Pre-conference workshop: Visual mismatch negativity
Tuesday, Sep 8, 2015
SKH Z005

Multi-way data decomposition and analysis for vMMN

Fengyu Cong1 & Piia Astikainen2

1Department of Biomedical Engineering, Dalian University of Technology, Dalian, China
2Department of Psychology, University of Jyvaskyla, Finland

In order to study time-frequency representation (TFR) of ERP data, usually, ERP data is first represented by a single vector from each recording channel. The wavelet transform is then applied to it to obtain its TFR. Next, the region of interest (ROI) is defined (a time window and a frequency band). Subsequently, the magnitude of responses within the ROI is calculated and submitted for statistical analysis. Defining a ROI with a rectangular shape is convenient, but may not be the most optimal or objective method.

Indeed, TFR of ERP data can contain multiple channels, multiple stimuli and multiple subjects and can naturally compose a multi-way array (called as tensor), including the modes of time, frequency, space, stimulus and subject. The interactions among those modes can be simultaneously exploited and ROI can be adaptively extracted if the multi-way array is factorized (called as tensor decomposition).

Here, we applied tensor decomposition on the TFR of visual ERP data recorded in a passive oddball paradigm using neutral faces as standard stimuli and fearful and happy faces as deviant stimuli. In terms of the conventional ERP analysis for visual mismatch negativity (vMMN), no significant difference was found in any interactions among the three factors of Group (control vs depressed), Hemisphere (P7 vs P8), and Deviant type (fearful vs happy). Using tensor decomposition, the interaction between Group and Hemisphere was significant (p = 0.015), reflecting a right-lateralized vMMN in the participants with depressive symptoms, and a bilateral vMMN in the control participants.