The Autonomic Nervous System Differentiates Between Levels of Motor Intent and Hand Dominance

While attempting to bridge motor control and cognitive science, the nascent field of embodied cognition has primarily addressed intended, goal-oriented actions. Less explored however, have been unintended motions. Such movements tend to occur largely beneath awareness, while contributing to the spontaneous control of redundant degrees of freedom across the body in motion. We posit that the consequences of such unintended actions implicitly contribute to our autonomous sense of action ownership and agency. We question whether biorhythmic activities from these motions are separable from those which intentionally occur. Here we find that fluctuations in the biorhythmic activities of the nervous systems can unambiguously differentiate across levels of intent. More important yet, this differentiation is remarkable when we examine the fluctuations in biorhythmic activity from the autonomic nervous systems. We find that when the action is intended, the heart signal leads the body kinematics signals; but when the action segment spontaneously occurs without instructions, the heart signal lags the bodily kinematics signals. We posit that such differentiation within the nervous system, may be necessary to acquire the sense of action ownership, which in turn, contributes to the sense of agency. We discuss our results while considering their potential translational value.


32
The field of embodied cognition (EC) has provided a powerful theoretical framework amenable 33 to bridge the gap between research probing our mental states and research investigating our physical 34 actions [1][2][3]. Indeed, within the framework of EC, the construct of agency conceived as a cognitive 35 movement phenomenon [4][5][6], may provide a way to finally connect the disparate fields of cognitive 36 science and motor control. An important component of agency is action ownership [5,7,8], i.e. the 37 sense that sensory consequences of the actor's action are intrinsically part of the actor's inner 38 sensations. When the actor owns the action, s/he has full control over those sensations that are 39 internally self-generated and self-monitored by the actor's brain, and yet extrinsically modulated by 40 external sensory goals. A critical aspect of this internal-external loop is the identification of the level 41 of actor's intent, and its differential contribution to the action's intended and unintended sensory 75 motor variability in fundamentally different ways (if we compare the signatures of variability derived 76 from the spontaneous samples to those derived from deliberately staging the same movement 77 trajectories [12,25].) More importantly, the fluctuations in the motor variability of these spontaneous 78 motions can forecast symptoms of Parkinson's disease before the onset of high severity [29,30]. They 79 have also aided in evoking the sense of action ownership and agency in young pre-verbal children 80 [31]. For these reasons, here we posit that deliberate and spontaneous segments of complex actions 81 ought to differentially contribute to our sense of action ownership and to our overall sense of agency.

82
To examine this proposition, we follow a phylogenetically orderly taxonomy of the nervous systems' 83 maturation ( Figure 1B) and examine all levels of neuromotor control -from autonomic to deliberate 84 -necessary to coordinate voluntary motions ( Figure 1A).

85
More specifically, since autonomic systems are vital to our survival and wellbeing, they may 86 remain impervious to subtle distinctions between deliberate and spontaneous motions that take place 87 across the body, as the end effector completes goal-directed actions. Here we explore the interplay

113
During the experiment, movement kinematics and heart signals were recorded from each 114 participant. However, one participant's recording had too much noise (i.e., inaccurate sensor position 115 with error larger than 10cm), so we excluded this participant's data in the analysis. For that reason, 116 eight participants' motor and heart signals were analyzed.

122
These sensors were secured with sports bands to allow unrestricted movement during the recordings.

127
Electrocardiogram (heart data): Three sensors of electrocardiogram (ECG) from a wireless Nexus-10 128 device (Mind Media BV, The Netherlands) and Nexus 10 software Biotrace (Version 2015B) were used 129 to record heart activity. At a sampling rate of 256Hz, the sensors were placed across the chest 130 according to a standardized lead II method.

136
As shown in Figure 2, for each trial, the participant was presented with a circle on the tablet 137 screen. This circle served as a prompt for the participant to touch the tablet screen within five seconds.

138
After the touch, either 100ms, 400ms, or 700ms elapsed, and the participant heard a tone at 1000Hz 139 for 100ms. Then, on the tablet screen, the participant was presented with a sliding scale, ranging from 140 0 to 1 (second), to indicate how long he/she perceived the time elapsed between the touch and the 141 tone. The response was to be made within five seconds upon the display of the sliding scale. The five 142 seconds time-window was considered enough for the participant to provide a response, as it took 143 approximately 1 s to touch the screen and retract the hand back to its original position. There was a 144 total of three conditions -control, low cognitive load, high cognitive load -and each condition 145 consisted of 60 trials. In the control condition, the participant simply performed each trial with no 146 additional task; under the low cognitive load condition, the participant performed each trial while 147 repeatedly counting forward 1 through 5; under the high cognitive load condition, they counted 148 backwards from 400 subtracting by 3 while they performed each trial. Participants counted forward 149 and backward at their own comfortable pace, and they took breaks in between each condition. The

153
In this study, we extracted the kinematics (i.e., linear speed, angular acceleration) and heart data 154 during time segments when the participant made a pointing motion towards the circle presented on 155 the tablet screen; and combined them across the three conditions. As a result, we analyzed the 156 kinematics and heart data recorded while the participant made 180 pointing motions (less any trials 157 that were deemed noisy; the most trials we excluded per participant due to instrumentation noise 158 were 12 trials).

