Multi-modal data fusion using group-structured sparse canonical correlation analysis: A simulation study

Recent studies in the field of multi-modal data fusion demonstrate that considering the prior spatial structure and groupness information available from many years of intensive neuroimaging research can increase interpretability and detection accuracy of the hidden phenomena within multi-modal datasets. Although recent functional neuroimaging data analyses indicate that certain brain regions may participate in multiple groups (networks), the extent of group overlap is never considered in data fusion frameworks. To address this issue, here we propose a group-structured sparse canonical correlation analysis (gssCCA) method by employing groupness and sparsity constraints in a unified fusion framework. To investigate the performance of the proposed algorithm, we compare the effect of overlapping and disjoint gssCCA in a simulation study, to evaluate their ability in detecting multi-modal data associations. The results demonstrate that considering an overlapping group-structure constraint can increase the sensitivity and specificity of the true associations between multi-modal datasets.