Movements in infants have age-specific characteristics. Fetal/preterm general movements (GMs) show large variations (Prechtl, 1990) until 36–38 weeks gestational age (GA). Writhing movements are then observed between 36 weeks GA and 2 months post term. These movements are proximal, characterized by a small to moderate amplitude, and of slow to moderate speed (Prechtl et al., 1997). Fidgety movements appear between 2 and 5 months of age: they are more distal, of smaller amplitude and lower speed, and more variable in acceleration. In addition to GMs, infants exhibit various other spontaneous movements that increase in number and variety with age (Prechtl et al., 1997). At 3 months of age, voluntary, goal-directed, functional movements progressively replace spontaneous movements. These developmental changes in motor upper limb behaviors indicate changes in developmental stage: the infants' muscle strength increases to overcome the force of gravity and the use of environmental information induces an increase in varied, goal-oriented movements. These changes are associated with a shift from subcortical to cortical processing (Hitzert et al., 2015). Infants start to reach out at 3–5 months of age to explore objects and the environment with their hands (Thelen et al., 1996). By 9 months, infants begin to develop clear communicative gestures, such as pointing, which paves the way for language development (Iverson and Goldin-Meadow, 2005; Iverson, 2010).

Recent technical advances have provided opportunities to automatically detect, assess, and track clinical developmental features, such as HM in infants. Kinematics is referred to as the “geometry of motion” by describing the position, velocity, and acceleration/deceleration of points of the described object and movement units, of which the combination measures the trajectory, curvature or jerkiness, smoothness, straightness, complexity, and repertoire of movements (von Hofsten, 1991; Thelen et al., 1996; Berthier and Keen, 2006; Meinecke et al., 2006). Kinematics of GMs are characterized by smoothness, assessed by constant changes in the angular values (Karch et al., 2008) and high complexity, without regular periodicity (Meinecke et al., 2006). Little longitudinal data concerning kinematics of spontaneous HM have been collected on infants. Gima et al. (2011) showed that infants' spontaneous movement involves chaotic dynamic systems that are capable of generating voluntary skill movements around 3–4 months. Waldmeier et al. (2013) showed a decreasing trend of statistical persistence in the acceleration time series with maturation, from birth to 2–3 months of age. Zoia et al. (2013) have studied the kinematics of upper limb movements from fetal life to 12 months of age. They particularly found two results: the longer movement duration and deceleration time at 22 weeks of gestation for movements toward the eye rather than the mouth, re-emerge at 4 months of postnatal life. They also showed at 4 months a reorganization of the length of the deceleration phase of the movement according to environment differences. Their results possibly suggest reorganization phases according to the novel environment.

Kinematics have largely been used to study infant reaching and prehension (Berthier and Keen, 2006; Lynch et al., 2008; Gonçalves et al., 2013). A period of pre-reaching, specific to toy oriented gestures, has been described (Bhat and Galloway, 2006). The smoothness of the reaching movement improves from 2 to 7 months of age (Lee et al., 2011), but independently of straightness (Thelen et al., 1996). Konczak et al. (1995) described this second period of reaching as the “fine-tuning” of the sensorimotor system that begins at 6–7 months, with smoother movements, fewer movement units, and an increase in straightness. Although reaching parameters become more stable by 8 months (Thelen et al., 1996), the time period for reaching kinematics remains controversial. Infants reliably grasp for objects within their workspace 3 ± 4 months after the onset of reaching, stable patterns of temporal coordination across arm segments begin to emerge at 12–15 months of age and continue to develop up to the 3rd year (Konczak and Dichgans, 1997).

Marcroft et al. (2014) reviewed the advantages and limitations of various methods applied to assessing spontaneous GM in high-risk infants. Indirect (video cameras or 3D motion capture) or direct sensing (through sensors worn by the patient) both capture limb movements, but differ in their temporal and spatial resolution. The main advantage of video-based movement analysis is the recording of spatial data in addition to temporal measurements, in a natural setting. Sensor approaches have much higher temporal resolution, but are less feasible in everyday clinical practice and carry on artifact-related issues.

