Explaining graph convolutional network predictions for clinicians-An explainable AI approach to Alzheimer's disease...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Neuroimaging Analysis

Background:

• Graph-based representations are increasingly used in medicine to model patient relationships and predict disease.
• Graph Convolutional Networks (GCN) can analyze complex patterns in neurocognitive, genetic, and brain atrophy data for cognitive status prediction.
• Elucidating GCN predictions is crucial for clinical adoption and physician trust in diagnostic decision support systems.

Purpose of the Study:

• To introduce a decomposition-based explanation method for individual patient classification using GCNs.
• To understand the contribution of various features and relationships to diagnostic predictions.
• To enhance the interpretability and trustworthiness of GCN models in medical applications.

Main Methods:

• Utilized a GCN model on the Alzheimer's Disease Neuroimaging Initiative (ADNI) database to predict cognitive status (Normal Cognition, Mild Cognitive Impairment, Alzheimer's Disease).
• Developed a decomposition-based approach to analyze output variations based on input value changes, assessing feature and edge impact.
• Investigated relational data by selectively silencing edges to obtain neighborhood-level explanations.

Main Results:

• The explanation method demonstrated stability with minor input changes, particularly for edge weights above 0.80.
• Comparative analysis against SHAP values showed comparable results with significantly reduced computational time.
• A survey of 11 domain experts revealed that 71% confirmed the correctness of the explanations, rating understandability above six out of ten.

Conclusions:

• The proposed explanation method provides stable and understandable insights into GCN predictions for cognitive status.
• The approach offers a computationally efficient alternative to existing methods like SHAP.
• Addressing limitations, such as GCNs' reliance on demographic data, will further facilitate clinical adoption and build physician trust.