Self-Explainable Graph Neural Network for Alzheimer Disease and Related Dementias Risk Prediction: Algorithm...

Area of Science:

• Utilizing advanced machine learning for medical data analysis.
• Developing explainable artificial intelligence (XAI) for healthcare applications.

Background:

• Alzheimer disease and related dementias (ADRD) are a leading cause of death, necessitating accurate risk prediction.
• Current ADRD risk prediction models often rely on medical imaging, which is not universally accessible.
• Claims data offers a valuable, underutilized resource for identifying ADRD risk factors and their interconnections.

Purpose of the Study:

• To employ graph neural networks (GNNs) with claims data for enhanced ADRD risk prediction.
• To introduce a novel, self-explainable method for interpreting GNN predictions and identifying influential relationships.
• To evaluate the model's performance across different prediction time windows (1, 2, and 3 years).

Main Methods:

• A variationally regularized encoder-decoder GNN (VGNN) was developed and integrated with a relation importance method.
• The self-explainable approach provides feature-importance explanations within the context of ADRD risk prediction.
• Model performance was assessed against baseline models, including Random Forest (RF) and Light Gradient Boost Machine (LGBM), across three prediction scenarios.

Main Results:

• The VGNN model consistently outperformed RF and LGBM across all prediction windows, achieving superior Area Under the Receiver Operating Characteristic (AUROC) scores.
• For instance, in the 1-year prediction window, VGNN achieved AUROC scores of 0.7272 and 0.7480, outperforming baselines by over 9%.
• The integrated interpretation method provided insights into paired factors influencing ADRD risk and progression.

Conclusions:

• The proposed self-explainable GNN method significantly enhances ADRD risk prediction using claims data.
• This approach offers valuable insights into the complex relationships between medical codes and ADRD.
• The methodology demonstrates potential for broader applications in medical predictions beyond ADRD, including image analysis.