Body Posture Modulates Action Perception

Experimental paradigms.

The experiment consisted of three parts completed in a fixed order, spread over two sessions. A bar-grasping task (see below for task descriptions) was performed during the first session only. An action-prediction task was performed during both sessions. To collect behavioral data, the first session took place in a dummy MR scanner identical in appearance to a real MR scanner. Several days later (average, 3.2 d) the action-prediction task was performed in a functional MR scanner.

Bar-grasping task.

The purpose of the bar-grasping task was to familiarize participants with the actions they were about to observe in the prediction tasks later on. The participants were seated at a table with three cradles positioned next to each other at 5 cm distance between adjacent cradles. Participants were instructed to use a power grip to grasp the bar (length, 25 cm; diameter, 2.5 cm; one end black, one end white), positioned horizontally on the middle cradle, and place it on either the left or right cradle according to instructions presented on a screen. Instructions involved both a direction (i.e., whether to place the bar on the left or right cradle) and a goal orientation of the bar (i.e., where the white and black ends of the bar should point).

Some actions required a translation of the bar from the middle cradle to the left or right cradle (16 trials). Other actions required an additional clockwise or counterclockwise rotation of the bar by 90° (16 trials) or 180° (16 trials). All actions were performed using the right hand, and participants were free to choose whether to use an overhand or underhand power grip when grasping the bar. Task duration was ∼15 min.

Action-prediction task.

Participants performed the action-prediction task both outside and inside the MR environment. First, they performed the task in a dummy MR scanner, where we collected behavioral data concerning their predictions on the observed actions. In the second session, the participants performed an adapted version of the task in the MR scanner, where we measured BOLD responses. Below we describe the task in general, followed by a description of the aspects that differed between the two versions of the task.

In the action-prediction tasks (Fig. 1), participants watched short videos of actions while they were asked to predict the goal state of the observed actions as quickly as possible. The stimulus videos lasted ∼2 s. In each video, an actor sitting at a table grasped and moved a bar with his right hand to one of the two cradles. Each video started with a static image of the actor in a rest position with his right hand on the table and his left hand out of view (below the table, on the actor's lap). After a variable delay (250–500 ms), the video started, showing the actor moving his right arm to grasp the bar with either an overhand or an underhand grip. Subsequently, the actor moved the bar to the left or right cradle, using either a rotation or translation movement. It has been shown that participants choose between different grip configurations depending on the action goal (Table 1) (Zimmermann et al., 2012). Namely, translation actions are more likely to be executed with an overhand grip; rotation actions to the left (actor's perspective) are more likely to be performed with an underhand grip of the bar; and rotation actions to the right are more likely to be performed with an overhand grip of the bar. We refer to these action preferences as low-frequency and high-frequency grip strategies. The set of videos used in this study displayed combinations of rest posture (overhand, underhand), grasp posture (overhand, underhand), initial bar orientations (black end on the left or on the right), movement direction (left, right), and action types (translation, rotation) in an equiprobable distribution, including both highly frequent and less frequent grip strategies. Time until the bar was grasped (∼800 ms) and total duration of the grasping movement (∼1600 ms) were standardized across trials. Videos were stopped when the goal was achieved (i.e., the actor's hand rested on the bar in its final configuration) and the last frame was shown until 2 s from video onset were elapsed.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

Action prediction task. A, B, D, E, In each trial, participants were shown videos of an action (A, D: eight representative still frames) that involved either a bar translation (schematically illustrated in B) or a bar rotation (E). C, F, Participants were lying in a scanner while the spatial relation between the posture of their right hand and the start/goal posture of the observed action was manipulated. Participants used their left hand to indicate their prediction of the goal state of the observed action when required (100% of the trials during the behavioral session, 10% of the trials during the fMRI session).

Participants were asked to predict the goal state of each observed action as quickly as possible. Goal state was defined as the final orientation of the bar on the cradle to which the actor moved the bar. Therefore, for each trial there were four possible goal states (black bar end “pointing” up, left, down, or right). Participants indicated their decision using one of four buttons on a button box held in their left hand. Each button was assigned to one final state, defined as the bar orientation on the target cradle (white end pointing up, left, down, or right), irrespectively of the movement used to achieve that final state. The mapping between final states and buttons was constant throughout the experiment. The mapping was displayed during practice and during breaks between trials.

