Participants

Seventeen participants (9 females, mean age 28.17 years, range 21–39) were recruited for the experiment. All participants were right handed, had normal or normal to corrected vision, and no history of neurological disorders. The research study was approved by local ethical committee (Comitato Etico Regione Liguria) and was carried out in accordance with the principles of the revised Helsinki Declaration (World Medical Association General Assembly, 2008). All the participants provided written informed consent.

Apparatus and Procedure

The participants were seated on a height-adjustable chair with the elbow and wrist resting on a table (length = 110 cm; width = 100 cm). In order to guarantee a repeatable start position across participants, they were asked to maintain the forearm pronated, the right arm oriented in the parasagittal plane passing through the shoulder, and the right hand in a semi-pronated position, with the tips of the thumb and index finger on a tape-marked point placed on the working space. A glass bottle was positioned on the table at a distance of about 46 cm from participants’ body midline. Depending on condition, participants were instructed to reach towards and grasp the bottle with one of the following intentions:

1. grasp-to-pour: grasp the bottle to pour some water into a small glass (diameter: 5 cm; height: 8.5 cm) positioned on the left side of the bottle (at a distance of 25 cm);
2. grasp-to-place: grasp the bottle and place it in a cardboard box (length: 17 cm; width: 17 cm; height 12.5 cm) positioned on the left side of the bottle (at a distance of 25 cm);
3. grasp-to-drink: grasp the bottle to drink some water from it.

The immediate spatial demands and the accuracy requirements were kept constant for each intention across trials. The experimenter visually monitored the performance for each trial to ensure participants’ compliance to task requirements. Participants performed two sessions of 30 trials each, in 3 separate blocks of 10 trials for a total of 2 blocks of 10 trials for each of the three intentions. The blocks were pseudo-randomized to ensure that the last block of the first session was different from the first block of the second session. The first session was preceded by a practice session to familiarize participants with the task (12 practice trials; 4 trials for each intention).

Kinematics and Video Recording

A near-infrared camera motion capture system (frame rate, 100 Hz; Vicon System) equipped with nine cameras was used to track the hand kinematics. Each participant was outfitted with lightweight retro-reflective hemispheric markers (4 mm in diameter) placed on the radial aspect of the wrist, the metacarpal joint and the tip of the index finger, the metacarpal joint of the little finger, the trapezium bone of the thumb, and the tip of the thumb. Movements were also filmed from a lateral viewpoint using a digital video camera (Sony Handy Cam 3-D, 25 frames/sec). The video camera was placed at about 120 cm from participant’s hand starting position with the camera view angle directed perpendicularly to the agent’s midline. Video camera position and arrangement were kept constant for the entire duration of the study in order to ensure that only the hand and the bottle were in full view from the beginning up to the end of the movement.

Kinematics Data Processing

After data collection, each trial was individually inspected for correct marker identification and then run through a low-pass Butterworth filter with a 6 Hz cutoff. Kinematics variables of interest were computed for the reach-to-grasp phase of the movement, defined as the phase from reach onset (i.e. the first time at which the wrist velocity crossed a 20 mm/s threshold) to grasp offset (i.e. the time at which the wrist velocity dropped below a 20 mm/s threshold). Within this time window, a custom software (Matlab; MathWorks, Natick, MA) was used to compute two sets of variables: F_global and F_local variables. F_global variables were expressed with respect to the global frame of reference, i.e., the frame of reference of the motion capture system. Within this frame of reference, we computed the following variables:

• wrist velocity, defined as the module of the velocity of the wrist marker (mm/sec);
• wrist height, defined as the z-component of the wrist marker (mm);
• wrist horizontal trajectory, defined as the x-component of the wrist marker (mm);
• grip aperture, defined as the distance between the marker placed on thumb tip and the one placed on the tip of the index finger (mm).

