Mining the posterior cingulate: Segregation between memory and pain components

Abstract

We present a general method for automatic meta-analyses in neuroscience and apply it on text data from published functional imaging studies to extract main functions associated with a brain area—the posterior cingulate cortex (PCC). Abstracts from PubMed are downloaded, words extracted and converted to a bag-of-words matrix representation. The combined data are analyzed with hierarchical non-negative matrix factorization. We find that the prominent themes in the PCC corpus are episodic memory retrieval and pain. We further characterize the distribution in PCC of the Talairach coordinates available in some of the articles. This shows a tendency to functional segregation between memory and pain components where memory activations are predominantly in the caudal part and pain in the rostral part of PCC.

Introduction

Functional neuroimaging methods such as positron emission tomography (PET) or functional magnetic resonance imaging (fMRI) map function to brain anatomy by manipulating behavior and measuring a change in regional activity. The resulting maps are used to infer the principles of brain function. The interpretability and value of these functional maps however depend critically on the specificity of the behavioral paradigm and on the sensitivity of the analytical tools for detecting the interesting signal. In any individual study, it is very difficult, if not impossible, to avoid the particular subject's in part irreproducible behavior, and it is not possible a priori to describe the spatiotemporal characteristics of the signal necessary for accurate modeling and detection. Investigating the consensus among results from a large number of brain imaging studies can overcome the problem of false positive and false negative results. Furthermore, neuroimaging studies may discover unexpected changes in activity in areas that so far have not been associated with a given behavior, thus generate new hypotheses. Unexpected activations may be genuine, showing a novel and interesting association between brain and behavior, or spurious, for instance, a false positive activation from some uninteresting behavior that the paradigm did not control for. If such changes are repeatedly encountered over a large number of different paradigms that addressed the same behavioral function, then it is probable that they are of significance and deserve further investigation.

The potential for data mining across neuroimaging studies has already been recognized and exploited to investigate the consistency of activation patterns associated with a given brain function (Cabeza and Nyberg, 2000) or to generate new hypotheses about the function of a given brain area (Bush et al., 2000, Maddock, 1999, Maguire, 2001). In these papers, the authors manually generate tables and figures that summarize results across a large number of studies in order to inspect for consensus. The expanding number of published studies, however, makes it increasingly time-consuming for the individual researcher to generate exhaustive result summaries. Here, we propose an automatic method which can extract from the neuroimaging literature the consensus about the functions of a given brain area. This method is thought to assist researchers by providing a quick summary of a large number of scientific publications. To illustrate the approach, we apply this method to extract functions that are consistently associated with the posterior cingulate cortex (PCC). The studies we include are not restricted to examine any particular behavioral function: we simultaneously examine multiple functions. To this end, we have used article abstracts indexed by PubMed under the keywords that relate to both PCC and functional imaging methods.

Despite the increasing attention that the PCC has attracted over the years (see Fig. 1), there is no “textbook” consensus about the functions undertaken by this area. Different reviews associate this area with a variety of brain functions, e.g., evaluative functions (for spatial orientation and memory) (Vogt et al., 1992), successful episodic memory retrieval (Cabeza and Nyberg, 2000), emotion (Maddock, 1999), navigation (Maguire, 2001), visuospatial attention (Mesulam et al., 2001, Small et al., 2003), pain (Bromm, 2001), and “resting state” (Binder et al., 1999, Mazoyer et al., 2001, Raichle et al., 2001, Shulman et al., 1997).

Our method is two-stage: in the first stage, we will make unsupervised text mining in the form of clustering based on the words in abstracts of PubMed that mention our specific target area (the PCC), and we compare these results with the results reported by previous manual reviews to assess the validity of this automatic analysis. In the second stage, we use the Talairach coordinates (Talairach and Tournoux, 1988) within the clustered articles and describe the spatial distribution in terms of the cluster labels from the first stage. This is in order to test whether the functional clusters are anatomically segregated within the posterior cingulate cortex.

That is, in terms of PCC, we ask whether our machine-based methodology is able to capture themes in alignment with major reviews and whether there is functional segregation within the PCC.