Biomarkers.

Area of Science:

• Neuroimaging and Artificial Intelligence
• Alzheimer's Disease Research
• Brain Volumetric Analysis

Background:

• Brain volumetric changes detected by MRI correlate with cognitive decline in Alzheimer's disease (AD).
• Previous studies using univariate methods explored individual brain-region associations with cognition.
• Multivariate relationships between whole-brain (WB) volumetric changes and cognition in AD remain under-explored.

Purpose of the Study:

• To investigate complex multivariate WB-cognition relationships in the AD continuum using explainable AI.
• To predict global cognition (Mini-Mental State Examination - MMSE) using WB regional features from MRI data.
• To identify key brain regions contributing to cognitive prediction and AD severity (Clinical Dementia Rating-Sum of Boxes - CDR-SB).

Main Methods:

• Employed machine learning (ML) and deep learning (DL) models for cognition prediction.
• Integrated an optimal DL model with Shapley Additive exPlanations (SHAP) for feature importance (DL-SHAP).
• Validated DL-SHAP on semi-simulated (n=1108) and experimental (n=668) MRI datasets, assessing MMSE and CDR-SB.

Main Results:

• The DL model significantly outperformed conventional ML models in MMSE prediction.
• DL-SHAP demonstrated robust performance on semi-simulated data (Spearman's correlation=0.94) and experimental data (Spearman's correlation=0.96).
• Identified hierarchically dominant brain regions associated with MMSE prediction and AD severity.

Conclusions:

• The explainable AI method DL-SHAP effectively predicts global cognition from large MRI datasets.
• DL-SHAP successfully identifies multivariate whole-brain-cognition relationships in the AD continuum.
• The method provides compelling evidence for predicting clinical severity and highlights critical brain regions involved.