Immersive virtual reality alters selection of head-trunk coordination strategies in young children

Developing coordinated motor control is essential for competent interactions with the surrounding world and requires a balanced multisensory integration. This integration can be challenged under altered sensory feedback, as is the case for vision in immersive virtual reality (VR). While recent works suggest that a virtual sensory environment alters visuomotor integration in healthy adults, little is known about the effects on younger individuals. Here, we assessed the development of head-trunk coordination in children aged 6 to 10 years and young adults using an immersive flight simulator and a virtual joint angle reproduction task. Contrarily to previous results, when vision was decoupled from the steering body part, only older children and adults displayed a joint (‘en-bloc’) head-torso operation mode. Our results reveal that immersive VR affects the coordination strategy in younger children and highlight the immaturity of postural control through the inability to implement a simplified coordination strategy. These findings have implications for pediatric applications of immersive VR, and reveal its usability as an investigation tool for sensorimotor maturation.


Introduction
Coordinated motor behavior and efficient integration of stimuli from different sensory modalities are necessary for successful interactions with the surrounding environment (1).
The development of these abilities follows a long-lasting and elaborate process, starting long before birth and extending into early adulthood. At the motor development level, the skills are usually grouped into two categories. First, gross motor skills comprise postural control and locomotion and require the use of axial and proximal muscles. The maturation of these abilities shows a steep increase until the age of 2 years and continues to refine until later childhood (2-5). Conversely, fine motor skills include precise actions such as functional hand 35 the individual modalities have matured (20,21), unless additional feedback on the reliability 68 of each cue is provided (22). Younger children will thus favor the information provided by the 69 modality with the highest context-dependent reliability (19,23). In the case of postural 70 control, children and adolescents until 15 years standing on an oscillating platform displayed 71 better stabilization with open than with closed eyes, thus indicating a strong reliance on vision 72 (3,24). The display of optic flow patterns to elicit automatic postural movements led to 73 stronger responses in children and adolescents when compared to adults, and the ability to 74 stabilize these movements improved with age until late adolescence (25). This effect was 75 further enhanced when the participants were standing on a sway-referenced platform 76 (26,27). When standing on the unstable platform, which attenuates the proprioceptive 77 feedback, adults use primarily vestibular information to stabilize their posture, and this ability 78 matures only during late adolescence (26). 79 Interestingly, children aged 7-10 years have been shown to display spatiotemporal muscle 80 activation patterns similar to those observed in adults in response to platform oscillations 81 (28), revealing an earlier development of automatic postural responses. Similarly, the 82 predominance of visual cues over self-motion has been observed in children up to 11 years in 83 a navigation task (29,30). The late maturation of visual-vestibular and visual-proprioceptive 84 integration has been correlated with the individual development of these modalities when 85 these are presented in conflict. While adult levels were observed as early as 3 years for 86 proprioception and from 14 years for vision, 15-year-olds still displayed lower levels of 87 vestibular function than adults (31). 88

89
The reliance on visual cues can be further challenged by the use of immersive VR, where the 90 participants are immersed in a digital environment through a head-mounted display (HMD). 91 This paradigm led to stronger sensory recalibration (32) and recruited different adaptation 92 mechanisms (33) than non-immersive sensory alterations. Thanks to the recent development 93 of lightweight HMDs, the use of VR has expanded to numerous applications designed for 94 children, including neurodevelopmental research (30,(34)(35)(36), neurorehabilitation (37-40), or 95 distraction from painful medical procedures (41,42). Yet, the majority of these applications 96 offer none or limited interactions with the virtual environment. Therefore, with the exception 97 of two studies showing that children displayed stronger and longer-lasting responses than 98 teenagers to prism adaptation in immersive VR (43), but generally tolerate this kind of 99 Post-hoc Tukey tests revealed that 6-year-olds performed better in the head-than in the 132 torso-controlled trials in all phases (Before: p = 0.002, d = 1.17; After: p = 0.009, d = 0.83; Day 133 After: p < 0.001, d = 1.17). This difference was also significant for 8-year-olds Before (p < 0.001, 134 d = 1.45), but not during the other phases, although large effect sizes were observed (After: 135 p = 0.83, d = 3.8; Day After: p = 0.652, d = 2.15). Similarly, large effect sizes suggested a 136 superiority of the head over the torso in all phases for 9-and 10-year-olds and After training 137 for adults (Table S1).

