= The Chicago Electrical Neuroimaging Analytics (CENA): = = Microsegmentation Suite Tutorial = ''Author: Stephanie Cacioppo, Ph.D.'' ''The University of Chicago Pritzker Medical School'' <> == Introduction == The [[https://hpenlaboratory.uchicago.edu/page/cena|Chicago Electrical Neuroimaging Analytics (CENA]]) is a suite of tools for adavanced dynamic spatiotemporal brain analyses that allows you to automatically detect event-related changes in the global pattern and global field power of your event-related potentials (ERPs). == CENA functions == Cena functions include: * Difference wave function; * High-performance microsegmentation suite (HPMS) which consists of three specific analytic tools: * a root mean square error (RMSE) metric for identifying stable states and transition states across discrete event-related brain micro states; * a similarity metric based on cosine distance in n dimensional sensor space to determine whether template maps for successive brain microstates differ in configuration of brain activity; * a global field power (GFP) metrics for identifying changes in the overall level of activation of the brain. * Bootstrapping function for assessing the extent to which the solutions identified in the HPMS are robust. == Description of the toolbox functions == {{{ 1. Difference wave function }}} The CENA function constructs a “difference waveform” that putatively represents physiological processes that are different between two conditions. The CENA difference wave function offers users the possibility to create a difference waveform configuration between two n-dimensional ERPs by subtracting the ERP waveform elicited by one condition (e.g., ERP_A) from the ERP waveform elicited by another condition (ERP_B). The output of this difference waveform function is computed as ERP_A – ERP_B, which results in a T x n matrix with T as the number of timeframes and n as the number of electrodes. When processing two ERPs via the Brainstorm routine window at the bottom of the Brainstorm interface, ERP_A will be the ERP at the top of the list and ERP_B will be the ERP second in the list. {{{ 2. High-Performance Microsegmentation Suite (HPMS) }}} __CENA toolbox currently allows users to perform two types of HPMS__: 1. HPMS for one condition (HPMS single) or 1. HPMS to com-pare two or more conditions (HPMS multiple). Menu options of the HPMS function (either HPMS single or HPMS multiple) allow users to select two different levels (either a 95% or 99%) of confidence interval (CI) for: * (i) thresholding RMSE peaks and valleys, and * (ii) performing a cosine metric analysis to determine whether time-adjacent microstates differed in configuration. A 95% CI is recommended in between-subjects contrasts, while a 99% CI is recommended in within-subjects contrasts (S. Cacioppo & Cacioppo, 2015; S. Cacioppo et al., 2014). In addition, the menu options allow users to specify the duration of their baseline (e.g., period prior a stimulus onset) and to tune the size of the RMSE lag for the HPMS at a minimum duration that isappropriate to their study. '''We recommend that the baseline to be time-jittered in your experiment, variable in length, and corrected to ensure the best possible model of noise. (Time-jittering the baseline is typical in fMRI research and is done to ensure the baseline is a reasonablemodel of the background noise level for the signal of interest.)''' Because most ERP research (and most microstate analyses) focuses on post-stimulus event-related brain states, post-stimulus brain microstates (evoked brain microstates) are the primary focus of the present version of CENA. However, if the experimenter were interested in pre-stimulus states, a straightforward modification of the experimental design would be sufficient to permit investigation of these pre-stimulus (e.g., anticipatory) microstates. If you are interested in evoked brain microstates, here is a typical trial structure we recommend: * (i) jittered, variable-length baseline, * (ii) stimulus onset, and * (iii) post-stimulus period during which evoked brain microstates are identified and investigated. If one were interested in the event-related anticipatory microstates and wished to use CENA, the trial structure could be modified as follows: (i) jittered, variable-length baseline, (ii) a fixed-interval pre-stimulus period that makes it possible for the subject to anticipate the stimulus onset (and during which evoked anticipatory microstates can be identified and investigated), (iii) stimulus onset, and (iv) post-stimulus period (during which evoked microstates can be identified and investigated). '''Root Mean Square Error (RMSE):''' The first HPMS step uses a root mean square error (RMSE) analysis that decomposes the n-dimensional ERP waveform based on noise levels detectedduring the baseline period into two types of event-relatedbrain states: (i) discrete stable microstates, and (ii) transition states between these microstates transitions are not immediate. For more details about the RMSE, see S. Cacioppo & Cacioppo, 2015 and S. Cacioppo et al., 2014. '''Lag parameter: ''' The Lag parameter in CENA corresponds to the the minimum duration for a putative microstate. Setting up a lag allows you to set the distance between topographical maps that areto be compared. L is the minimum duration for a putative evokedbrain microstate, which means the time interval between topo-graphical maps (i.e., map x and mapˆx) that are to be compared. Because a brain microstate must have a minimum duration of a few consecutive time points to be meaningful of a functional brain processing, we recommend an L lag of approximately 8 ms (for basic visual tasks, such as a passive reversal checkerboard) and at least 12 ms for more complex cognitive task. '''Cosine similarity metric: ''' To confirm whether the microstates identified in the RMSE differ in the configuration of brain activity, CENA employs a multi-dimensional cosine similarity metric based on the cosine distance between template maps for successive evoked brain microstates (S. Cacioppo et al., 2014). Although the cosine similarity metric resolves ambiguities left by the RMSE analysis, the RMSE analysis is a necessary first step to identify candidate brain microstate based on the ERP configuration across n-dimensional sensory space. Specifically, the RMSE analysis identifies significant changes in the stable event-related pattern of EEG activation across the n-dimensional sensor space. However, there are two reasons such