An explainable autoencoder with multi-paradigm fMRI fusion for identifying differences in dynamic functional connectivity during brain development (Q6053366)

To better identify the dynamic functional connectivity (dFC) in brains with significant differences during brain development, a multi-paradigm fusion-based explainable deep sparse autoencoder (MF-EDSAE) is proposed. The construction of the MF-EDSAE is based on a deep sparse autoencoder (DSAE). For the processing of multi-paradigm information two strategies are used: instead of a combination of layers of different paradigms, a nonlinear fusion layer is defined and to preserve the high-order relationships within each paradigm and between paradigms, a hypergraph regularization is introduced. In the second section of the article the new MF-EDSAE is introduced. These autoencoder consists of following parts: the nonlinear fusion layer, the hypergraph regularization, a multi-paradigm hypergraph regularization and a feature selection layer. The MF-EDSAE architecture and the learning process are extensively explained. One introduces the hypergraph regularization and the multi-paradigm hypergraph regularization. In the training section, the mathematical model to the training with single-paradigm data and training with multi-paradigm data is derived. As future selection layer one uses an already known \(k\)-means procedure with \(k=2\). The developed MF-EDSAE is applied to selected fMRI (functional magnetic resonance imaging) data from the Philadelphia Neurodevelopmental Cohort (PNC) which is a collaborative project between the Brain Behavior Laboratory at the University of Pennsylvania and the Children's Hospital of Philadelphia. One investigates intrinsic functional changes during brain development from resting-state fMRI, fMRI of working memory and emotion identification tasks. The third section of the article is devoted to data collection, data processing, data reconstruction, dFC selection and analysis of dynamic functional connectivity states. A discussion on proposed MF-EDSAE and conclusions to the work are contained in the last two sections. Additional computations are to be found in the Appendix A and B.