Predicting Cognitive Declines Using Longitudinally Enriched Representations for Imaging Biomarkers

Area of Science:

• Neuroimaging
• Biomarker Discovery
• Machine Learning

Background:

• Longitudinal neuroimaging data is crucial for studying Alzheimer's Disease (AD) progression.
• Missing patient records present a significant challenge in analyzing this data.

Purpose of the Study:

• To develop a novel method for learning fixed-length biomarker representations from longitudinal neuroimaging data.
• To capture information from both baseline scans and progressive variations despite missing follow-up records.

Main Methods:

• Proposed a new formulation to create enriched, fixed-length imaging biomarker representations.
• Developed a robust objective function maximizing l1-norm distances for unaligned, missing data.
• Derived an efficient, non-greedy iterative algorithm with proven convergence.

Main Results:

• Achieved significant performance gains in predicting ten cognitive scores compared to baseline representations.
• Demonstrated that the method effectively utilizes longitudinal information from incomplete datasets.
• Identified top biomarkers consistent with existing Alzheimer's Disease research knowledge.

Conclusions:

• The novel method provides effective enriched representations for studying Alzheimer's Disease progression.
• This approach improves the predictive power for cognitive decline using longitudinal neuroimaging data.
• The findings validate the method's effectiveness and potential for AD research.