Even though video based analysis allows assessment of movements under natural conditions, infants are generally assessed lying or sitting, without any specific task. Very few studies have described the kinematics of HM during interactions with the caregiver, although these are the most relevant situations to assess genuine interactive processes (Avril et al., 2014; Leclère et al., 2016).

The literature on the kinematics of HM mostly concerns preterm infants. The variation of these patterns (intrinsic movement features and their repertoire) has been used to predict later neurological outcome. Atypical motor development in preterm and full-term infants (limited variation, poor repertoire, fluency, and complexity), especially the presence of cramped synchronized GMs which are more stereotypical, predict later developmental impairment and cerebral palsy (Hadders-Algra, 2004; Einspieler et al., 2012). The poor repertoire of GM has shown its utility in predicting neurodevelopmental outcome at two (Beccaria et al., 2012) and two-and-a-half years of age (Kodric et al., 2010). In particular, poor quality of fidgety GM at 3–5 months has been reported to predict minor neurological dysfunction later, at school age (Hadders-Algra and Groothuis, 1999). However, earlier GMs, during the writhing period, are considered to be less reliable in predicting cerebral palsy (Prechtl et al., 1997).

The clinical assessment of hand movements in at-risk infants requires training and expert clinicians. Nonetheless, kinematics is an important tool to detect specific cues of HM, which can then be more precisely clinically assessed. Ohgi et al. (2008) found that HM in infants with brain injuries were more unstable and less predictable than the movements of low-risk infants in a longitudinal study of spontaneous movements of preterm infants. Adde et al. (2009) classified non-fidgety vs. fidgety movements in preterm infants using computer-based video analysis. Kinematics have shown superiority over clinical assessment to detect specific features of HM: clinical assessment correctly identified all infants with neurodevelopmental impairment, including cerebral palsy, by 3 months, but only the stereotypy score of limb movement, derived from kinematics, discriminated the appearance of cerebral palsy from other developmental impairments (Philippi et al., 2014).

Infants at risk of autism spectrum disorders (ASD) are particularly studied during their first year of life. The detection of early signs is the first step to intervene and improve the prognosis of neurodevelopmental disorders (Rogers et al., 2014). Although later communicative gestures in ASD children are well-known, relatively little attention has been given to specific motor and movement development as a potential diagnostic risk marker for ASD during the first year of life, until recently. Such studies are mostly retrospective (e.g., home movies) and lack systematic motor assessment (Zwaigenbaum et al., 2013). Only one assessed early movements by video analysis, but none used kinematics. Phagava et al. (2008) found a preponderance of writhing movements in ASD children (70% in ASD vs. 12.5% in controls), and the absence of or abnormal fidgety movements were found in 50% of ASD vs. 11% of control children. However, no statistical significance was reported due to the small sample size. Two other studies coded home videos using a standardized movement analysis system. Esposito et al. (2009) assessed 0 to 5-month-old infants in a supine position with “manual” frame-by-frame video rating of movements. Infants with ASD exhibited lower levels of both “static” and “dynamic symmetry” than both infants with intellectual disability and controls. Teitelbaum et al. (1998) assessed GMs in 4 to 6-month-old infants on home videos using the Eshkol–Wachman Movement Analysis System and found disturbances in the shape of the mouth and development milestones (lying, righting, sitting, crawling, and walking) in infants who later showed autistic features, but did not focus on HM. No specificities have yet been found for motor function nor HM.

This longitudinal and observational study aimed to assess HM in infants at risk for developmental disorders and controls, followed from the age of 2 to 10 months, during mother-infant interactions. We hypothesized that HM may reflect early specificities of gesture and communicative development, which are clinically difficult to assess. We selected five groups of at-risk infant populations: two populations for neurodevelopmental risk, including ASD (epileptic infants with West syndrome; premature infants); two for specific environmental factors [infants being hospitalized immediately after birth and infants with visually impaired mothers (VIMs)]; one for both risks (infants with orality disorders); and one control population. We examined the longitudinal development of HM under natural interactive conditions with the mother, during three communicative situations, to differentiate toy-directed and person-directed movements.

We explored the kinematics of HM using several kinematic descriptors to characterize the spatial use, curvature, acceleration, and velocity of HM asking the following research questions:

Do HM differ with age in all cohorts?