To provide a better characterization of the hand joint movements, the second set of variables was expressed with respect to a local frame of reference centered on the hand (i.e., F_local; see [11] for a detailed description of the F_local). Within F_local, we computed the following variables:

• x-, y-, and z-thumb, defined as the x-, y- and z-coordinates for the thumb with respect to F_local (mm);
• x-, y-, and z-index, defined as the x-, y- and z-coordinates for the index with respect to F_local (mm);
• x-, y-, and z-finger plane, defined as the x-, y- and z-components of the thumb-index plane, i.e., the three-dimensional components of the vector that is orthogonal to the plane, providing information about the abduction/adduction movement of the thumb and index finger irrespective of the effects of wrist rotation and of finger flexion/extension;
• x-, y-, and z-dorsum plane, defined as the x-, y- and z-components of the radius-phalanx plane, providing information about the abduction, adduction and rotation of the hand dorsum irrespective of the effects of wrist rotation.

All kinematics variables were expressed with respect to normalized (%) rather than absolute (ms) movement durations. After normalizing the duration of each reaching-grasping movement, the data were resampled at intervals of 0.1 of the normalized movement time. This procedure was applied to allow comparison of hand postures across trials and participants.

Clustering parameters

To avoid incorporating the effect of intersession variability for clustering, we selected correct trials i.e. trials with correct marker identification, only from the second experimental session (grasp-to-pour = 147 trials, grasp-to-drink = 154 trials, grasp-to-place = 160 trials). The resulting kinematic parameters for these trials were normalized and submitted to ‘two step clustering procedure’ as implemented in SPSS (version 22) [12] separately for each intention. The two step clustering procedure is a time efficient, agglomerative hierarchical clustering approach that performs clustering in two steps:

1. a pre-cluster step that clusters cases into many sub-clusters;
2. a cluster step that clusters the sub-clusters resulting from pre-cluster step into the desired number of clusters.

The clustering in this procedure is based on the ‘BIRCH’ algorithm [13]. Briefly, the algorithm creates a Cluster Feature Tree (CF) that, instead of storing the actual values, stores the statistics: Number of points (N), Linear Sum (LS), and Square Sum (SS) for all clusters. These features have enough information to calculate the intra-cluster distances. This algorithm has also been shown to be relatively stable both to the input order as well as to selected parameters [13–15]. That is, unlike k-means algorithm or CLARANS, results do not vary based on the order of data points or some initial parameter setting thus, producing stable clusters.

Utilizing this clustering procedure, we first determined, separately for each intention, the optimal number of clusters by calculating Bayesian information criterion (BIC) (BIC difference ratio between successive numbers of clusters less than criteria value of 0.04). This initial estimate of the number of clusters was then refined by finding the largest relative increase in inter-cluster distance between the two closest clusters in each hierarchical clustering stage. Finally, we performed the clustering utilizing the optimal number of clusters obtained by the previous procedure and obtained the cluster number for each trial. The relative importance of each kinematic feature was also calculated. Predictor importance (PI) provides a measure of how well a specific variable differentiates between clusters. The PI of each variable i was calculated as where Ω denotes the set of predictor and evaluation fields, sig_i is the significance or p-value computed based on an F-test.

Participants

Eighteen participants (9 females, mean 24.78 years, range 20–32) were recruited for the experiment. All participants were right handed, had normal or normal to corrected vision, and did not have any neurological disorders. All the participants gave written informed consent to participate in the experiment. None of them participated in Experiment 1.

