Multi-View Hilbert Curve-Based Hierarchical Information Aggregation for Incomplete Multimodal Alzheimer's Disease...

Area of Science:

• Artificial Intelligence in Medicine
• Neurodegenerative Disease Diagnostics
• Medical Data Fusion

Background:

• Early Alzheimer's disease (AD) diagnosis relies on multi-modal data (neuroimaging, biomarkers, cognitive tests).
• Data unavailability due to cost, compliance, or risks is a common clinical challenge.
• Existing AI models (CNNs, Transformers) struggle with heterogeneous medical data and capturing both local/global dependencies.

Purpose of the Study:

• To develop a novel hierarchical information aggregation and dynamic fusion (HI-AD) framework for incomplete multi-modal AD diagnosis.
• To address limitations in current AI architectures for handling high-dimensional, heterogeneous medical data with missing modalities.
• To improve the robustness and generalizability of early-stage AD screening in diverse clinical settings.

Main Methods:

• Developed a HI-AD framework incorporating a multi-view Hilbert curve-guided Mamba block for spatial feature extraction and long-range dependency modeling.
• Implemented hierarchical spatial feature extraction to preserve spatial continuity and integrate local context from neuroimaging.
• Utilized a unified mutual information-driven learning objective with active confidence evaluation to balance semantic alignment and modality-specific information, preventing modality collapse.

Main Results:

• The HI-AD framework demonstrated superior performance compared to state-of-the-art methods in extensive experiments on real-world datasets.
• Consistent outperformance was observed across various scenarios with missing data modalities.
• The method proved effective and generalizable for early-stage AD screening in heterogeneous clinical data environments.

Conclusions:

• The proposed HI-AD framework offers an effective and robust solution for Alzheimer's disease diagnosis using incomplete multi-modal data.
• The novel approach successfully handles data heterogeneity and missing modalities, outperforming existing methods.
• HI-AD establishes a promising direction for improving early AD detection in resource-limited or complex clinical settings.