Brain activity during a motor learning task: An fMRI and skin conductance study

Abstract

Measuring electrodermal activity (EDA) during fMRI is an effective means of studying the influence of task‐related arousal, inferred from autonomic nervous system activity, on brain activation patterns. The goals of this study were: (1) to measure reliable EDA from healthy individuals during fMRI involving an effortful unilateral motor task, (2) to explore how EDA recordings can be used to augment fMRI data analysis. In addition to conventional hemodynamic modeling, skin conductance time series data were used as model waveforms to generate activation images from fMRI data. Activations from the EDA model produced significantly different brain regions from those obtained with a standard hemodynamic model, primarily in the insula and cingulate cortices. Onsets of the EDA changes were synchronous with the hemodynamic model, but EDA data showed additional transient features, such as a decrease in amplitude with time, and helped to provide behavioral evidence suggesting task difficulty decreased with movement repetition. Univariate statistics also confirmed that several brain regions showed early versus late session effects. Partial least squares (PLS) multivariate analysis of EDA and fMRI data provided complimentary, additional insight on how the motor network varied over the course of a single fMRI session. Brain regions identified in this manner included the insula, cingulate gyrus, pre‐ and postcentral gyri, putamen and parietal cortices. These results suggest that recording EDA during motor fMRI experiments provides complementary information that can be used to improve the fMRI analysis, particularly when behavioral or task effects are difficult to model a priori. Hum Brain Mapp, 2007. © 2007 Wiley‐Liss, Inc.

INTRODUCTION

The autonomic nervous system (ANS) innervates the skin and produces measurable electrodermal activity (EDA). This activity manifests in the sweat glands, causing the secretion of water and salts, and leads to changes in the skin conductance. These changes in EDA may occur slowly over time (tens of seconds), as defined by changes in the skin conductance level (SCL), or they may be more rapid changes (seconds), as defined by the skin conductance response (SCR, also known as the galvanic skin response). The temporal features of the SCR have been well characterized, with an onset latency of ∼1.5 s and a rise in skin conductivity thereafter that is proportional to the degree of synchronization of sweat gland secretions [Lim et al., 2003]. When measured during the performance of specific tasks, this measure of ANS activity provides an index of task‐related arousal that is linked to task difficulty or effort.

Interestingly, aspects of the blood oxygenation level dependent (BOLD) fMRI hemodynamic response [Ogawa et al., 1992] are similar to features described earlier during EDA recordings. The onset of the hyperemia period, for instance, is delayed with respect to the stimulus by an amount of 1–2 s. Next, the BOLD time‐to‐peak is ∼4–5 s, which is comparable to the times‐to‐peak of 3.4 and 5.1 s for SCR changes due to auditory stimuli when measured from the fingers and toes, respectively [Lim et al., 2003]. In cases when EDA and BOLD fMRI data correlated in time during a fear conditioning task [Cheng et al., 2003] and a gambling task [Patterson et al., 2002], EDA was used as a model to derive statistical parametric fMRI maps.

Previously, skin conductance information has been used in conjunction with fMRI in studies of cognition or emotion that elicit phasic changes in autonomic arousal [Shastri et al., 2001; Williams et al., 2000, 2004, 2005] but the applicability and utility of the EDA recordings is not restricted to such experiments. In the present study, EDA recordings were measured concurrently with fMRI to investigate how changes in the autonomic nervous system affect the fMRI BOLD signals associated with performing a motor task. For example, one aspect of the ANS, attention, has been inferred to affect motor and sensory signals during fMRI [Johansen‐Berg et al., 2000]. The purpose of the present study is to demonstrate that strong correlations can exist between BOLD fMRI motor‐related signals and EDA.

