Functional and Clinical: An Explainable Deep Learning Model for Multimodal Alzheimer's Disease Classification.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Clinical Diagnostics

Background:

• Deep learning models using fMRI show promise for Alzheimer's disease (AD) classification.
• Challenges include small datasets, lack of explainability, and reliability issues like data leakage, hindering clinical adoption.

Purpose of the Study:

• To develop a reliable and explainable multimodal deep learning model for AD classification.
• To address limitations of fMRI-based models by integrating clinical data and employing explainable AI (XAI).

Main Methods:

• A 3D convolutional neural network was trained on fMRI data from the default mode network and five clinical tests.
• Multimodal data integration and a strict leave-one-out cross-validation were used to overcome data limitations and prevent data leakage.
• Perturbation ranking was applied for feature importance analysis.

Main Results:

• The multimodal model achieved 90% accuracy in classifying AD from controls, significantly outperforming models using only fMRI (58% accuracy).
• Feature importance varied by diagnostic group, with clinical tests like the MoCA showing differential relevance for controls versus AD patients.
• Explainable AI revealed distinct patterns of feature importance across clinical and fMRI data.

Conclusions:

• Combining fMRI and clinical data in a multimodal deep learning model enhances AD classification accuracy and provides insights into disease characteristics.
• The developed model demonstrates potential for refining diagnostic tools for Alzheimer's disease.
• Further external validation with larger sample sizes is recommended.