Classification of psychosis spectrum disorders using graph convolutional networks with structurally constrained...

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Classifying psychosis spectrum conditions, including subthreshold psychotic-like experiences (PLEs), presents challenges due to subtle neurobiological differences.
• A transdiagnostic approach aligns with modern understanding of psychosis neurobiology.
• Distinguishing individuals with PLEs from healthy controls requires advanced analytical methods beyond traditional diagnostic categories.

Purpose of the Study:

• To explore the utility of machine learning, specifically Support Vector Machines (SVMs) and Graph Convolutional Networks (GCNs), for classifying individuals with psychosis spectrum conditions.
• To identify neurobiological markers differentiating individuals with PLEs from healthy controls using brain network features.
• To investigate the potential of deep learning for improving diagnostic accuracy and enabling early intervention in psychosis.

Main Methods:

• Utilized functional and structural brain network data from individuals with diverse psychosis spectrum conditions and healthy controls.
• Employed Support Vector Machines (SVMs) and Graph Convolutional Networks (GCNs) with various edge selection techniques for classification.
• Applied the MultiVERSE algorithm for generating network embeddings as input features for SVMs.

Main Results:

• The best classification models (SVMs and GCNs) achieved accuracies exceeding 63%.
• Identified a region within the right inferior parietal cortex (PGi) as a crucial network hub differentiating individuals with PLEs from controls.
• Class activation mapping highlighted distinct salient regions in the PLE group, including the dorsolateral prefrontal, orbital, polar frontal, and lateral temporal cortices.

Conclusions:

• Deep learning methods show promise in distinguishing individuals with subclinical psychosis from healthy controls.
• Findings suggest potential for enhancing diagnostic accuracy, establishing neurobiological bases for diagnoses, and informing early intervention strategies.
• The identified PGi region and frontal/temporal cortical patterns offer insights into the neurobiology of the psychosis spectrum.