During difficult motor tasks, it is hypothesized (1) that EDA provides a useful surrogate measure of task difficulty and sense of effort, and (2) that features in EDA recordings can be used to improve fMRI data analysis, such as by helping to account for heterogeneity of the fMRI data within the measurement session and across participants. The intent of these hypotheses is not to replace conventional hemodynamic modeling in fMRI. Rather, support for these hypotheses will illustrate that monitoring of relevant physiological signals can help to explore the rich features in fMRI datasets. To test these hypotheses, a novel, difficult motor task was developed that involves fractionated finger movements with the nondominant hand. Similar to previous demonstrations [Grafton et al., 2002], this task is likely to elicit motor sequence learning effects that vary among participants. Furthermore, to explore the EDA and fMRI temporal features that are associated with performance of this task in detail, univariate analysis using the general linear model (GLM) approach and multivariate analysis using partial least squares (PLS) [McIntosh et al., 2004a, b] were both performed.

There is an additional clinical motivation behind the current preliminary study. Pathologies that result in movement disorders often impact the ANS. For instance, in assessing neurological impairment due to stroke, EDA physiological data would be beneficial to monitor the ANS as the paretic limb is engaged. One of the earliest fMRI stroke studies described that enhanced attention could be a compensatory strategy during movement of the affected hand [Pineiro et al., 2001]. There is strong potential to use EDA recordings to provide a supplementary measurement that is sensitive to such a strategy and that subsequently informs interpretation of the associated fMRI data. Similarly, recent stroke research suggests that as motor function improves during recovery, the attention demands and the corresponding activity from attention networks decrease [Johansen‐Berg et al., 2002; Ward et al., 2003]. Data analysis strategies identified in the present study hopefully will permit optimal integration of EDA recordings with fMRI data sets, and will reduce the possibility of misinterpreting fMRI results.

METHODS

Participants

Twelve healthy right‐handed adults participated in this study (mean age: 29 ± 11 years, 7 females). Functional MRI experiments were performed with the informed consent of the participants and approval from the research ethics board at Sunnybrook Health Sciences Centre. All participants were naÿve of the task prior to commencing the experiment. Participants received instructions and a visual demonstration of the task, immediately after anatomical MRI. Two participants were excluded from subsequent analysis. One subject found the task very easy and elicited no task‐related EDA, whereas skin conductance measurements from another participant were unsatisfactory owing to poor electrode contact with the skin.

Task

Participants performed sequential fractionated finger movements in a block design that consisted of 20 s of the task condition (7 blocks) and 20 s of rest (7 blocks). Participants were cued visually to perform adduction then abduction of digits 3 and 4, followed by adduction and abduction of digits 2 and 4, in sequence with the left hand as quickly as possible (Fig. 1A). This motor task is based on a paradigm that has been developed to investigate perceived sense of effort during motor tasks [Giannoylis et al., 2003]. After fMRI, participants were asked to rate their initial level of task difficulty on a scale from 0 (easiest) to 10 (hardest), in accordance with the Borg rating scale for exertion [Borg, 1985]. Participants were instructed to keep their right hand perfectly still throughout the experiment to avoid introducing movement artifacts in the EDA data.

Figure 1.

(A) Sequential fractionated finger movements during one task block. (B) Task waveform (light gray) and the corresponding hemodynamic waveform (dark gray). Average EDA data (black; N = 10) show temporal features that are not depicted by hemodynamic waveform. (C) Individual EDA data over time, presented for each participant.

