Temporal-multimodal consistency alignment for Alzheimer's cognitive assessment prediction

Area of Science:

• Neuroscience
• Medical Informatics
• Biostatistics

Background:

• Alzheimer's disease (AD) is a prevalent neurodegenerative disorder significantly impacting cognition and behavior.
• Early and accurate prediction of cognitive decline is critical for effective AD intervention.
• Existing prognostic models have limited predictive power due to insufficient exploration of longitudinal multimodal data associations.

Purpose of the Study:

• To propose the Multi-Modality fusion with DUal-gRanularity Alignment (MM-DURA) framework for simultaneous modeling of longitudinal correlations and modality interactions.
• To forecast cognitive assessment scores using temporal MRI scans, clinical diagnostics, and genomic data.
• To enhance the accuracy of Alzheimer's disease progression prediction.

Main Methods:

• Developed a coarse-to-fine feature representation learning approach for subject and visit-level modality congruence.
• Implemented a hierarchical multimodality fusion (HMF) block to exploit inter-modal relationships.
• Utilized an LSTM-based regression head for forecasting cognitive ability based on fused multimodal embeddings.

Main Results:

• Validated MM-DURA on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset (707 subjects).
• Achieved superior performance in forecasting cognitive assessments (e.g., RMSE of 1.099 for CDRSB, 5.601 for ADAS-Cog).
• Outperformed six comparison methods, including state-of-the-art multimodal temporal models, affirming the effectiveness of temporal-multimodal alignment.

Conclusions:

• Proposed a novel framework unifying multimodality fusion with dual-granularity alignment for cognitive assessment forecasting.
• Demonstrated superior performance of MM-DURA compared to existing methods through extensive numerical results and visualizations.
• Highlighted the clinical potential for cognitive assessment prediction and released the code for public access.