A Multi-Domain Connectome Convolutional Neural Network for Identifying Schizophrenia From EEG Connectivity Patterns

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Altered brain functional connectivity patterns are key indicators in neuropsychiatric disorders.
• Previous deep learning approaches for fMRI-based functional network classification were limited, often focusing on single connectivity measures.

Purpose of the Study:

• To propose a deep convolutional neural network (CNN) framework for classifying schizophrenia (SZ) using electroencephalogram (EEG)-derived brain connectome.
• To integrate diverse connectivity features, including time/frequency-domain metrics and network topology, for improved classification accuracy.

Main Methods:

• Developed a novel multi-domain connectome CNN (MDC-CNN) utilizing parallel 1D and 2D CNNs.
• Integrated various connectivity features: effective connectivity metrics (vector autoregressive model, partial directed coherence) and complex network measures.
• Explored an extension for dynamic brain connectivity using recurrent neural networks.

Main Results:

• Hierarchical latent representations from EEG connectivity revealed distinct group differences between SZ and healthy controls (HC).
• The proposed CNNs significantly outperformed traditional support vector machine classifiers on a large resting-state EEG dataset.
• MDC-CNN achieved a remarkable accuracy of 91.69% with decision-level fusion, outperforming single-domain CNNs.

Conclusions:

• The MDC-CNN effectively discriminates SZ from HC by integrating information from diverse brain connectivity descriptors.
• This framework shows significant potential for developing diagnostic tools for schizophrenia and other neuropsychiatric disorders.