Skin Conductance Measurements

An initial exfoliation was performed using sand paper on the palmar surface of distal phalanges of digits 2 and 3 of the right hand, followed by alcohol wipes to clean the skin surface. Electrodermal electrodes designed for placement on fingers (TSD 203, Electrodermal Response Electrode Set, BIOPAC Systems, Goleta, CA) were filled with conductive paste and fastened onto digits 2 and 3 using Velcro straps. In preliminary experiments, the EDA system was tested and baseline skin conductance values were measured for 10–30 s during fMRI. With the goal of minimizing gradient‐induced artifact in the electrodermal recordings, the electrode cables used were cut, shortened, wound as a twisted pair and altered to connect to a custom preamplifier using Touchproof safety plugs (Plastics One, Roanoke, VA). The EDA preamplifier box was positioned on top of a cushion and covered with a cotton fabric in an attempt to reduce the effects of vibrations induced by the gradients on the electrodermal recordings. The circuit design was based on the Wheatstone bridge preamplifier described elsewhere [Shastri et al., 2001] with a system gain of 1,000. After preamplification, voltage values were converted to optical signal and transmitted using a multiplexer originally developed for a parallel eight channel electromyography system [MacIntosh et al., 2006]. The optical signals were passed through the waveguide using fiber optic cable. Outside the magnet room, the signals were demultiplexed, further amplified, and digitized at 1,000 Hz using LabVIEW (National Instruments, Austin, TX) software written in the laboratory. A calibration procedure using known resistors (500 kΩ, 1 MΩ, 2 MΩ) was used to determine skin conductance levels from the voltage measurements. Using Matlab (Mathworks, Natick, MA) software, EDA data were subsequently band‐pass filtered (0.01–0.3 Hz) using a fourth‐order Butterworth filter. Filtered EDA time series for each participant was resampled to match the temporal resolution of the image acquisition and used in the fMRI data analysis in two ways. First, the full EDA time series (310 s) were used as a complimentary fMRI model waveform, with the intent of contrasting activation maps produced using the hemodynamic model. Second, the minimum and maximum EDA values for each of the task block conditions were used to generate peak‐to‐peak EDA changes (Fig. 2). These physiological data were used as “behavioral metrics” in the multivariate analysis described below.

Figure 2.

Normalized peak‐to‐peak EDA versus task block averaged across all participants (N = 10). Error bars denote standard deviation.

MRI Acquisition

Imaging data were collected using a 3.0 T whole‐body scanner (General Electric Medical Systems, Waukesha, WI). Axial anatomical images were collected using a three‐dimensional T₁‐weighted spoiled gradient recall echo sequence with the following parameters: TI/TR/TE/FA = 300 ms/7.0 ms/3.1 ms/15°, voxel dimensions of 0.86 mm × 1.15 mm × 1.4 mm (220‐mm field‐of‐view (FOV), 256 × 192 × 128). Next, participants underwent a single 6‐min spiral fMRI scan with 20‐cm FOV (in‐plane resolution 3.1mm × 3.1 mm), TR/TE/FA = 2000 ms/30 ms/70°, spanning 25 slices that were 5‐mm thick. Whole‐brain axial T${}_{\text{2}}^{\text{*}}$‐weighted BOLD contrast images were acquired using single shot spiral‐in then spiral out k‐space trajectories (spiral IO). As is described elsewhere [Glover and Thomason, 2004], spiral IO data with signal‐weighted averaging increases the signal‐to‐noise ratio significantly compared to conventional spiral out sequences, without sacrificing temporal resolution. The first five fMRI volumes (i.e. 10 s) were excluded to ensure the magnetization had reached steady state in all subsequent analyses.

fMRI Data Analysis

Univariate analysis

Processing of functional images was first done using AFNI software [Cox, 1996]. Retrospective coregistration was performed and motion plots indicated that head motion did not exceed 1 mm in translation throughout. Images were temporally smoothed using a 3‐point median filter, spatially smoothed using a Gaussian full‐width half‐maximum of 6 mm and detrended to remove linear and quadratic terms. The resultant data were analyzed in several different ways to assess consistency. Activation maps were generated using a univariate general linear model (GLM) and two model waveforms: (1) a standard hemodynamic model, created by convolving the task waveform with a hemodynamic response function, HEMO*TASK (of the form t ^8.6 exp(−t/0.547) [Cox, 1996]), and (2) an EDA model consisting of the normalized and filtered electrodermal time series for each individual participant. Activation maps for each participant were inspected, and then converted from t‐values to Z‐scores to facilitate comparison of the histogram distributions. FMRI data were converted to Talairach coordinate space to facilitate a mixed‐effect group ANOVA from which the main effects and interaction effects were observed. The ANOVA was performed with temporal smoothing and detrending both included and both excluded to ensure that fMRI results were not confounded by temporal processing steps.