147
When steering with their torso, 6-year-olds performed better After than Before training (p = 148 0.013, d = 0.85) and on the Day After than Before (p = 0.014, d = 0.97). The same improvement 149 was observed in 8-year-olds between the evaluations Before and After training (p = 0.001, d 150 = 1.26) and from Day After compared to Before training (p = 0.002, d = 1.28). While not 151 reaching statistical significance, large effect sizes were observed for 9-and 10-year-olds from 152 Before training to Day After (p = 0.998, d = 0.89; p = 0.998, d = 0.74 respectively). Interestingly, 153 large effect sizes suggest that 9-year-olds and adults improved their steering precision in 154 head-controlled trials from Before to After and Day After (Table S1). 155

156
In the torso-controlled trials, 6-year-olds showed significantly lower performance than 10-157 year-olds Before training (p = 0.023, d = 1.34) and on Day After (p = 0.02, d = 1.07) and than 158 adults in all phases (Before: p = 0.006, d = 1.66; After: p = 0.042, d = 1.09; Day After: p = 0.001, 159 d = 1.1.57). Likewise, 8-year-olds performed worse than 10-years-olds and the adults Before 160 training (p = 0.015, d = 1.45 and p = 0.005, d = 1.78 respectively). In the head-controlled trials, 161 6-year-olds displayed higher errors than the adults After training (p = 0.001, d = 1.55), and 162 than 10-year-olds and the adults on Day After (p = 0.013, d = 1.07 and p = 0.002, d = 1.52 163 respectively). Non-significant differences with large effect sizes suggest a gradual 164 development of head-torso motor patterns, particularly between the two older children 165 groups and adults (Table S2).   (Table S3). In particular, torso movements were executed with larger yaw amplitude (p < 184 0.001, p 2 = 0.67) and higher average velocity (p < 0.001, p 2 = 0.79) in the torso-controlled 185 trials ( Figure 2C). Head movements were more similar to trunk movements in torso-than in 186 head-controlled trials, as assessed by the head-torso correlation in the roll plane (p < 0.001, 187 p 2 = 0.89) or the dynamic time warp (DTW) distance between both segments in the yaw 188 plane (p < 0.001, p 2 = 0.82). Interestingly, the higher pitch head anchoring index (AI) in the 189 torso-controlled trials (p < 0.001,  p 2 = 0.85) reveals that the head is preferentially stabilized 190 to the external space than to the trunk in these trials.   To extract the specific variability inherent to torso steering, we repeated the procedure 203 described above, using only the data from the corresponding trials. On this partial dataset, 204 PCA revealed an age-based separation in the space spanned by the first two PCs, accounting 205 respectively for 25.91% and 19.38% of the total variance ( Figure 3A). Individually, both PC1 206 and PC2 showed a decreasing trend with age ( Figure 3A). 207

208
The selection of relevant descriptive variables yielded five functional clusters: Cluster 1 (PC1) 209 and Cluster 2 (PC2) holding variables describing the torso movements, Cluster 2 (PC1) 210 corresponding to head movements, Cluster 1 (PC1) characterizing head-torso correlation and 211 finally Cluster 3 (PC2) containing only the error ( Figure 3B). All the identified variables showed 212 a significant effect of Age and/or Age:Phase interaction (Table S4). Younger children displayed 213 larger vertical head movements (p = 0.004, p 2 =0.42, Figure 3C) and smaller torso 214 movements (p = 0.003, p 2 = 0.44). Remarkably, the similarity between head and torso 215 movements augmented with age, as revealed by the increased correlation in the roll plane (p 216  For head-controlled trials, PCA revealed a soft age-based separation along with the first 229 principal component, accounting for 25% of the total variance ( Figure 4A). Clustering the 230 variables with normalized loadings larger than 0.75 yielded one single cluster describing torso 231 movements ( Figure 4B). All the identified variables showed a significant effect of Age and/or 232 Age:Phase interaction (Table S5). The amplitude of the torso movements decreased with age 233 in the pitch (p = 0.016, p 2 = 0.31, Figure 4C) and yaw planes (p = 0.015, p 2 = 0.32), as well as 234 the average (p = 0.016, p 2 =0.3) and maximal torso velocity (p = 0.015, p 2 = 0.32). 235 236

