Combined ICA-LORETA analysis of mismatch negativity

Abstract

A major challenge for neuroscience is to map accurately the spatiotemporal patterns of activity of the large neuronal populations that are believed to underlie computing in the human brain. To study a specific example, we selected the mismatch negativity (MMN) brain wave (an event-related potential, ERP) because it gives an electrophysiological index of a “primitive intelligence” capable of detecting changes, even abstract ones, in a regular auditory pattern. ERPs have a temporal resolution of milliseconds but appear to result from mixed neuronal contributions whose spatial location is not fully understood. Thus, it is important to separate these sources in space and time. To tackle this problem, a two-step approach was designed combining the independent component analysis (ICA) and low-resolution tomography (LORETA) algorithms. Here we implement this approach to analyze the subsecond spatiotemporal dynamics of MMN cerebral sources using trial-by-trial experimental data. We show evidence that a cerebral computation mechanism underlies MMN. This mechanism is mediated by the orchestrated activity of several spatially distributed brain sources located in the temporal, frontal, and parietal areas, which activate at distinct time intervals and are grouped in six main statistically independent components.

Introduction

Mismatch negativity (MMN) is an involuntary auditory event-related potential (ERP), which peaks at 100–200 ms when there is a violation (deviant tone) of a regular pattern (standard tone sequence) (Näätänen et al., 1978). The MMN mechanism appears to correspond to a “primitive intelligence,” as the wave produced with the violation of regularities—even those of an abstract nature—in the auditory stream pattern (Näätänen et al., 2001). Moreover, the deviant information provided by the change needs to be compared with a memory trace of the regularity previously stored in the brain. Thus, MMN provides a promising in vivo simplified model for studying abstract brain processing and related memory mechanisms.

Several cerebral MMN sources have been described in the literature, with global right hemisphere dominance. Studies performed with EEG (Alain et al., 1998), MEG (Rosburg et al., 2004), functional magnetic resonance (fMRI) (Jääskeläinen et al., 2004, Kircher et al., 2004), and positron emission tomography (PET) (Müller et al., 2002) have shown that the main generators of MMN are located in supratemporal cortex. Other contributions to MMN have been described in frontal cortex using electroencephalographic techniques (Giard et al., 1990, Opitz et al., 2002). PET studies (Müller et al., 2002) have revealed that these frontal sources could be in the prefrontal cortex. Similar results have also been found using fMRI (Doeller et al., 2003, Opitz et al., 2002) and intracranial recordings (Liasis et al., 2001). Electrophysiological recordings have also revealed possible frontal sources in the anterior cingulate cortex (Jemel et al., 2002, Waberski et al., 2001).

Additional sources of MMN have been located in the inferior parietal cortex (Kasai et al., 1999, Levänen et al., 1996), but its contribution is not identified in all studies. Moreover, as the relationship between structures and their temporal dynamics is still a question to be resolved, the location and nature of neural contributions to cerebral processing underlying MMN remain open questions.

In order to study the spatiotemporal dynamics of MMN cerebral sources, a two-step approach analysis has been devised. First, independent component analysis (ICA) (Makeig et al., 1996), a powerful data-driven mathematical tool that blindly separates signals' statistically independent contributions, was used to find temporally independent and spatially fixed components of ERPs. In this study, we use the Infomax ICA Algorithm (Jung et al., 2001a), in which components are obtained through minimization of mutual information among output components. ICA has recently been used to separate mixed information into spatially stationary and temporally independent subcomponents in some ERP studies (Jung et al., 2001b, Makeig et al., 1999, Makeig et al., 2002) and in other branches of neuroscience and medicine (Brown et al., 2001, Calhoun et al., 2001, Nakai et al., 2004).

In the second step, the spatial maps associated with each ICA component were analyzed, with use of low-resolution tomography (LORETA) (Pascual-Marqui, 1999, Pascual-Marqui et al., 1994), to locate its cerebral sources. LORETA is a tomographic technique that gives a single solution to what is known as the inverse problem of location of cerebral sources. It is based on two constraints: it searches for the smoothest of all possible solutions, using cortical gray matter and the hippocampus of the Talairach human brain model. In the past 5 years, this tomography approach has been used in several neuroscience studies (for example, see Gomez et al., 2003, Kounios et al., 2001, Mulert et al., 2001, Pizzagalli et al., 2001). The LORETA version used in this study reconstructed the sources of activation in 2394 voxels distributed in the Talairach human brain (Pascual-Marqui, 1999).

In summary, the main aim of this study is to uncover the spatiotemporal pattern of brain activations underlying MMN, separating statistically independent components by preprocessing data with ICA, and identifying the cerebral sources of each ICA component using LORETA.