159
To analyze the ECG and kinematics data in tandem, we up-sampled the kinematics data from 160 240Hz to 256Hz using piecewise cubic spline interpolation. Note, the ECG signals were not 161 synchronized with the kinematics data but were manually time stamped at the start and end of each 162 experimental condition. For that matter, we expect a presence of lag between the two modes of signals 163 -kinematics and ECG -but the lag would not exceed 1 second.

164
To exclude effects of muscle motion from the ECG heart data, we bandpass filtered the data with 165 Butterworth IIR for 5-30Hz at 2nd order. This filter was effective in identifying QRS complexes and 166 extracting R-peaks in previous studies [13,32]. Here, the filter excluded the dominant frequency 167 range where typical kinematics signals are present (see Appendix Figure A1). We performed our 168 analyses using both filtered and non-filtered EKG data and found similar trends and patterns.

169
However, the paper only presents the results from using the filtered data, as it is a better reflection of 170 the heart activity.

172
We used the rationale in Figure 1 to structure our analyses, with a focus of two main axes 173 denoting the level of motor intent and awareness that the brain may have during complex tasks 174 ( Figure 3A). More precisely, one axis explores possible differentiations between time segments of the 175 pointing movements that are deliberately aimed at an external target (forward/high motor intent) vs.

176
segments that are consequential to the deliberate ones (backward/low motor intent). The latter may 177 occur when the hand retracts back to rest, or when after touching the target the person transitions the 178 hand in route to another goal-directed motion. These segments have been studied in our lab across 179 very complex motions in sports (boxing, tennis) and in the performing arts (ballet, salsa dancing). We 180 have coined them spontaneous movements and discovered that they have precise signatures that 181 distinguish them from the deliberate ones. For this reason, we hypothesized here that these

188
After the touch, the participant heard a tone. The duration between the touch and the tone was 189 randomly set to be 100ms, 400ms, or 700ms. In the next 5 seconds, the participant was presented with  side of the body. Furthermore, we explored how other body parts (also co-registered within the 204 sensors' network) contributed to the overall performance of this task.

205
These two axes were explored at the voluntary level of motor control interleaving deliberate 206 goal-directed (forward) actions and spontaneous (backward) segments of the full pointing loop. We 207 also included in our analyses the autonomic level of control in the taxonomy of Figure 1A. And to 208 that end, we co-registered the heart activity and incorporated it into the bodily kinematics activity 209 ( Figure 3B). We next explain how to overcome challenges in sensors' data fusion from disparate 210 systems along with new approaches to analyze these multi-modal data.   Allometric effects: Another issue is that when examining such data from different participants with 240 different anatomical sizes, allometric effects may confound our results. This is so because e.g. the 241 speed ranges that a person attains depend on the length of the arm. Longer arms tend to broaden the 242 ranges of speed and contribute to the distribution of speed values that the person attains in any given 243 experiment. As such, we need to account for these possible allometric effects.

244
Assumption of normality: Another related matter to the ranges of speed and their distributions is 245 that they vary from person to person according to multiple factors (e.g. age, body mass, sex, fitness,

256
Assessing similarity in probability space: Going beyond significant hypothesis testing models, one 257 may need to assess the differences between probability distributions. To that end, one may need a 258 proper similarity metric. Yet, when our data represents points in probability space, and the 259 distributions are not symmetric, it is challenging to assess their similarity in a consistently proper 260 way. Measures like the Fisher information metric are designed to compare symmetric distributions 261 and the Kullback-Leibler divergence is computed asymmetrically between distributions (one-sided).

262
We would like to have a proper (two-sided) distance metric to assess change and its rate when points 263 are related to non-symmetric continuous probability density functions, or to their discrete 264 approximations.

265
Degrees of freedom across intent levels of motor control: Multiple locations of the grid of sensors, co-266 registering biorhythms from different nervous systems, contribute differently to the overall behavior 267 of the system. Some may be more directly related to action success, while others may provide 268 support. Separating the bodily region within a kinematics-heart network can be challenging because

294
Either way, these fluctuations ought not be averaged out by assumptions of normality. Whereas in 295 the extant literature these fluctuations are considered noise, or superfluous, here we treat them as 296 important signal.