Study 2 237
To further elucidate the mechanisms underlying the observed behavior, in particular the 238 importance of mature and reliable proprioceptive inputs when the visual feedback is altered, 239 we designed a second study in which the participants were immersed in a virtual landscape 240 as previously and asked to execute a joint angle reproduction (JAR) test using their head or 241 their torso. The JAR paradigm is an active test for proprioception that reflects the functional 242 use of this sensory pathway and relies on kinesthetic memory (47,48), a necessary 243 competence for the proficient use of the flight simulator tested in study 1.      Figure 5A). C Difference between final head orientation and target 324 orientation (see Figure 5F). D Difference between final head orientation and target orientation (see Figure 5G). Dots In this work, we investigated the development of head-torso coordination when challenged 330 by an alteration of the visual feedback through immersive VR. We first evaluated the ability 331 of children aged 6-10 years and young adults to steer an immersive flight simulator using 332 either their head or their torso (Study 1), followed by a virtual JAR task to decipher the 333 behaviors observed during the steering task (Study 2). 334 All the participants were able to steer the simulator using their head in study 1. However,  year-olds showed lower performances than the oldest children and adults, and while this 337 difference was maintained even after practicing the task, the scores were in a comparable 338 range. When using their torso, 6 and 8-year-olds initially struggled to control the simulator 339 but substantially improved their performance with training. Yet, their average error remained 340 higher than the 10-year-olds' and adults'. Overall, 6-year-olds performed worse with the torso 341 than with the head. Kinematic data revealed a stronger involvement of the torso and a stiffer 342 head-torso link during torso-based steering, particularly for the older age groups. Age-related 343 differences in the torso-controlled trials were attributable to an increase of the torso 344 movements, a decrease of the head movements and an increase in the head-torso 345 correlation. Conversely, the age-dependent changes in the head-controlled trials were 346 predominantly caused by a decrease of superfluous torso movements. 347

348
The virtual JAR test carried out in study 2 revealed that in the absence of explicit visual 349 feedback, all participants except the 6-year-olds did not reach the target position with their 350 head while exceeding it when performing the task with their torso. The younger children 351 instead failed to reach the desired orientation with both body parts, overestimating their 352 displacement. During the torso JAR, older children and adults decoupled their heads from 353 their torso, maintaining the head close to the vertical during sideward trials. When explicit 354 feedback was given on the torso position, the 6-year-olds had the tendency to overshoot the 355 target orientation with their head. Lastly, we found that for this age group, the amplitude of 356 unnecessary head movements during the torso JAR correlated with their performance in the 357 torso-controlled flight game. 358

359
The comparable performances observed for all age groups in the head-controlled JAR and 360 steering task indicate that children as young as 6 years are able to use and interact with an 361 immersive body-machine interface both for simple and more complex tasks, in line with a 362 recent study (44). The earlier maturation of the head control is not surprising, as this condition 363 does not require the mastery of an articulated control of the head-trunk unit, which develops 364 from 7 years onwards (10). However, even in this simpler experimental condition, younger 365 children still display a higher error variability and a larger overshoot, confirming the 366 incomplete development of robust internal models as observed in standard experimental 367 frameworks (2,49,50) 368 369 Kinematic analyses of the head-controlled trials showed that the major age-related difference 370 could be attributed to differences in the torso movements, with rotation amplitudes and 371 mean and maximum rotation velocities are decreasing with age. The ability to decouple head 372 from torso movements thus develops along with childhood, confirming previous results 373 obtained during obstacle avoidance during locomotion (1,15), where adults display 374 anticipatory head movements (15). However, mature coordination patterns appear later with 375 our experimental setup when compared to simple locomotion. This is in line with 376 observations revealing that developing children tend to increase their head-body stiffness 377 with increasing task difficulty (9), and to involve their trunk in situations where such 378 movements are not necessarily required (51,52). In our case, the increased difficulty can be 379 imputed to the use of immersive VR, which provides altered visual information and requires 380 higher cognitive processing abilities to appropriately interpret the displayed environment 381 (53,54). Here, immersive VR appears to increase the contribution of proprioceptive and 382 vestibular inputs to postural control over vision (55). 383

