A hierarchical Bayesian model to find brain-behaviour associations in incomplete data sets.

Area of Science:

• Neuroimaging
• Statistical modeling
• Machine learning

Background:

• Canonical Correlation Analysis (CCA) is common for neuroimaging, but has limitations.
• CCA struggles with robust statistical inference and modeling within-modality associations.
• Existing methods often require data imputation or removal for missing values.

Purpose of the Study:

• Extend Group Factor Analysis (GFA) to handle missing data in neuroimaging.
• Demonstrate GFA's capability as a predictive model for multimodal data.
• Address limitations of CCA in analyzing brain imaging and behavioral data.

Main Methods:

• Applied an extended Group Factor Analysis (GFA) model.
• Utilized synthetic and Human Connectome Project (HCP) data (brain connectivity and non-imaging measures).
• Evaluated GFA's performance on complete and incomplete datasets.

Main Results:

• GFA successfully identified shared and specific factors in synthetic data, predicting missing modalities.
• In HCP data, GFA revealed associations between behavior and brain networks (default mode, frontoparietal, etc.).
• GFA accurately predicted non-imaging measures from brain connectivity, consistent across datasets.

Conclusions:

• Extended GFA provides a robust method for analyzing associations between and within neuroimaging data modalities.
• GFA handles missing data effectively and serves as a powerful predictive tool.
• GFA is a promising approach for complex analyses in large-scale datasets like HCP.