MG-TCCA: Tensor Canonical Correlation Analysis Across Multiple Groups.

Area of Science:

• Neuroimaging
• Biostatistics
• Machine Learning

Background:

• Traditional Tensor Canonical Correlation Analysis (TCCA) struggles with heterogeneous tensor data, potentially leading to biased results in group-specific analyses.
• Real-world datasets, like brain imaging, often exhibit heterogeneity due to factors such as sex and race, necessitating advanced analytical approaches.

Purpose of the Study:

• To introduce Multi-Group TCCA (MG-TCCA), a novel method for joint analysis of multiple subgroups within tensor datasets.
• To address data heterogeneity and leverage cross-group information for identifying consistent signals in complex datasets.

Main Methods:

• Developed MG-TCCA incorporating a dual sparsity structure and a block coordinate ascent algorithm.
• Applied MG-TCCA to analyze correlations between AV-45 and FDG PET imaging modalities in an Alzheimer's disease cohort.
• Quantified shared and individual structures, reduced dimensionality, and enabled visual exploration.

Main Results:

• MG-TCCA demonstrated superior performance compared to traditional TCCA and Sparse TCCA (STCCA).
• The method successfully identified sex-specific cross-modality imaging correlations in Alzheimer's disease.
• MG-TCCA effectively handled heterogeneity and leveraged information across subgroups.

Conclusions:

• MG-TCCA offers a robust solution for analyzing heterogeneous tensor data, particularly in neuroimaging.
• The approach provides valuable insights into multimodal imaging biomarkers for Alzheimer's disease by revealing subgroup-specific correlations.