Percent BOLD signal changes were also calculated by correlating the fMRI and EDA data during two distinct periods, an early period corresponding to the first two task blocks and a late period corresponding to the last two task blocks. A voxel‐wise group t‐test was performed to compare early versus late % BOLD signal changes The results of this group t‐test were compared visually with the results of the mixed‐effects ANOVA contrast map that compared the activation maps from the hemodynamic model with the activation maps from the EDA models. Finally, regions‐of‐interest (ROIs) analysis was performed to produce average % BOLD signal changes during the early and late conditions for functional neuroanatomical regions, as circumscribed by the brain atlas provided in AFNI, and shown to be active on the basis of the main effect group maps of brain activity obtained by ANOVA. The ROIs analyzed in this manner were larger in volume than the regions identified from activation maps, thus potentially providing more statistical power to detect differences between early and late conditions.

Multivariate analysis

To characterize the relationship between BOLD fMRI and EDA time series, behavioral PLS analysis was also performed [McIntosh et al., 1996]. The PLS software is freely available and has recently been described in the context of fMRI data analysis [McIntosh et al., 2004a]. A brief summary of the technique is given below. EDA data were normalized for each participant, such that the data ranged from 0 to 1, to account for the inherent differences in skin conductance across participants. Peak‐to‐peak changes in EDA were used to represent behavioral metrics during the seven task blocks according to the calculation outlined earlier.

The behavioral PLS technique attempts to reduce the dimensionality of the correlation data matrix between voxel time series and behavioral data using singular value decomposition (SVD). The result of the data reduction operation produces latent variables (LVs) that can be ranked according to the amount of the variance in the data that they explain (analogous to a principal component analysis). Visualizing the significant brain regions identified is achieved by selecting a threshold for the bootstrap ratio (BSR). As is described elsewhere [McIntosh and Lobaugh, 2004b], the BSR is approximately equivalent to a Z‐score. Brain scores are supplementary aggregate measures that are used to summarize the PLS results. They are calculated by projecting each of the LVs onto the original data and indicate how strongly individual participants express the patterns expressed by each LV [McIntosh et al., 2004a]. Thus, a brain score was plotted for each participant at each condition, producing a scatter plot that represents the brain‐behavior relationship for each task block. Five‐hundred permutations were performed to determine the statistical significance of each LV by randomly resampling the order of the task blocks for each participant. Next, to assess the reliability of each LV, one‐hundred bootstrap resampling iterations were performed to assess the contribution that each task block has on the LV maps. The bootstrap enables calculation of confidence intervals for the correlation value between brain and behavior for each task block by resampling the data with replacement; effectively one row of the design matrix consisting of all participants at a given task block is replaced by another. Further methodological and theory description of the PLS technique are discussed and reviewed elsewhere [McIntosh et al., 1996; McIntosh and Lobaugh, 2004b].