During the task, we manipulated the arm posture of each participant's right arm (Fig. 1). Participants could either have their arm in a prone posture (i.e., palm down), or in a supine posture (i.e., palm up), lying to the right side of their body on the scanner table. Posture was changed after every block of nine trials. The posture manipulation resulted in different patterns of congruency between participants' own arm posture and the observed arm posture(s) in the videos. During translation trials, participants' posture could either be “overall congruent” or “overall incongruent” with the observed action (because start posture and goal posture are the same for these actions). During rotation trials, the participant's posture could either be in a “goal-posture congruent” state or in a “goal-posture incongruent” state. After each arm posture change instruction, there was a short break (5 s) to allow for arm repositioning.

Participants engaged in a total of 432 trials. On average, ∼11% of trials were filler trials. In these trials, the bar was placed vertically rather than horizontally on a cradle. These trials, in which the bar was rotated 90° (instead of 0° or 180° as occurring during the experimental conditions), were introduced to increase the number of possible observed movements and reduce predictability. These filler trials were excluded from subsequent analyses. Of the remaining trials (N = 372), half were translation trials (N = 186). The other half were rotation trials. In each group, half of the trials (N = 93) were goal-posture congruent (rotation trials) or overall congruent (translation trials), and half were goal-posture/overall incongruent. Sessions were divided into six blocks of 72 trials each, with self-paced rest breaks between blocks. Trials were presented in pseudorandom order, such that each block consisted of the same number of trials of each condition, and the same action was not presented twice in a row.

The goal of the prediction task performed in the dummy scanner was to examine whether postural congruency affected decision speed on the prediction task. Therefore, in this session, participants were asked to give a response as quickly as possible after they inferred the action goal. The video was stopped when the subjects pressed the button to indicate that they could predict the action goal. The intertrial interval (ITI) varied between 0.5 and 1 s. Before the task, participants practiced the task until they could correctly predict 8 of 10 consecutive trials, with a reaction time <2 s. The behavioral session lasted ∼40 min.

During functional imaging, we were interested in how postural congruency affected neural responses during the prediction task, while avoiding any motor preparation processes related to responding (i.e., button presses). Therefore, participants were probed to respond only to a small number of “catch” trials (10% of all trials), during which the stimulus video was replaced by a green exclamation mark at an unpredictable moment during the video, between 1000 and 1500 ms after stimulus onset. Participants then, using one of four buttons, had to choose the likely goal of the observed action. These catch trials (as well as other trials where participants mistakenly pressed a button) were modeled separately in the fMRI analysis. The ITI varied between 2 and 4 s. Before the fMRI session, participants engaged in a number of practice trials until they could correctly respond to 8 of 10 consecutive catch trials within 2 s. During the neuroimaging session, eye movements were measured using an MR-compatible infrared camera (MRI-LR, SensoMotoric Instruments). Muscle activity of participants' right forearms (approximately above musculus pronator teres and musculus supinator, to optimally detect pronosupination of the forearm) was measured using an MR-compatible EMG system (Brain Products) and silver/silver-chloride (Ag/AgCl) electrodes (Easycap). The fMRI session lasted ∼55 min.

EBA localizer task.

As detailed in the introduction, we wanted to test for the presence of posture congruency effects in the EBA. To functionally localize the EBA we used a set of previously validated stimuli (http://pages.bangor.ac.uk/∼pss811/page7/page7.html). This set consisted of 20 pictures of human bodies without heads and 20 pictures of chairs. Stimuli were presented in an alternating blocked design with stimuli presentation time of 300 ms on and 450 ms off, and 20 stimuli per block. Two stimuli of each block were presented twice in succession. Participants were instructed to detect stimulus repetitions (1-back task) to ensure attention to the stimuli. To prevent low-level adaptation, the location of each stimulus on the screen was slightly shifted at random. The functional localizer took ∼10 min and was administered after the prediction task was completed.

Analysis of behavioral data.

We obtained the time required to predict the goal state of observed actions [prediction time, (PT)] and error rate from the button box responses. Trials with prediction times exceeding 2.5 SDs above a participant's condition mean were removed from the analysis (on average, 1.7% of the trials were removed by this procedure). Mean PTs were computed from all remaining, correct responses. Given the low error rate (7.5%), we did not analyze error trials.

