A Comparison of ERP Data Cleaning Strategies for Neuroergonomic Error Detection

Abstract

Event Abstract Back to Event A Comparison of ERP Data Cleaning Strategies for Neuroergonomic Error Detection Ben D. Sawyer1*, Waldemar Karwowski2, 3, Petros Xanthopoulos4 and P. A. Hancock3 1 Massachusetts Institute of Technology, United States 2 University of Central Florida, Department of Industrial Engineering, United States 3 University of Central Florida, Department of Psychology, United States 4 Stetson University, United States The decision to employ postprocessing on electroencephalographic (EEG) data, toward the removal of undesirable artifacts, is associated with concerns of inadvertently filtering brain process data of interest to the research question. The rich data provided by multichannel EEGs supports a variety of postprocessing approaches. Brain process characteristics are often already well-studied1,2, and so the approach often impractical terms involves applying a postprocessing technique, and determining if the aggregate signal representing the brain process of interest matches those previously reported in the literature. However, as increased interest in real-time approaches to characterizing brain processes dominates the applied neuroergonomic literature, it is worth considering the absolute merits of various postprocessing techniques. For example, in event related potential/evoked response potential (ERP) work analyzed after collection it is common to utilize independent component analysis (ICA), which relies upon this statistical independence of variance accounted for by artifacts and separates them from variance accounted for by brain activity. ICA techniques, in effect, “clean” the waveform for analysis, preserving epics of interest. This is, however, a relatively computationally “expensive” approach for real-time applications. A relatively simple technique, moving window peak-to-peak amplitude detection (P2PW), uses differences between the highest and lowest voltages within successive epics of time to flag artifacts for removal. P2PW, therefore, does not preserve epics of interest, instead removes them entirely. The present work compares the performance of these two approaches in data collected by Sawyer et al.2,3 during an experiment which, for the first time, demonstrated the detection of the error related negativity (ERN) ERP in visual search for complex stimuli. In this work, participants completed tasks during 8 channel EEG recording, which was then analysed using ICA post-processing3. Successfully elicitation and detection of this ERN in visual search of complex images opens the door to applied neuroergonomics ‘in the field’ (as in Fedota & Parasuraman, 2010) 1,3. The question of how best to process data “on-the-fly”, however, is relevant specifically because of the context: computation costs power, which is heavy and expensive to carry in the field. For example, in event related potential/evoked response potential (ERP) work analyzed after collection it is common to utilize independent component analysis (ICA), which relies upon this statistical independence of variance accounted for by artifacts and separates them from variance accounted for by brain activity. ICA techniques, in effect, “clean” the waveform for analysis, preserving epics of interest. This is, however, a relatively computationally “expensive” approach for real-time applications. A relatively simple technique, moving window peak-to-peak amplitude detection (P2PW), uses differences between the highest and lowest voltages within successive epics of time to flag artifacts for removal. P2PW, therefore, does not preserve epics of interest, instead removes them entirely. The present work compares the performance of these two approaches in data collected by Sawyer et al.2,3 during an experiment which, for the first time, demonstrated the detection of the error related negativity (ERN) ERP in visual search for complex stimuli. In this work, participants completed tasks during 8 channel EEG recording, which was then analysed using ICA post-processing3. Successfully elicitation and detection of this ERN in visual search of complex images opens the door to applied neuroergonomics ‘in the field’ (as in Fedota & Parasuraman, 2010) 1,3. The question of how best to process data “on-the-fly”, however, is relevant specifically because of the context: computation costs power, which is heavy and expensive to carry in the field. Figure 1: Waveform data for errors and non-errors are shown across a simple letter flanker task and a complex motorcycle conspicuity task, separated by three postprocessing strategies, A) ICA B) P2PW, and C) raw data. These average time-locked ERP waveforms for are represented with negitive plotted down, and relative to a 50ms baseline time-locked against participant response by keypress. A full 100ms of pre response activity is shown here for evaluative purposes. The waveforms for ICA (A) show the clearest ERN pattern, but the negative trend of erroneous results and separation between correct and error trials can be clearly seen in the P2PW (A) and raw data (C) waveforms. Table 1 Task data loss by post-processing type Figure 1 Image 1