RESULTS

Skin Conductance Responses

All 10 participants experienced difficulty in performing the motor task with an average Borg rating of 7 ± 1.4 s.d. (range of metric: 0–10). Some participants reported that the task became less difficult as the experiment progressed, while others claimed to struggle throughout. Figure 1B illustrates the task and rest versus time, the hemodynamic model, and the group‐averaged EDA versus time for comparison. Figure 1C shows the EDA waveforms for each participant. The time to initial EDA peak was estimated to be 5.6 ± 0.88 s (s.d.) after the visual cue across participants (range: 4.7–6.9 s). The average EDA waveform shows rise‐times at the onset of each task block that appear similar to that of the modeled BOLD signal. However, there are also clear distinctions between the hemodynamic model and the average EDA waveform. While the fMRI hemodynamic model remained approximately constant over a single task block, the EDA attenuated rapidly. Furthermore, the amplitude of the peak‐to‐peak EDA showed a decreasing trend over all the task blocks, as is reflected in the normalized, average data shown in Figure 2. Comparing the peak‐to‐peak EDA values from task block 1 with the remaining 6 task blocks produced a significant difference in all cases (e.g. P _{task1–task2}(t = 4.17, DOF = 9) < 0.0024). The average Pearson correlation value between the two model waveforms for the participants was nonsignificant, r = 0.30 ± 0.17 (P = 0.19, range: 0.04–0.55). The average correlation value reflects the waveform similarities and differences just outlined, while the range in correlation shows the heterogeneity across participants (Fig. 1B). Seven of the participants exhibited decreasing peak‐to‐peak EDA amplitude over the task blocks, but showed variation in the maximum amplitude and rate of decrease. The remaining three participants exhibited no decrease over the successive task blocks. Finally, a nonsignificant correlation was found (r = −0.18, P = 0.32) when attempting to correlate the variance in the EDA data with the reported Borg scores.

Univariate fMRI Analysis

Individual fMRI activation maps using the HEMO*TASK model produced activation in brain regions that are known to be implicated in motor execution [Wu et al., 2004]. These regions include: the bilateral pre‐ and postcentral gyri, cingulate gyrus, putamen, secondary somatosensory cortex, precuneus, right inferior frontal gyrus, and the left cerebellum. Similarly, the activation maps from the EDA model produced activation in the bilateral pre‐ and post‐central gyri, cingulate gyrus, insula, cerebellum, and the right putamen. Using the EDA model, smaller activation was also found in the medial frontal gyrus and the ventromedial prefrontal cortex. There was a significant overlap between the maps produced from these two models. This was evident from a nonsignificant difference between the Z‐score distributions (P = 0.31). Given that these results were unremarkable, activation maps and histograms are not shown.

However, significant differences were observed at the next level of scrutiny. Figure 3 shows fMRI and EDA results from a representative participant. The EDA for this participant was plotted together with the HEMO*TASK model to illustrate differences in the time series (Fig. 3A). For comparison, the BOLD signal for the right insula (R ins) and the right primary motor cortex (R M1) are plotted (Fig. 3B), illustrating differences analogous to those observed in the two models. Fig. 3C shows the brain regions identified when using the EDA data for this participant (N1) as the fMRI model.

Figure 3.

FMRI and EDA data from a representative participant (N1). (A) Model waveform, HEMO*TASK and EDA₁, are shown for this participant. (B) Voxel time courses are shown for two brain regions, the right Insula (ins) and the right primary motor cortex (M1). Standard deviation values represent the variance in the % BOLD change time course for three contiguous voxels. (C) An activation map using the participants EDA recordings as the model waveform shows significant brain regions identified (P < 0.001).

The group voxel‐wise t‐test comparing % BOLD signal change during early and late task blocks is shown in Figure 4A. Significant differences were found in the insula, cingulate gyrus, pre‐ and postcentral gyri, and thalamic regions (P < 0.005, uncorrected). The group ANOVA contrast map, comparing EDA and HEMO*TASK models, served as a complimentary test. A large cluster of activation was found in the bilateral insular regions when performing the contrast EDA > HEMO*TASK (P < 0.005, uncorrected; Fig. 4B), whereas the opposite contrast, HEMO*TASK > EDA, produced no significant voxels. The brain regions identified are similar to those shown in Figure 4A. These findings were robust, as similar brain regions were found when the group ANOVA was performed (1) without temporal smoothing or detrending of the fMRI time series data, as well as (2) using within subject contrasts, i.e. [EDA, HEMO*TASK] = [1, −1]. The Z‐scores and Talairach coordinates of maximum % BOLD signal changes associated with Figure 4 are shown in Table I. Finally, ROI analysis revealed that all brain regions implicated in the motor network showed a significant decrease in % BOLD signal changes during late task blocks when compared with the changes observed in early task blocks (P < 0.05, Fig. 5).