At risk participants were recruited from the Necker Enfants-Malades Hospital, Paris, and healthy or TD infants from maternal and infant protection institutions, pediatric consultations, or by proxy, from February 2004 to March 2013. Six cohorts of infants with and without risk were selected (Table 1):

West Syndrome (WS): The WS cohort was particularly important because it leads to ASD in 7–25% of cases and to developmental delay or learning disorders in 70–90% of cases. Prior studies have already shown that early assessment of interactions is an early predictor of later developmental delay or ASD (Ouss et al., 2014).

Developmental Quotient (DQ) was measured by the Brunet-Lezine (BL) scale at 12 months, which estimates the developmental quotient in four domains: posture, coordination, language, and socialization, based on standardized developmental quotients available for 0- to 30-month-old French toddlers (Josse, 1997).

Exclusion criteria were as follows: parents' refusal to consent to follow-up assessment and/or to the research protocol, families who lived too far from the hospital, or families who could not complete the questionnaires due to an insufficient knowledge of French. This study was approved by the Institutional Review Board (Comité de Protection des Personnes from the Groupe-Hospitalier Necker Enfants Malades, IRB registration number 00001072), and parents gave written informed consent after receiving verbal and written information on the study. CNIL procedures (Commission Nationale Informatique et Libertes) for data anonimity processing were respected.

The infants' HM were assessed during standardized play sessions in interaction with the mother. They were recorded at various timepoints from the age of 2 to 10 months. Two synchronized cameras (face and profile, see Figure 1, top panel) recorded the movements in two dimensions while the infant was sitting in a baby-chair. The standardized situation was divided into three sequences: 3 min of free play in which the mother was instructed to interact “as usual” without any toy (sequence 1), 3 min of play with a cuddly giraffe (sequence 2), and 3 min of play with the mother singing a well-known nursery rhyme accompanied by rotating her own hands (sequence 3).

One minute was extracted from each 3-min video sequence (thus 3 min per infant per session) according to two criteria: the infant's hands should be visible for at least part of the sequence (e.g., the mother is not leaning toward the infant) and the minute was that with the greatest amount of interaction between the mother and the infant. This duration does not reflect the entire complexity of interactive behaviors, but in the field of social and clinical psychology, Ambady and Rosenthal (1992) have shown that predictions based on recorded observation of even under 30 s did not differ significantly from predictions based on 4–5 min observation.

The coordinates of the hands were computed from the video images using a tracking framework implemented in a software that was designed and developed for this purpose (software and documentation available in Supplementary Material). The tracking framework comprised three steps: prediction, observation, and estimation. The prediction step estimated potential locations of the hand using a dynamic model. The observation step estimated the probability associated with the predicted location, using a computed likelihood measure, based on the data provided by the camera. Finally, an estimation of the state of the hand (here limited to its positions and velocities in the picture) was performed. A yellow wristband was used for the right hand for the observation step. The observation was based on the Bhattacharyya distance, which measures the likelihood between color histograms computed on a region-of-interest around the predicted location and on a hand-selected one in the first frame of the sequence (Czyz et al., 2005). For the other two steps, we developed an approach using a bootstrap-based particle filter with a first-order model as the HM were highly non-linear with abrupt changes in direction and speed (Isard and Blake, 1998; Hue, 2003).

The infants' hands were often occluded in the video, as the parent could interact freely with the infant. We used an approach combining tracking with detection (Czyz et al., 2005) to minimize this problem. This was performed by adding a Boolean variable to the state vector associated with each particle: its value was set to zero if the particle was in detection mode or to one if the particle was in the tracking mode. The transition from one mode to the other was managed via a transition matrix. In the tracking mode, the state of the particle was determined according to a classical particle filter-based algorithm. In the detection mode, the state of the particle was randomly determined using a computed probability map, based on the entire image.

