IA-GCN: Interpretable Attention based Graph Convolutional Network for Disease Prediction.

Area of Science:

• Medical Imaging Analysis
• Machine Learning
• Computer Vision

Background:

• Interpretability in Graph Convolutional Networks (GCNs) is crucial but underexplored in the medical domain.
• Existing GCN interpretability methods often use post-hoc analysis of model outputs.
• There is a need for GCN models that provide inherent interpretability for clinical applications.

Purpose of the Study:

• To develop an interpretable attention module (IAM) for Graph Neural Networks (GNNs) that directly explains input feature relevance.
• To improve GNN performance in medical tasks by leveraging feature interpretations.
• To support clinical decision-making in diagnosis and treatment planning.

Main Methods:

• Proposed an interpretable attention module (IAM) that operates directly on input features.
• IAM learns feature attention using unique interpretability-specific losses.
• Applied the model to disease classification on the Tadpole dataset and age/sex prediction on the UK Biobank (UKBB) dataset.

Main Results:

• Achieved an average accuracy increase of 3.2% for Tadpole disease classification.
• Improved UKBB sex prediction accuracy by 1.6% and age prediction by 2%.
• Demonstrated exhaustive validation and provided clinical interpretations of the results.

Conclusions:

• The proposed IAM enhances GNN interpretability and performance in medical applications.
• IAM's direct feature relevance explanation aids clinical experts in decision-making.
• The model shows significant improvements over state-of-the-art methods on public medical datasets.