Figure 4.

(A) Results of a group t‐test comparing Early versus Late % BOLD signal change. The thalamus, cingulate (cing) and the insula (ins) exhibited strong differences between the conditions (Z > 3.2, P < 0.005). (B) Mixed‐effects ANOVA results contrasting the EDA model with HEMO*TASK model (Z > 3.7, P < 0.005).

Table I.

Summary of brain regions found to be significant according to the three group analyses

	Talairach coordinates
	Z‐score	X (mm)	Y (mm)	Z (mm)	Volume (μL)
ANOVA model contrast
R Insula (anterior)	4.806	39	5	−2	3512
L Insula (posterior)	4.456	−35	19	−4	2456
t‐test comparison: early vs. late contrast
B Thalamus	3.7485	7	−17	−2	1280
B Cingulate gyrus	3.6483	5	−11	44	800
R Insula (anterior)	3.8933	39	3	14	584
R Insula (posterior)	3.9254	45	−25	16	472
R Precentral gyrus	3.7672	51	−27	44	448
L Insula (posterior)	3.7016	−53	−23	14	232
R Insula (anterior)	3.6319	47	15	6	224
L Precentreal gyrus	3.3628	−31	−25	44	168
	Positive correlation
Behavioral partial least squares	BSR	X (mm)	Y (mm)	Z (mm)	Volume (μL)
R Insula (anterior)	7.7422	42	0	9	567
L Superior parietal	7.3342	−27	−57	45	3699
R Middle frontal gyrus	6.0816	39	42	27	675
R Precuneus	5.933	24	−57	30	1242
R Middle frontal gyrus	5.754	24	12	24	648
R Postcentral gyrus	5.2252	27	−30	54	702
	Negative correlation
L Precuneus	−8.5969	−6	−69	21	4185
L Putamen	−7.5774	−21	18	9	999
R Thalamus	−7.5159	3	−12	12	729
L Postcentral gyrus	−5.1999	−45	−21	54	1269
R Superior frontal gyrus	−4.7944	9	15	60	1215

Z‐scores are shown for ANOVA tests and bootstrap ratios, BSR (BSR > 3.3, P < 0.001), are shown for PLS, with the maximum intensity coordinates in Talairach coordinate space and the cluster size (in μL). (B, bilateral, L, left, R, right; note: contiguous bilateral regions produced one total volume).

Figure 5.

ROI Analysis comparing the % BOLD change according to the EDA model for early and late task blocks, listed by brain region (B, bilateral; L, left; R, right; R and L Insula: 14 mL, B Cingulate: 52 mL, B Thalamus 14 mL, R and L Precentral Gyrus: 28 mL, R and L Postcentral Gyrus: 20 mL; *P < 0.05, ^§ P < 0.005, **P < 0.0001).

Multivariate fMRI Analysis

Behavioral PLS that used normalized peak‐to‐peak EDA changes for each of the seven task blocks and 10 participants produced one significant LV (P < 0.005), accounting for 31% of the variance in the correlation matrix during the task periods. Brain regions that were positively correlated with task‐related EDA changes were: right dorsolateral prefrontal cortex (DLPFC), right insula, bilateral parietal cortex, right MI (Fig. 6A). A subset of these brain regions was common with the brain regions identified by univariate analysis (Fig. 4, Table I). Brain regions that were negatively correlated with task‐related changes in EDA were: left precuneus, left putamen, left postcentral gyrus, right thalamus, and superior frontal gyrus (Table I).

Figure 6.

(A) Brain regions implicated by the behavioural PLS (P < 0.0027; ins, insula, DLPFC, dorsolateral prefrontal cortex). (B) Scatter plots of brain score versus peak‐to‐peak EDA change for each participant. Plots are shown for task blocks 1 through 7. Strong linear relationships are observed from the third task block onward, as supported by Pearson correlation coefficient value (bottom right).

