Schizophrenia recognition based on three-dimensional adaptive graph convolutional neural network.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning models have advanced schizophrenia research but often overlook EEG's 3D spatial properties and dynamic node interactions.
• Existing methods may rely on subjective criteria for feature selection and brain network construction.

Purpose of the Study:

• To propose a 3D adaptive graph convolutional neural network (3D-AGCN) model for enhanced schizophrenia classification using EEG signals.
• To dynamically learn interactions between brain network nodes and integrate spatial, feature, and frequency band dimensions.

Main Methods:

• EEG data segmented by length and frequency bands.
• Attention mechanism for integrating multi-dimensional node features.
• Construction of adaptive brain functional networks.
• Classification using a Graph Attention Network (GAT) + Graph Convolutional Network (GCN) model.

Main Results:

• The 3D-AGCN model achieved a highest classification accuracy of 87.64% for first-episode schizophrenia patients.
• Optimal performance was observed with 6-second EEG segments and combined time- and frequency-domain features.
• The adaptive approach outperformed traditional feature selection and brain network modeling techniques.

Conclusions:

• The proposed 3D-AGCN offers a robust and adaptive method for schizophrenia classification.
• This approach provides new insights for early diagnosis and recognition of schizophrenia.
• The model's ability to capture dynamic, multi-dimensional EEG characteristics is crucial for improved diagnostic accuracy.