PTs were defined as the time elapsed between the first video frame when the actor grasped the bar and the moment the participant pressed a button. We investigated the influence of three task-related factors on PT. The effect of action complexity was assessed by comparing PTs during translation actions with PTs during rotation actions. To probe the orthogonal effect of action frequency on performance, we compared PTs of actions performed with high-frequency and low-frequency grip strategies. Finally, we assessed the effect of postural congruency during translation and rotation actions on PTs. For translation actions, we compared PTs for translation trials with overall congruent and overall incongruent body posture. For rotation actions, we compared PTs for rotation trials where participants' own posture was either congruent or incongruent with the goal posture of the observed action.

We used two-tailed paired-sample t tests for all comparisons on behavioral data. Comparisons that exceeded t values corresponding to p values <0.05 were considered significant.

To assess performance during the fMRI version of the prediction task, we analyzed the error rate during the catch trials as a function of viewing duration (i.e., the time before video playback was stopped and the catch trial signal was presented). We calculated the error rate for trials depending on viewing duration in bins of 100 ms. Note that we cannot calculate PTs during the fMRI version of the prediction task, since the decision moment was imposed by the experimenter, rather than the participant.

Eye movement and EMG data.

To regress out potential interpretational confounds related to cerebral effects of eye and muscle movements during the action-prediction task, regressors describing eye movement and EMG activity recorded during the fMRI session of the prediction task were included in the first-level fMRI analysis. For eye movements, we computed trajectory length and number of eye blinks for each MR volume. For EMG activity, we computed the root mean square (RMS) activity for each MR volume. These eye-movement and EMG time series were included as additional nuisance regressors in the first-level analysis of imaging data (see below).

The eye-movement recordings were also used to compare eye movements between conditions (see analysis of behavioral data) by segmenting the recordings into trials and time-locking each segment to video onset.

Image data acquisition.

We used a 3 T Trio MR-scanner (Siemens) with a 32-channel head coil for signal reception to acquire whole-brain T2*-weighted multiecho echo-planar images (TR, 2070 ms; TE(1), 9.4 ms; TE(2), 21.2 ms; TE(3), 33.0 ms; TE(4), 45.0 ms; voxel-size, 3.5 × 3.5 × 3.0 mm; gap size, 0.5 mm) during all functional scans. For each participant, we collected ∼1400 volumes for the prediction task and 180 volumes for the EBA localizer. The first 30 volumes of each scan were used for echo weighting (see Imaging data analysis) and were discarded from the analysis. This also ensured signal equilibration of T1. Anatomical images were acquired with a T1-weighted MP-RAGE sequence (TR/TE, 2300/3.03 ms; voxel size, 1.0 × 1.0 × 1.0 mm) after the EBA localizer task.

The head of each participant was carefully constrained using cushions on both sides of the head. Participants were instructed to remain as still as possible during the experiment. For additional somatosensory feedback on head movements, the forehead of each participant was taped, with tape extending to both sides of the head coil. Data inspection showed that no head movements of participants ever exceeded 2 mm.

Imaging data analysis.

Imaging data were analyzed using MatLab (MathWorks) and SPM8 (Wellcome Department of Cognitive Neurology). First, functional images were spatially realigned using a sinc interpolation algorithm that estimates rigid body transformations (translations, rotations) by minimizing head movements between the first echo of each image and the reference image (Friston et al., 1995). Next, the four echoes were combined to form a single volume. For this, the first 30 volumes of each scan were used to estimate the best echo combination to optimally capture the BOLD response over the brain (Poser et al., 2006). These weights were then applied to the entire time series. Subsequently, the time series for each voxel were temporally realigned to the acquisition of the first slice. Images were normalized to a standard EPI template centered in Talairach space (Ashburner and Friston, 1999) by using linear and nonlinear parameters and resampled at an isotropic voxel size of 2 mm. The normalized images were smoothed with an isotropic 8 mm full-width-at-half-maximum Gaussian kernel. Anatomical images were spatially coregistered to the mean of the functional images and spatially normalized by using the same transformation matrix applied to the functional images. The ensuing preprocessed fMRI time series were analyzed on a subject-by-subject basis using an event-related approach in the context of the general linear model.