In Figure 6B, scatter plots are shown representing the brain score for each participant as a function of the normalized peak‐to‐peak EDA for each of the seven task blocks. Pearson correlation values for each scatter plot are shown in bottom right panel in Figure 6B. Task blocks 1 and 2 show large scatter and no statistically significant correlation between brain score and peak‐to‐peak EDA. However, a pattern begins to emerge from task block 3 onwards (i.e. r > 0.57, P < 0.05, for each task block 3–7). Brain score versus the normalized peak‐to‐peak EDA changes for task block 3–7 each showed a strong linear relationship. In essence, the linear relationship between the brain score and normalized peak‐to‐peak EDA changes in these latter task blocks produced the positive brain activations shown in Figure 6A. Participants that continued to show large normalized peak‐to‐peak EDA changes also showed greater activation of the DLPFC, insula and superior parietal regions.

DISCUSSION

This study demonstrates that temporal features in electrodermal activity (EDA) recordings can be used as a surrogate measure of the autonomic nervous system (ANS) to identify brain regions that are strongly engaged during difficult or effortful motor tasks. Examples in the literature illustrate the utility of measuring EDA during fMRI in the context of decision‐making [Critchley et al., 2000; Patterson et al., 2002], biofeedback [Critchley et al., 2001], and fear conditioning [Cheng et al., 2003; Knight et al., 2005; LaBar et al., 1998], the application of EDA to fMRI of a visually cued self‐initiated motor task is novel. Compared to examples where the task pattern was not synchronized with the onset of the EDA [Critchley et al., 2000], by virtue of experimental design, the present study found the EDA model to be comparable to the HEMO*TASK model. Informative distinctions between the EDA and HEMO*TASK models were also observed and thus consistent with anatomically‐derived ROI analysis.

Regions of the brain that showed a correlation with the EDA model were found to be significantly more activated when using the individual EDA recordings as the reference waveform compared to the hemodynamic waveform, and produced large % BOLD changes among early compared to late task blocks. The comparison between early and late task blocks suggests that the reduction in the peak‐to‐peak EDA amplitude is related to either short‐term learning effects, decreases in sustained attention, or the level of vigilance and/or arousal, with recent work providing support for the latter [Nagai et al., 2004]. An extension of this work would be to introduce a motor task that is novel at each successive task block, thereby separate the effects of learning from those associated with attention. A subsequent preliminary experiment that attempted to characterize EDA during sequence learning motor fMRI has been presented in abstract form [MacIntosh et al., 2006b]. The intent of this experiment was to understand better the brain's regional specificity associated with EDA during motor paradigms. For example, one strategy for isolating EDA signals that are specific to the execution of a motor command would be to contrast EDA occurring during the task period with EDA occurring at rest.

The EDA data facilitated multivariate analysis of the fMRI data using PLS. It enabled observation of brain regions additionally implicated in within‐participant differences in behavior, representing a novel way in which EDA data can be used to interpret fMRI data. Such an approach is not possible using conventional rating scales to estimate perceived effort, such as the Borg score [Borg, 1985], given its lack of sensitivity and the impracticality of recording ratings across task blocks. The behavioral PLS results of Figure 6 are dependent, to some extent, on the procedure used to quantify EDA changes. Some exploration of different procedures was conducted, and the peak‐to‐peak metric reported in the Methods section proved to be more sensitive than a metric based on the area under the SCR curves. The peak‐to‐peak procedure produced one statistically significant LV that represented brain regions where % BOLD changes were correlated with the peak‐to‐peak changes in EDA at successive task blocks across all participants.

