A multilayer multimodal detection and prediction model based on explainable artificial intelligence for Alzheimer's...

Area of Science:

• Computational neuroscience and machine learning applied to neurodegenerative disease diagnostics.

Background:

• Alzheimer's disease (AD) diagnosis and progression detection are critical but hindered by reliance on single modalities, separate analyses, and lack of model interpretability.
• Physicians require trustworthy, explainable models for clinical adoption in diagnosing dementia and predicting its progression.

Purpose of the Study:

• To develop an accurate and interpretable model for Alzheimer's disease diagnosis and mild cognitive impairment (MCI) to AD progression detection.
• To provide physicians with accurate decisions and clear explanations for enhanced trust and clinical applicability.

Main Methods:

• Developed a two-layer Random Forest (RF) model integrating 11 modalities from 1048 subjects in the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
• Layer 1: Multi-class classification for early AD diagnosis. Layer 2: Binary classification for MCI-to-AD progression within three years.
• Employed SHapley Additive exPlanations (SHAP) for global and instance-based RF explanations, complemented by decision trees and fuzzy rule-based systems presented in natural language.

Main Results:

• Achieved cross-validation accuracy of 93.95% and F1-score of 93.94% in the first layer (AD diagnosis).
• Achieved cross-validation accuracy of 87.08% and F1-score of 87.09% in the second layer (MCI-to-AD progression).
• Explanations derived from SHAP and other methods were consistent with each other and existing Alzheimer's disease literature.

Conclusions:

• The developed model is accurate, interpretable, trustworthy, and medically applicable for Alzheimer's disease diagnosis and progression.
• The system provides detailed insights into modality effects, enhancing clinical understanding and potentially improving patient care pathways.
• Explainable AI in this context bridges the gap between complex models and clinical decision-making for neurodegenerative diseases.