297
Distribution-free approach to counter current assumption of normality: We do not assume normality 298 in the data. Instead, we gather enough data to empirically estimate the best family of probability 299 distributions that fits the data. To that end, we here use maximum likelihood estimation (MLE) with 300 95% confidence intervals and seek the best continuous family that fits our data.

301
Distance metric to assess similarity in probability space: We here introduce the use of the Earth 302 Mover's Distance Metric (EMD) [36-39] to approximate (using the frequency histograms of the MMS 303 amplitudes) the stochastic shifts in probability space that occur for different movement types. This is 304 an appropriate similarity metric that allows us to examine the extent to which different levels of 305 motor control change the stochastic patterns. We briefly describe it below:

306
The EMD, also known as the Kantarovich-Wasserstein distance [40], measures the distance 307 between two discrete probability distributions. Given two discrete distributions P = {(p1,wp1), … 308 (pm,wpm)} [13,14], where pi is the cluster representative and wpi is the weight of the cluster; and Q = 309 {(p1,wp1), … (pn,wpn)}, EMD computes how much mass is needed to transform one distribution into As there are infinite ways to do this, the following constraints are imposed to yield EMD values: Network connectivity analyses to assess degrees of freedom recruitment across modalities of motor 321 control: We use graph theory to examine the inter-relations across the nodes of the multilayered 322 kinematics-heart network. To that end, we derive an adjacency metric of pairwise quantities 323 reflecting the cross-correlation between any pair of nodes in the grid. We then construct weighted 324 directed networks and borrow connectivity metrics from brain-related research. We extend these 325 methods to represent the peripheral network using the bodily biorhythms from multiple layers of the 326 nervous systems' functioning, spanning from voluntary to autonomic ( Figure 1A).

333
To assess spatial components, we use the scalar speed (

343
As we need many spikes for our distribution-fitting and stochastic analyses, we used the angular

352
We also examined temporal components of the data. To that end, we used the linear speed As a first step, we separated the kinematics data obtained from all 10 body parts, using the start 360 and end time of the dominant hand making a forward-deliberate motion, and the hand making a 361 backward-spontaneous motion ( Figure 4A). This is possible to do (automatically) because (1) the 362 speed is near 0 at the onset of the motion towards the target; (2) the distance to the target 363 monotonically decreases and once again the hand pauses at the target at near 0 speed. As the 364 deliberate (forward) segment is completed, the speed rises again away from 0 and the distance to the 365 target increases as the hand follows the backward segment of the full pointing loop. The two 366 segments can be automatically differentiated also because the deliberate (forward) one is less variable 367 than the spontaneous (backward) one [11,29,34,42].

368
For the connectivity analysis centered on spatial aspects of the signal amplitude, we pooled the which were visualized as edges in the schematics of the network in Figure 4K. The intuition behind 379 taking the absolute difference in angular acceleration time series from two body parts is that this 380 reflects the change in positional distance between those two body parts, and thus represents the 381 connectivity (physical distance) between those two.

417
For the spatial domain of connectivity, we took the segmented data of angular acceleration and 418 EKG data, and extracted MMS from both signals, and plotted a histogram of the MMS. Because the 419 MMS of EKG signals did not follow a Gamma distribution, in order to assess the connectivity between 420 the two, we computed the earth mover's distance (EMD) between the histogram from a single body 421 part and from the EKG data ( Figure 5A-D).   For all these metrics, we compared the medians between different movement segments (i.e., 457 forward vs. backward) and different hand dominance (i.e., right vs. left arm/hand), to understand how 458 stochasticity and temporal dynamics changed across varying levels of motor intent between the heart 459 (from ANS) and kinematics (from PNS/CNS).

481
Motor intent in the context of our experimental assay specifically refers to the level of 482 deliberateness (or spontaneity) of the movement segment in route to an external target (away from 483 it). An instructed pointing action to touch the target is a goal-directed reach with high level of intent.

484
In contrast, the uninstructed spontaneous retraction away from the target carries lower motor intent 485 than the goal-directed one.

486
As a first set of analysis, the MMS extracted from the angular acceleration data from each body

537
When we compared between different motion segments, median cross-correlations were higher 538 for forward motions than for backward ones for all but two participants. When we compared between 539 different dominance side, all participants showed higher correlation on the dominant side than the 540 non-dominant one. The median CC showed to be higher for forward motions than for backward 541 segments for all participants, and higher for the dominant side than the non-dominant side for all but 542 two participants. For all participants, both measures showed statistical significance in their difference 543 (see Table A1 of Appendix for detailed statistical results).

555
The distinctions that we observe from these findings, on how different levels of motor intent 556 have separable network connectivity patterns based on temporal aspects of the kinematics data, are 557 consistent with the patterns uncovered using spatial aspects of the kinematics data. Specifically, when 558 we exert higher intent on our body, regardless of the physical trajectory of the motion, there is a 559 stronger connectivity across our body parts. However, we note that this pattern is not as uniform 560 across all participants, as we had found in the spatial aspect of the network analysis.