384
When the control of the flight game was based on torso movements; instead, younger 385 children struggled to use the system, even after practicing the task. Assessing the kinematics 386 during this task and the JAR reveals an underlying twofold behavior. First, the age-related 387 increase of the torso amplitude in the steering task and the evolution of the torso JAR error 388 indicate that the immaturity of the torso proprioception leads younger children to 389 overestimate their torso movements. This complements a previous study showing an increase 390 in torso positioning accuracy with age (2). Second, the larger head movements displayed by, 391 the younger participants during the flight game and the amplitude of their head movements 392 during the torso JAR with visual feedback suggest that these children attempt to resolve the 393 visual discrepancy by compensatory head movements. This is likely due to weaker reliability 394 of the neck proprioception, which is not mature yet at this developmental stage (31,56,57). The joint display of these two behaviors led to the unexpected observation that only the older 401 participants favorably selected an 'en-bloc' strategy with a stiff intersegmental link during the 402 steering task. This comes in opposition to previous studies, where such behavior was 403 preferentially observed in younger children (1,13,15). One study found a similar behavior in 404 adults, who displayed a head-to-torso stabilization in dimensions in which independent head 405 movements were not beneficial (14). This is concomitant with our results, as head movements 406 in the torso-controlled trials tended to disturb the participants' spatial orientation. Younger 407 children instead failed to use this simpler coordination pattern, which suggests that the 408 This study shows that young children are able to understand and to operate a body-machine 417 interface to interact with immersive VR, but that 6-to 8-year-olds fail to successfully use such 418 a system when decoupling of vision and steering commands is required. In such a sensory 419 environment, these children do not resort to the simpler 'en-block' control strategy usually 420 resorted to at a younger age in challenging conditions, but instead use a less efficient 421 segmental control, overestimating their torso displacement and attempting to correct the 422 visual discrepancy through head movements. This suggests that at these ages, the 423 proprioception at the neck and torso levels is not yet mature enough to be robust to an 424 alteration of the visual feedback, thus preventing an effective visual-vestibular-425 proprioceptive sensory integration, and confirms that the maturation of motor control 426 extends beyond childhood. 427 The results of this study indicate the potential of immersive VR to characterize complex 428 aspects of sensorimotor maturation, but that this technology should be used with care for 429 applications such as motor rehabilitation as it alters the selection of postural strategies in a 430 developing population. 431 432

Methods 433
Subjects 434 Thirty-six typically developing children participated in the first study, grouped as follows: nine 435 6-year-olds (5 girls), eight 8-year-olds (2 girls), four 9-year-olds (1 girl) and eleven 10-year-436 olds (2 girls). Two children (aged 6 and 8) asked to stop the experiment and two other ones 437 (aged 8 and 10) did not comply with the instructions; their data were excluded from further 438 analyses. In addition, 13 healthy adults participated in the study (3 women, age 28.53.4 439 years). Twenty-four typically developing children participated in the second study, grouped 440 as follows: ten 6-year-olds (7 girls), ten 8-year-olds (5 girls), and ten 10-year-olds (5 girls), as 441 well as 10 healthy adults (4 women, age 27.03.2 years). Two 6-year-olds did not complete 442 the session with the flight simulator, their data are reported only for the JAR task. Both studies 443 were approved by the local ethical committees and were carried out in accordance with the 444 Helsinki declaration. All the participants or their legal representative gave their written 445 consent to take part in this study. 446 447

Experimental setup 448
The participants were equipped with a head-mounted display (HMD, Oculus Rift) through 449 which they were shown the virtual environment, and an inertial measurement unit (IMU, X-450 sens MTw Awinda) placed in their back between the scapulae and maintained with a custom 451 harness to acquire their trunk's 3-dimensional (3D) rotation (see Figure 1B). The IMU 452 embedded within the HMD was used both to control the view in the virtual environment and 453 to acquire the head rotations. The kinematic data were acquired at a sample period of 68 ms. 454 455