Stimuli

Video clips of grasp-to-pour and grasp-to-drink movement recorded in Experiment 1 served as stimuli for Experiment 2. After removing videos with technical problems (e.g., hand temporarily out of sight due to trajectory height; water reflection and movement in the bottle; n = 51), we equated the number of movements for each intention based on the intention with least number of trials (grasp-to pour, n = 128). Then, we performed linear discriminant analysis (LDA) using the kinematics variables as predictors and the intention as target and, based on LDA results, we selected, for each intention, the 50 movements that minimized the distance from their own centroid (i.e. the mean variate score for each intention). This procedure allowed us to identify, for each intention, 50 representative movements. As shown in Table 2, cluster distribution profiles (in % of trials) comprised an unequal distribution of movements per cluster similar to that obtained in Experiment 1. This suggests that the selected movements reflected the natural prevalence of movement styles over the entire distribution of movements. The corresponding videos were edited using Adobe Premiere Pro CS6 (.mp4 format, disabled audio, 25 frames/s, resolution 1280 × 800 pixel). To ensure that only advance sources of information were made available to participants as to judge the agent’s intention, video clips were temporally occluded at the time the fingers contacted the object. Each video clip started therefore with the actual reach onset and ended at grasp offset, with the duration of the video varying according to the actual duration of the movement (from 760 ms to 1640 ms). Neither the second part of the movement nor other objects on the table were visible (see sample S1 Video). The actors in the videos were all Caucasian and wore a black colored shirt. The percentage of movements performed by male and female agents was also controlled to be similar across intentions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Cluster distribution profiles for representative movements in Experiment 2 and Experiment 3.

https://doi.org/10.1371/journal.pone.0165297.t002

Procedure

The experiment was carried out in a dimly lit room. Participants sat in front of a 17- inch computer screen (1280 x 800 resolution, 75 Hz), at a viewing distance of 50 cm. Stimuli presentation, timing, and randomization procedures were controlled using E-prime (version 2.0.10.242). Task structure conformed to a binary choice design (Fig 1a). Participants were asked to watch videos of grasp-to-pour and grasp-to-drink movements and to decide if the observed movement was performed with the intent to pour or drink. Each trial began with the presentation of a white central fixation cross (1500 ms), followed by the viewing of a grasping movement. Participant were instructed to respond as accurately and quickly as possible. They could indicate a decision at any time after reach onset by pressing one of two active buttons on a keyboard. When no response was given during the video, a red central fixation cross was displayed until response or 3000 ms had elapsed. After response, participants were also requested to rate confidence of their decision on a 4-level scale by pressing a key (data not analyzed in the current study). Each participant completed a total of 400 trials (200 trials for each intention), with 4 repetitions of each movement. The videos were pseudo-randomized over 4 blocks so that each block included one repetition of each movement. The experimental session lasted approximately 50 minutes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

a) Experimental design for Experiments 2 and 3. Participants were exposed to 4 blocks of 100 trials (50 trials for each intention). Each trial started with the onset of two cues specifying the possible intentions. Then, a fixation cross appeared, followed by the video stimulus. Participants were free to respond either during the video presentation or in the subsequent 3000 ms. b) Representation of a drift diffusion model. The model defines parameters of rate of evidence accumulation, drift rate (‘v’), and separation between the decision boundaries, decision threshold (‘a’).

https://doi.org/10.1371/journal.pone.0165297.g001

To familiarize with the type of stimuli, before the beginning of the experiment, participants were presented with the action execution set up and shown some videos displaying occluded grasp-to-pour (n = 8) and grasp-to-drink movements (n = 8) followed by two unoccluded videos displaying the agent pouring the water into the glass and bringing the bottle to the mouth to drink. No response was requested from participants during this practice.

Dependent measures and Statistical Data Analysis

We modeled the accuracies and reaction times of participants using a drift diffusion model (DDM). Drift diffusion models have been used to model behavioral data from two-choice tasks by modeling decisions based on the accumulation of evidence from the stimuli [16,17]. DDMs specifically model different components of the cognitive decision process (rate of information uptake, bias, etc.) and represent them as different parameters. Modeling of the cognitive decision process underlying the choices with DDMs is thus useful for dissociating task parameters like task difficulty (indicated by drift rate) and effect of instructions (indicated by threshold) [16,18]. The model accesses the performance of the participants in terms of measures like speed of reaching a decision boundary (drift rate, ‘v’) and distance between the two boundaries (decision threshold, ‘a’) (Fig 1b).