Heterogeneity in fMRI data during task blocks can occur as a result of a premeditated experimental design, such as changing the level of difficulty or altering the response time during a sensorimotor task. Such design issues can reasonably be accounted for in a GLM analysis. However, task‐related changes may be more inconspicuous, subtle, and heterogeneous across participants, and difficult to model accurately based on a priori information; as is the case in the present study. The healthy right‐handed participants had no motor or neurological impairment and all participants reported a consistently high Borg score for perceived task difficulty (mean 7 ± 1.14). However, this metric was not found to be correlated with any summary statistic, such as standard deviation, in the EDA data. One participant was excluded from analysis because they found the task exceedingly easy. Furthermore, all subjects were able to perform the fractionated finger movements, as was confirmed by visual inspection for each participant. Unfortunately, finger kinematics or electromyographic data were not collected in this study, so it is possible that subjects performed the task differently as the experiment progressed. Based on post‐fMRI interviews with participants this was unlikely to be the case.

Each of the analyses undertaken provided complimentary findings. The strongest effect observed was the insular brain regions that were detected using the univariate, ROI, and multivariate analyses, not surprising since the insula plays an integral role in attention [Nagai et al., 2004]. Furthermore, the insula has been shown to be implicated in the production of SCRs [Knight et al., 2005; Patterson et al., 2002]. Brain regions that were negatively correlated with the EDA data were also found, primarily in brain regions of the ipsilateral hemisphere: left precuneus, left putamen, left postcentral gyrus, and the right thalamus. These brain regions may play a role in sustained activity level of the ipsilateral basal ganglia‐thalamo‐motor loop during self‐initiated movement [Taniwaki et al., 2003]. The one significant LV from the behavioral PLS produced positive brain scores in the right dorsolateral prefrontal cortex (DLPFC), superior parietal lobe and the right insula (Table I). It is well established that the DLPFC is implicated in motor sequence learning [Jenkins et al., 1994; Wu et al., 2004]. In our study, decreases in DLPFC activity occurred over minutes as participants performed the task with what can be interpreted as increasingly automatic movements. The acquisition of fine motor control needed to perform fractionated finger movements occurring over a range of competency levels seems plausible. Intersubject variability may explain why the mixed‐effects ANOVA, whereby participants are treated as random factors, did not produce a similar result. The EDA model main effect produced activation of frontal lobe regions, such as the ventromedial prefrontal cortex (VMPFC), which have been shown to be critical for the generation of SCRs [Nagai et al., 2004; Patterson et al., 2002].

Notable distinctions in statistically significant brain regions were observed. For example, the cingulate gyrus has been shown to be coupled with task difficulty [Paus et al., 1998] and was identified in the (1) ROI analysis, (2) early versus late t‐test, and (3) univariate model comparison, but not in the behavioral PLS. Differing outcomes from these analyses may reflect methodological and philosophical differences between methods. Interestingly, a subsequent task PLS analysis was performed with the intent to resolve some of the discrepancies by closely matching the conditions prescribed by the early versus late t‐test. The task PLS incorporated blocks 1, 2, 6, and 7 and found that the cingulate gyrus and right insula were detected from an LV that explained 56% of the variance during these four task blocks (P < 0.05). Considering the collected results, the present study provides support for the use of parallel hypothesis and data‐driven strategies to better characterize motor fMRI data sets.

One motivation for measuring EDA during motor tasks stems from the need to characterize behavior in the presence of and recovery of neuropathology. Functional MRI of stroke recovery is a pertinent example [Staines et al., 2001]. In longitudinal fMRI studies, EDA could be used to monitor autonomic processes, such as a patient's sense of effort. A recent example found that nodes in the attention network, specifically the insula (in 5 of 8 patients) and the cingulate gyrus (significant across the group of 8 patients), were negatively correlated with recovery after stroke [Ward et al., 2003]. Ward and colleagues also found that stroke patients with more profound motor deficits are more likely to engage attention networks during early stages than when compared to later stages of recovery [Ward et al., 2004]. In conclusion, this study provides support for the use of EDA as a means to monitor the influence of the autonomic nervous system in experiments that attempt to better characterize the spatial and temporal features in motor fMRI datasets.