Virtual environment and navigation task 456
We created a virtual environment (VE) using the game engine Unity3D, which represented a 457 FPV flight on a bird's back at a constant speed of 12 m/s, (45,46). A succession of coins to 458 catch (distance between consecutive coins: 58m) represented a path to follow, randomly 459 alternating simple forward motion and one of four directional maneuvers (right turn, left turn, 460 ascent, descent). The coins' initial diameter was 1 m, and every time one coin was caught, 461 the next one was enlarged to 2 m. To minimize possible effects of path planning abilities, we 462 additionally displayed a colored line smoothly connecting the coins, computed as a . Similarly, to provide the participants with a visual cue of their own position 464 in space, an eagle was displayed below their visual horizon (see Figure 1A). Finally, to keep 465 the experiment engaging, a tinkling sound was played when the coin was caught at a distance 466 smaller than 10 m, which also added points to a total score for the trial, displayed at the top 467 of the screen. Upon arriving, the participants were shown the movements to control the simulator using the 491 head or the torso. They were equipped with the HMD and the IMU, and were seated on a 492 stool or on a chair and asked not to lean against the backrest. The participants were randomly 493 allocated to start the experiment using the head or the torso, using adaptive covariate 494 randomization with the gender as covariate (61). For the torso-controlled trials, the 495 participants were advised to keep their neck rigid as to move their entire upper body as a 496 whole. Similarly, before starting the head-controlled trials, the experimenter made the 497 participants aware that moving their trunk was unnecessary. 498 The recording sessions took place on two consecutive days. On day 1, the participants had to 499 steer the simulator along four paths with each body part. The first sequence contained 26 500 coins and was an initial evaluation of the performance (hereafter: Before). The second and 501 third sequences each contained 50 coins; these sequences were considered as training. The 502 fourth sequence contained 18 coins (hereafter: After). All the sequences controlled with a 503 given body part were executed successively. On day 2, one sequence containing 26 coins had 504 to be performed with each body part (hereafter: Day After). Breaks were allowed between 505 the sequences, at the participants' demand. 506 507 Experimental protocol study 2 508 The participants were equipped and seated as previously and were shown the JAR 509 movements by the experimenter. The conditions were tested in the following order: 510 Feedback, Still, Forward, while the participants were randomly allocated to start either with 511 the head or the torso, using covariate adaptive randomization with the gender as covariate 512 (61). The orientations were presented in a randomized order, totalling 5 repetitions for each 513 orientation in the Feedback condition and 10 repetitions for the Still and Forward conditions. 514 At the end of the session, the participants executed one flight sequence with the simulator 515 (Before session described above). 516 517

Data processing 518
The kinematic data acquired in study 1 was divided into segments corresponding to the 519 intervals between consecutive coins. Descriptive variables were computed on these segments 520 and averaged over each entire sequence (see Table 1). Principal component analysis (PCA) 521 was applied to the dataset containing the kinematic variables extracted from all trials, or from 522 the head-and torso-controlled trials, respectively. Outliers were detected as data points 523 whose Euclidean distance to the centroid of the z-scored dataset deviated from the average 524 value by more than 4 standard deviations. These points were given a weight of 0.5 in the PCA 525 computation. The variables with normalized loadings > 0.75 on the first (all trials, head-526 controlled trials) or the first two principal components (torso trials) were considered as 527 significant and were regrouped into functional clusters. 528

529
The data acquired during study 2 was separated into individual trials, and the final position 530 was averaged over the last 1.5 s of each trial. For each trial, we computed the signed error 531 with respect to the target orientation, the overshoot, the number of oscillations around the 532 final angle, and for the trials involving the torso, the head anchoring index (AI, computed over 533 the entire trial), the final angular difference of the head and the torso and the head alignment 534 "error" as the difference between the final head angle and the target orientation. 535 536

Statistical analysis 537
The statistical evaluations were performed using paired t-tests or repeated-measures 538 ANOVA, using the age as a between-subjects factor and the control type and/or experimental 539 phase as within-subject factors using custom Matlab routines (62). The p-values were 540 corrected using the Greenhouse-Geisser correction when Mauchly's test indicated a violation 541 of sphericity. Post hoc analyses were conducted using Tukey's honest significant differences 542 test, with a significance level of .05 for all tests. 543

544
Acknowledgments 545 We thank Stefania Saviotti, Elisa Freddi and Davide Esposito for their help in the data 546 collection. This work was supported by the Bertarelli Foundation. 547 548 549 References 550