Learning brain connectivity with the false-discovery-rate-controlled PC-algorithm

Discovering the connectivity networks in the brain, i.e. the neural influence that brain regions exert over one another, has attracted increasing research attention in studies on brain functions. An important error rate criterion on the discovered network is the false discovery rate (FDR), that is the expected ratio of the falsely “discovered” connections to all those “discovered”. Very recently, we have developed an algorithm that is able to control the FDR under a given level q at the limit of large sample size, and its modification that controlled the FDR accurately in simulations with moderate sample sizes [1]. However, the algorithms do not consider prior knowledge on the network structure, and can not be applied to models such as dynamic Bayesian networks. In this paper, we extend the algorithms to incorporate prior knowledge, and demonstrate how to apply the extended algorithm to learning the structure of dynamic Bayesian networks from continuous data. Its application to a real functional-Magnetic-Resonance-Imaging (fMRI) data set revealed that Parkinson's disease patients' brain connectivities are normalized by L-dopa medication. This result is consistent with the fact that L-dopa has dramatic effects against bradykinesia and rigidity.