In this way, theoretically distinct aspects of the cognitive decision process can be separated statistically. DDMs have thus been proposed to provide a highly detailed measure of participants’ performance, deeper insights into the observed behavior and consequently, drive theoretical advances [19,20]. Indeed, a major advantage of the DDMs is the high degree of information utilization. In contrast to conventional forms of data analysis, the diffusion model incorporates response times (RTs) for correct responses and errors, as well as the ratio of correct and erroneous responses. Additionally, DDMs provide precise, unambiguous quantification of performance (compared to accuracy and reaction times) and are immune to speed-accuracy tradeoffs [18].

We used a hierarchical Bayesian approach for the estimation of these DDM parameters (drift rate ‘v’, threshold ‘a’, and non-decision time ‘t’) as implemented in the toolbox—Hierarchical drift diffusion model (HDDM) [21]. HDDM allowed us to minimize potential confounds resulting from low stimulus probabilities in some movement styles as it can recover reliable estimates even from lesser number of trials per condition [21] To evaluate HDDM model performance, we used the Deviance Information Criterion (DIC). DIC is a commonly used goodness-of-fit measure for evaluating hierarchical models [22]. We compared DIC values (lower being better) for a model allowing drift rates to vary across clusters (‘Cluster model’) with those of a ‘Null model’ where drift rate was not allowed to change between clusters (for a similar approach see [21]). Following the usual rule of thumb, a difference of more than 10 between model DIC scores was interpreted as evidence in favor of the better (i.e., lower) scoring model [23]. Model performance evaluation showed that DIC values for the ‘Cluster model’ (DIC = 21986.42) were indeed lower than those for the ‘Null model’ (DIC = 22834.92) with a difference of 848.50. This indicates that the ‘Cluster model’ provided better fit than the ‘Null model’.

We therefore tested the significance of the estimated parameters for the ‘Cluster model’. Since HDDM toolbox utilizes a Bayesian framework, significance testing can be performed directly on the posterior distribution and results can be interpreted in terms of probabilities. Thus, we analyzed the posterior distributions for different clusters to evaluate differences between them. To this end, we calculated the proportion of the posteriors in which the drift rate for one cluster was higher that than the other. A difference of more than 5% in the posterior distribution overlap (P_p|D) was considered significant (suggesting a higher probability of difference between the conditions). We also evaluated whether cluster drift rates were significantly different from a test value of 0 to ascertain likelihood of drifting towards the correct alternative choice. A drift rate close to 0 corresponds to a process which is equally likely to move towards either of the choices, indicating a slow rate of evidence accumulation. On the contrary, a higher positive drift rate indicates a faster evidence accumulation towards the correct alternative. Since the hierarchical estimation procedure utilized violates the independence assumption, we did not analyze subject parameter estimates in a frequentistic test. Additional confirmatory analyses were also performed on accuracy and RTs separately. Results are provided in S1 File.

Participants

Eighteen participants (11 females, mean 26.17 years, range 22–34) were recruited for the experiment. All participants were right handed, had normal or normal to corrected vision, and did not have any neurological disorders. Written informed consent was obtained from the participants. None of them participated in Experiments 1 or 2.

Stimuli, Procedure and Analysis

Stimulus selection and experimental task procedures were similar as those utilized in Experiment 2, except that grasp-to-place movements (duration ranging from 720 ms to 1280 ms) replaced grasp-to-drink movements (cluster distribution profiles are reported in Table 2). As in Experiment 2, we compared DIC values for a model allowing drift rates to vary across clusters (‘Cluster model’) with those of a ‘Null model’, forcing drift rate to be equal across clusters. With a DIC difference of 1137, the ‘Cluster model’ (DIC = 14651.28) was clearly better than the ‘Null model’ (DIC = 15788.75).