Alzheimer's disease detection using data fusion with a deep supervised encoder.

Area of Science:

• Computational neuroscience
• Medical informatics
• Machine learning

Background:

• Alzheimer's disease (AD) diagnosis requires improved accuracy and early detection methods.
• Current computational diagnosis often relies on single data modalities, potentially missing crucial information.
• Integrating diverse data sources (multi-modal data) may offer a more comprehensive patient profile for diagnosis.

Purpose of the Study:

• To develop an optimal data analysis strategy for enhancing computational diagnosis of Alzheimer's disease.
• To investigate the impact of data fusion and dimensionality reduction techniques on diagnostic accuracy.
• To compare various fusion strategies (simple, early, intermediate) and dimensionality reduction methods.

Main Methods:

• A comprehensive comparison of over 80 statistical machine learning methods was conducted.
• Explored three data fusion strategies: simple concatenation, early fusion (concatenate then reduce dimensions), and intermediate fusion (reduce dimensions then concatenate).
• Evaluated common dimensionality reduction techniques (PCA, AE, LASSO) and a novel supervised encoder (SE).

Main Results:

• The supervised encoder (SE) demonstrated substantial improvements in prediction accuracy compared to PCA, AE, and LASSO.
• Intermediate data fusion, when combined with SE, yielded the highest accuracy for multiclass diagnosis prediction.
• Multi-modal data integration, coupled with effective dimensionality reduction, enhances computational diagnostic capabilities.

Conclusions:

• Optimal data fusion and dimensionality reduction strategies are crucial for accurate computational diagnosis of Alzheimer's disease.
• The supervised encoder (SE) shows significant promise as a dimensionality reduction technique for multi-modal AD data.
• This research provides a framework for developing more effective AI-driven diagnostic tools for neurodegenerative diseases.