a change in the RMSE function may occur (Cacioppo et al., 2014): 1. '''A different stable evoked brain microstate was elicited''', typically interpreted as meaning that one or more of the cortical sourcesunderlying the prior event-related microstate had changed; or 1. '''The same stable evoked brain microstate was maintained but GFP increased (or decreased)''', typically interpreted as meaning that the level of activation of the set of cortical sources underlying theevent-related microstate had increased (or decreased). Once the putative stable microstates have been identified by the RMSE, each topographical map within a microstate can be expressed within a n-dimensional (e.g., 128-dimensional) vector space, the template (i.e.,mean) map for the microstate can be expressed in this microstate, a confidence interval region can be determined around this. If the succeeding evoked brain microstate identified by RMSE is the result of a change in the location of the underlying neural sources of the n-dimensional event-related waveform, the cosine metric between the template map for anevent -related microstate and the template map for the succeedingmicrostate should differ. This is because different configurationsof activity produce different vector angles in n-dimensional vectorspace. However, if the succeeding evoked brain microstate identi-fied by RMSE is the result of a change in the level of neural activation (i.e., GFP) rather than a change in source location, then the represen-tation of these microstates in n-dimensional vector space differ inthe length of the vector but not in the angle of the vector (Cacioppoet al., 2014). Therefore, the RMSE is followed by an analysis based on a cosine similarity metric (for details, see S. Cacioppo et al., 2014). The results of a HPMS single provide two types of outputs: Oneoutput with “preliminary” results (provided users select the option“yes” to the question “plot preliminary results” (supplementary Fig.S2), and one output with final results (Fig. 2). * If the user is interested in changes in GFP, then outputting the preliminary results should be selected. The preliminary out-put includes information about microstates before they are merged using a multi-dimensional cosine similarity metric based on cosine distance function that determines whether template maps for successive brain microstates differ in configuration of brain activity. The final output includes, on the other hand, results after the merging of the brain microstates. A comparison of these outputs permits one to identify which microstates identified by the RMSE analysiswere subsequently determined by the analysis based on the cosinemetric as the same microstate but at a different GFP. Changes in GFP levels within the same microstate are provided in the GFP outputsfor the microstates in the preliminary results that were merged inthe final results. * If one has no interest in GFP, then there isno need to output the preliminary results. The Final outcomes suffice. Both the preliminary and final outputs are organized similarly. They both display two figures and five tables. See S. Cacioppo & Cacioppo (2015) for details, tables, and figures. '''Global Field Power (GFP) metrics: ''' Users interested in changes in magnitude (rather than changes in topography only) will find this GFP information useful. Initially introduced by Lehmann and Skrandies (1980), the GFP is equivalent to the standard deviation of the electrode voltages fora given timeframe (topographic map). As it was done for the RMSE values over the specified baseline interval, a CI is calculated for the GFP values around the mean GFP value over the same specified baseline interval. Meaningful changes in GFP levels are then determined in the same way as for RMSE (See S. Cacioppo et al., 2014 for details). '''Template maps: ''' Finally, the HPMS function allows users to export each template maps and estimate their brain source using Brainstorm tools and head models. For more details about this steps, see [[http://neuroimage.usc.edu/brainstorm/Tutorials|Brainstorm tutorial (steps 21-22)]]. {{{ 3. Bootstrapping function }}} The third function of the CENA are between-subjects and within-subjects bootstrappingprocedures. Typically, one assumes that the series of brain microstates evoked across trials or across participants is homogeneous. This assumption may not be justified, however. '''We therefore implemented a bootstrapping procedure to identify heterogeneities in the timing or number of microstates as well as their representative template maps across analysis trials, runs,or participants.''' This data-intensive analytic approach, made possible by the use of high-performance computing, promises to dramatically improve the spatiotemporal information provided by noninvasive electrical neuroimaging. This CENA function can be performed either within-subjectsor across groups of subjects. * Within-subjects bootstrapping: At each iteration, a unique ERP is “bootstrapped” by a process of random selection from the available trials in a given subject’s EEG recording for a given condition, with the selected trials then averaged to generate an ERP for that subject and condition. * Between-subjects bootstrapping: A pre-processing step must be performed in which each subject’s EEG recordings for a given condition are reduced to a within-subject ERP by averaging (see S. Cacioppoet al., 2014 for details). The rest of the between-subjects bootstrapping procedure is the same as the within-subjects procedure but instead of performing a random selection from the set of one subject’s available trials, the bootstrapped ERP is generated by selecting from the set of all subjects ERPs for the given condition. In either case, a random sample of r (without replacement) of the available N possibilities is used to generate the bootstrapped ERP. Following each bootstrap ERP generation phase, the resulting ERP (either within- or between-subjects) is subjected to the microsegmentation routine. These steps are repeated a large number of times (on the order of thousands to quadrillions, See S. Cacioppo et al., 2014 for details). == Download Sample Data Set == == Sample Data Set Material and Methods == Participants Participants were 22 volunteers (8 females) with a mean ageof 23.18 (SD = 3.92) years. All were right-handed (Edinburgh Han-dedness Inventory; Oldfield, 1971), and had normal or correctedto-normal visual acuity. None had any prior or current neurological == References == == License ==