Modeling longitudinal imaging biomarkers with parametric Bayesian multi-task learning.

Area of Science:

• Neuroscience
• Biomedical Imaging
• Statistical Modeling

Background:

• Longitudinal imaging biomarkers are crucial for understanding neurodegeneration progression.
• Existing models require robustness to under-sampling, measurement errors, and multi-modal data integration.
• Accurate trajectory modeling is key for early disease detection and tracking.

Purpose of the Study:

• To present a parametric Bayesian multi-task learning approach for modeling longitudinal biomarker trajectories.
• To develop a model robust to under-sampling and measurement errors, capable of integrating multi-modal information.
• To improve trajectory inference and prediction for neurodegenerative diseases.

Main Methods:

• A parametric Bayesian multi-task learning framework was developed.
• The model optimizes a combination of uncoupled, fully coupled, and kernel-coupled models for information sharing across subjects.
• Kernel-based coupling links subject trajectories using biomarker measures like APOE genotype, CSF, and amyloid PET.

Main Results:

• The model was demonstrated using Alzheimer's Disease Neuroimaging Initiative (ADNI) data, modeling MRI-derived cortical volumes.
• Established disease effects were detected, along with novel disease-related changes in the insula.
• The model showed competitive predictive performance and sensitivity in detecting disease effects.

Conclusions:

• The proposed Bayesian multi-task learning model offers a robust approach to modeling biomarker trajectories in neurodegeneration.
• Its ability to integrate multi-modal data and share information across subjects enhances trajectory inference and prediction.
• The model's sensitivity and predictive performance suggest its utility as an alternative or supplement to traditional linear mixed effects models.