564
To assess patterns of connectivity between biophysical signals derived from voluntary and 565 autonomic levels of motor control we examined the kinematics (generated by the CNS-PNS) and the 566 heart activity (generated by the ANS). The patterns of MMS stochasticity and temporal correlation 567 across these systems distinguished levels of motor intent and control.

568
The analyses involving EKG and kinematics revealed larger stochastic differences in MMS data 569 when higher motor intent and control are exerted. More precisely, the pairwise EMD showed higher 570 differentiation between these two signals in all but one participant when forward motion was made, 571 but only on the dominant side of the body. Furthermore, all but two participants showed higher EMD 572 on the dominant side of the body, but only during forward motions. On the other hand, however, 573 when backward motion is made, we find an opposite pattern, where all participants show higher 574 EMD on the non-dominant side. We infer that there may be a modulating factor that underlies the 575 stochastic relation between kinematics and heart signals.

576
When we examine the temporal relations between the two signals, by computing pair-wise 577 cross-correlations, we see higher cross-correlations when there is lower motor intent across all 578 participants -that is, during backward motions, and on the non-dominant side. Here we note the low 579 range of the correlation coefficient values, around 0.1. However, we see a similar trend when this is 580 based on the non-filtered raw EKG data, with a higher range around 0.6.

583
We also examined the lag values to assess which signal leads the other. We found that motions 584 under higher motor intent (i.e., during forward-deliberate motions performed with the dominant side 585 of the arm), EKG signals tend to lead the kinematics signal. On the other hand, in movements 586 performed under lower intent (i.e., during backward-spontaneous motion, and on the non-dominant 587 side of the arm), kinematics signals tend to lead the EKG signals. This is depicted in Figure 8.

604
We caveat that that because the EKG device and motion capture system was not exactly 605 synchronized, the absolute lag value may not be as meaningful. Nevertheless, as we analyze these 606 data in terms of the difference (i.e., the delta lag values between forward and backward motions, and 607 between dominant and non-dominant sides), it is indeed meaningful to find such patterns uniformly 608 across all participants.

609
Table1 summarizes the results that we showed in the sections above. We emphasize that

635
Our initial thought was that autonomic systems contributing to our brain's autonomy over the 636 body and to our overall embodied sense of agency would remain impervious to stochastic shifts at 637 the voluntary levels. We reasoned that given the vital role of these systems for survival, their robust 638 signal would not reflect subtle changes in levels of intent, motor awareness and voluntary control.

639
As such, our guess was that if during voluntary movements, there were stochastic differences 640 between deliberate and spontaneous segments of the reach, or between dominant and non-dominant 641 sides of the body, such shifts in patterns of variability would not be appreciable in the heart signals' 642 fluctuations. Our guess was altogether wrong. Not only were the heart signals' differences 643 quantifiable at the level of fluctuations in signal amplitude; these differences were appreciable as well 644 in the inter-dynamics of the kinematics and cardiac signals.

645
We found that when movements are intended and deliberately performed to attain the goal 646 defined by an external (visual) target, the heart signal leads the movement kinematics signal. Yet, 647 when these overt movements are spontaneous in nature, i.e. uninstructed and not pursuing the 648 completion of a specific externally defined task goal, the heart signal lags the movement kinematics 649 signal. Across spatial and temporal parameters, we found consistent trends and confirmed the trends 650 through different parameters. Indeed, deliberate motions, performed with the dominant effector, 651 carry higher levels of NSR, denoting higher fluctuations away from the empirically estimated mean.

652
We interpret these findings considering the principle of reafference [46]. Furthermore, we 653 discuss the possible contributions of these self-generated signals to the self-emergence of cognitive 654 agency from motor agency, namely, the sense that one can physically realize what one mentally 655 intends to do, confirm the consequences (both intended and unintended) and as such own the action.

662
particularly within the framework of internal models for action [21,47,48] and more recent models 663 of stochastic feedback control [49,50]. Central to all these conceptualizations of the control problem 664 has been the notion of anticipating the sensory consequences of impending intended actions.

665
Nevertheless, nothing has been said about the consequences of action segments that bear a lower 666 level of intent, that occur spontaneously, or that are altogether occurring autonomously. Modelers 667 and experimenters in motor control do not seem to be aware of the former (although see [11, 12, 51]) 668 and the latter are assumed to be far removed from cognitive processes (

686
The present work provides empirical evidence that (1)

729
Here we offer a unifying framework with a taxonomy of function and differentiable levels of

798
distribution and has a large sample size (n>1000) that may yield low statistical power.