Biomarkers.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Standardized multimodal neuroimaging markers are crucial for early detection of preclinical dementia states and improving clinical trials.
• The SCAN initiative harmonizes diverse biomarkers (PET, MRI, CSF) across Alzheimer's Disease Research Centers (ADRCs) for cross-site analyses.
• Integrating diverse data modalities to predict cognitive impairment remains a significant challenge.

Purpose of the Study:

• To develop an interpretable AI model leveraging SCAN data to predict the Clinical Dementia Rating (CDR) global score.
• To assess the accuracy of AI models in predicting CDR scores using multimodal MRI data.
• To establish a foundation for early Alzheimer's diagnosis through advanced AI techniques.

Main Methods:

• Analysis of 2,006 participants' MRI biomarkers and UDS-3 assessments from NACC data.
• Development of a deep neural network (DNN) with multimodal fusion and concept embedding to integrate diverse data types.
• Utilizing Shapley Additive Explanations (SHAP) for model interpretability and feature importance quantification.

Main Results:

• The Concept Embedding MLP demonstrated superior performance in CDR classifications compared to other machine learning models.
• SHAP analysis identified key predictors, including hippocampal volumes and anterior cingulate thickness.
• The model provides both global feature importance patterns and local explanations for individual predictions.

Conclusions:

• The developed AI approach successfully integrates multimodal neuroimaging to predict CDR scores, supporting early Alzheimer's diagnosis.
• Model interpretability highlights crucial neuroimaging features, guiding future research directions.
• Future work will focus on incorporating longitudinal data and external validation for enhanced generalizability.