Multimodal attention-based deep learning for Alzheimer's disease diagnosis

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Alzheimer's disease (AD) is a prevalent neurodegenerative disorder with complex pathogenesis, complicating accurate diagnosis and clinical decision support.
• Current diagnostic approaches often struggle with the subtle differences between AD, mild cognitive impairment (MCI), and healthy controls.

Purpose of the Study:

• To develop a novel multimodal deep learning framework for enhanced Alzheimer's disease diagnosis.
• To improve the accuracy and clinical utility of decision support systems for AD and MCI detection.

Main Methods:

• Development of the Multimodal Alzheimer's Disease Diagnosis (MADDi) framework integrating imaging, genetic, and clinical data.
• Implementation of cross-modal attention mechanisms to capture inter-modal interactions, a novel approach in AD diagnostics.
• Multi-class classification comparing MADDi against state-of-the-art models and evaluating attention mechanisms.

Main Results:

• MADDi achieved 96.88% accuracy in classifying MCI, AD, and controls on a held-out test set.
• The combination of cross-modal and self-attention yielded the best performance, outperforming models without attention layers by 7.9% in F1-scores.
• Structured clinical data was identified as a critical component for contextualizing and interpreting other data modalities.

Conclusions:

• Multimodal data integration via cross-modal attention significantly enhances diagnostic accuracy for Alzheimer's disease.
• The MADDi framework demonstrates the potential of deep learning in providing effective decision support for AD diagnosis.
• Structured clinical data is indispensable for robust performance in multimodal AI models for neurodegenerative disease detection.