Supervised multimodal fusion and its application in searching joint neuromarkers of working memory deficits in schizophrenia

Multimodal fusion is an effective approach to better understand brain disease. To date, most current fusion approaches are unsupervised; there is need for a multivariate method that can adopt prior information to guide multimodal fusion. Here we proposed a novel supervised fusion model, called 'MCCAR+jICA', which enables both identification of multimodal co-alterations and linking the covarying brain regions with a specific reference signal, e.g., cognitive scores. The proposed method has been validated on both simulated and real human brain data. Features from 3 modalities (fMRI, sMRI, dMRI) obtained from 147 schizophrenia patients and 147 age-matched healthy controls were included as fusion input, who participated in the Function Biomedical Informatics Research Network (FBIRN) Phase III study. Our aim was to investigate the group co-alterations seen in three types of MRI data that are also correlated with working memory performance. One joint IC was found both significantly group-discriminating (p=7.4E-06, 0.001, 7.0E-09) and highly correlated with working memory scores(r=0.296, 0.241, 0.301) and PANSS negative scores (r=-0.229, -0.276, -0.240) for fMRI, dMRI and sMRI, respectively. Given the simulation and FBIRN results, MCCAR+jICA is shown to be an effective multivariate approach to extract accurate and stable multimodal components associated with a particular measure of interest, and promises a wide application in identifying potential neuromarkers for mental disorders.