For each trial type, square-wave functions were constructed with a duration corresponding to the stimulus duration and convolved with a canonical hemodynamic response function (hrf) and its temporal derivative (Friston et al., 1996). Additionally, the statistical model included 34 separate regressors of no interest, modeling catch trials and false alarms, residual head movement-related effects by including Volterra expansions of the six rigid-body motion parameters (Lund et al., 2005), and compartment signals from white matter, cerebrospinal fluid, and out-of-brain regions (Verhagen et al., 2008). Volterra expansions consisted of linear and quadratic effects of the six movement parameters for each volume, and included temporal derivatives. Finally, to covary out any potential confounding effects of eye and muscle movements, hrf-convolved metrics of eye movements (path trajectory and number of eye blinks) and muscle activity data were included as additional regressors of no interest.

Parameter estimates for all regressors were obtained by maximum-likelihood estimation, using a temporal high-pass filter (cutoff, 128 s), modeling temporal autocorrelation as an AR(1) process. Linear contrasts pertaining to the main effects of the functional design were calculated based on parameter estimates of canonical hrfs.

For analysis of the experimental task, we looked at the same comparisons as those we looked at during the behavioral prediction task, including those related to action complexity, frequency, and postural congruency. Contrasts of the parameter estimates for these comparisons constituted the data for the second-stage analyses, which treated participants as a random effect (Friston et al., 1999). Contrasts were thresholded, if not otherwise specified, at p < 0.05 after familywise error (FWE) correction for multiple comparisons at the voxel level. Anatomical details of significant clusters were obtained by superimposing the structural parametric maps onto the structural images of the MNI template. Brodmann areas (BAs) were assigned based on the SPM anatomy toolbox (Eickhoff et al., 2005).

Apart from a whole-brain search for significant differences, we specifically focused on two predefined regions of interest (ROIs; spherical, radius: 5 mm). The first ROI consisted of individually localized EBA (on the basis of a separate EBA localizer session), to test whether action-prediction effects were visible in this area, which is sensitive to observation of body parts, as has been previously suggested (Downing et al., 2001). The second ROI was a region in the IPS, which has been found to be sensitive to body-posture manipulations during planning of goal-directed actions. Here we used previously published stereotactic coordinates (MNI: −22, −60, 58; Zimmermann et al., 2012) to extract the difference in brain activation for contrasts related to posture congruency.

Effective connectivity analysis.

After having identified that regions in parietal and dorsal premotor cortex and the left EBA are more strongly involved in predicting goals of low-frequency actions compared with high-frequency actions (see Results), we assessed whether there were changes in effective connectivity between EBA and parietal or premotor regions as a function of action frequency.

More specifically, we expected an increased connectivity between EBA and parietal/premotor cortex during prediction of unlikely (i.e., low frequency compared with high frequency) observed actions, under the hypothesis that EBA forms predictions about potential goal states during observation of another agent. Moreover, it has previously been shown that predictions about observed actions are influenced by one's own, previously executed actions (Cattaneo et al., 2011). Therefore predictions may also be influenced by the likelihood with which the observed action would be chosen to reach a particular goal state in general. With accumulating evidence, predictions in EBA can be updated, and this updated information may be forwarded to the parietal or precentral regions to inform the action plan. This would result in the hypothesized increase in connectivity.

To analyze changes in connectivity, we performed a psychophysiological interaction (PPI) analysis (Friston et al., 1997). PPI analysis tries to model regionally specific responses based on an interaction between a psychological factor and physiological activity of one specific (seed) brain region. Here, the analysis was set up to test for differences in connectivity (measured by correlation strength between activity of 2 areas) between left EBA and all remaining brain areas, depending on the grip strategy used in the observed video (low frequency or high frequency). To define activity in EBA, we used the peak location of the left EBA from the independent localizer task as a starting point. We drew a 5-mm-radius sphere around that voxel and extracted the first eigenvalue of voxels in this sphere that showed a relative increase in BOLD signal during observation of low-frequency actions (first level, p < 0.05 uncorrected). First, a PPI analysis was performed for each subject. Then, contrasts of parameter estimates for the interaction term constituted the data for the second-stage PPI analysis, treating participants as a random effect. Finally, contrasts were corrected for multiple comparisons by applying FWE correction at the cluster level (p < 0.05) over the search volume (whole brain, IPS-ROI), on the basis of an intensity-based voxelwise threshold of p < 0.001 uncorrected.