Graph Data Augmentation for Graph Convolutional Networks Learning in Robust Mental Disorder Prediction with Limited...

Area of Science:

• Biomedical informatics
• Machine learning
• Psychiatric research

Background:

• Graph neural networks (GNNs) excel in biomedical tasks but struggle with noisy and scarce data in psychiatric disease prediction.
• Existing methods lack effective solutions for these data challenges in mental illness prediction.

Purpose of the Study:

• To propose a graph data augmentation method to overcome data noise and scarcity in mental illness prediction.
• To enhance the robustness and accuracy of GNN models for psychiatric disease prediction.

Main Methods:

• Utilized edge predictors to refine graph topology, strengthening connections between similar nodes and removing noisy edges.
• Incorporated adversarial perturbations in the feature space to improve model robustness against label noise.
• Implemented a confident self-checking mechanism for accurate pseudolabeling to aid model training.

Main Results:

• The proposed graph data augmentation method demonstrated superior performance on two multimodal real mental illness datasets.
• Ablation studies confirmed the effectiveness of individual components within the framework.
• The framework proved effective and scalable for population-based disease prediction under noisy and sparse data conditions.

Conclusions:

• The novel graph data augmentation approach effectively addresses data noise and scarcity in psychiatric disease prediction.
• The method enhances GNN model performance and robustness, offering a scalable solution for real-world applications.
• Publicly available code facilitates further research and application in mental illness prediction.