schizophrenia.