Bayesian scalar-on-image regression with spatial interactions for modeling Alzheimer's disease

Area of Science:

• Neuroscience
• Biostatistics
• Medical Imaging

Background:

• Predictive modeling for cognitive impairment in Alzheimer's disease (AD) using neuroimaging shows progress.
• Existing models often overlook heterogeneity from interactions between brain imaging features and demographic, clinical, or genetic risk factors.
• Ignoring this heterogeneity can lead to inaccurate predictions and biased estimations in AD research.

Purpose of the Study:

• To develop a novel statistical framework that incorporates spatially varying interactions between brain imaging data and supplementary risk factors for improved prediction of cognitive impairment in AD.
• To address the limitations of current predictive models by accounting for complex interactions and heterogeneity in AD.
• To identify specific brain regions and their interactions with risk factors that are significantly associated with cognitive abilities in AD.

Main Methods:

• A Bayesian hierarchical model within a scalar-on-function framework using multi-resolution wavelet decomposition.
• Incorporation of spatially varying interactions between brain imaging features and demographic, clinical, and genetic risk factors.
• Application of a spike and slab mixture prior with latent class distribution for simultaneous sparsity and clustering to handle high dimensionality.
• Development of an efficient Markov chain Monte Carlo algorithm for posterior computation.

Main Results:

• The proposed model significantly improved the prediction of cognitive impairment in AD across multiple longitudinal visits compared to existing methods.
• The model successfully identified key brain regions in AD that show significant associations with cognitive abilities.
• Demonstrated the importance of considering interactions between brain imaging and risk factors for accurate AD prediction.

Conclusions:

• The novel Bayesian approach effectively models heterogeneity and interactions, leading to enhanced prediction of cognitive impairment in Alzheimer's disease.
• This method provides a more nuanced understanding of AD pathophysiology by highlighting region-specific interactions with risk factors.
• The findings suggest a promising direction for developing more personalized and accurate diagnostic and prognostic tools for AD.