Each extracted trajectory consisted of 1,500 pairs of x and y coordinates (25 frames per second, generating 1,500 pairs of coordinates during the 60 s, see Figure 1, second panel). The frames in which the hand was not visible were clearly indicated in each trajectory as missing coordinates for these timepoints. Incomplete trajectories missing more than 30% of the coordinates, either due to occlusion of the camera by the mother or losing the target wristband by the tracking framework, were removed from further analysis. The trajectories obtained were normalized using a fixed reference system present in the settings of each video recording (Figure 1) to account for differences in the camera zoom parameters. Normalization was performed on all trajectories and 95% of the normalization factors ranged between 0.8 and 1.22, with a few outlier trajectories that required greater correction. Forty-one percent of the trajectories required less than a 5% correction. The recordings between the two cameras were synchronized and, in principle, should have allowed reconstruction of the trajectory in three dimensions. Such reconstruction was not possible because of the accumulation of missing data, asynchronously observed by the two cameras, for each trajectory. However, it was recently reported that 2D motion capture with appropriately defined movement descriptors can be a powerful method for detecting clinically relevant changes (Marcroft et al., 2014). We thus decided to perform a combined, independent analysis of the videos recorded by each camera as the two cameras provided orthogonal views of the infants' HM.

The coordinates of the trajectories were used to calculate descriptors characterizing the infant's HM. Four types of descriptors were calculated (Figure 2), covering the main descriptors already reported in the literature (Marcroft et al., 2014).

The outlier values for each descriptor were evaluated and used to remove either aberrant points in the trajectory or aberrant values of a descriptor for a specific trajectory. The threshold to consider a value as aberrant was set at the median plus eight times the standard deviation. The points of the trajectory or the calculated descriptor were then considered to be outliers and were replaced with missing values.

The statistical analyses were performed on the calculated descriptors using the R statistical package (version 3.0.1). The similarity between different descriptors was assessed using hierarchical clustering based on the Pearson correlation index and only sixteen descriptors with correlation below 0.8 were retained. The normality of the distributions were evaluated using the Wilk–Shapiro test. None of the variables was considered as normally distributed and the differences between cohorts were thus assessed using the non-parametric Kruskal–Wallis one-way analysis of variance test based on the ranks. The test was followed by a Dunn test to more precisely determine the groups with different effects. The interactions between age and cohort and between age and sequence were tested with a type III ANOVA.

Infants with WS have a developmental age (DA) that is lower than the norm, which could bias the results for this specific cohort. We used the DA measured with the Brunet-Lezine scale to control for this effect.

A total of 94 infants (47 girls and 47 boys) were included in the study. For all infants, recordings were performed at various time points from the age of 2 to 10 months, totaling 289 entries. The recruitment and follow-up of the infants during these 8 months was heterogeneous across cohorts (Figure 3) with a weighted average age similar for all cohorts (~6 months), except for the WS cohort, for which infants were recruited at a slightly higher age, thus shifting the weighted average age to 8 months. The follow-up of the WS cohort was concentrated in the last few months of the 2–10 month period, as the symptoms of WS often appear between 4 and 7 months of life. The average number of visits per infant also varied across cohorts with the highest number of visits for the EH cohort (5.3 visits per infant) and the lowest number for the WS cohort (1.6 visits per infant).

Each recording session consisted of three video sequences that were followed by two cameras, generating a total of 1,734 video sequences. The tracking framework generated 1,446 trajectories (83% of successful coordinate extractions). The main reasons for unsuccessful extraction of the coordinates were deviations from the protocol, either in the compliance of the observational setup (e.g., the mother remaining between the infant and the camera for the entire duration of the recording or a wristband of the same color as the infant's clothes). When possible, manual tracking of missing episodes of the target wristband was conducted (wristband not visible but the hand not obstructed by the mother) to extract the coordinates. The obtained trajectories were equally distributed between cameras and sequences. The comparison between the y coordinates of the two orthogonally positioned cameras (Figure 1, second panel) suggested that the trajectory extraction was reproducible. This was further confirmed by the lack of effect on the camera on all measures, based on the vertical coordinates (min p-value: 0.5), as well as the velocity, acceleration, and pauses for which the two cameras were expected to capture similar information. After removing trajectories containing <70% of the expected coordinates, a total of 1,355 trajectories was obtained and used to derive the descriptors of the hand trajectories (Table 2).

Descriptors were extracted from the trajectories to characterize four aspects of the infants' HM: (i) their velocity and acceleration to assess smoothness, which characterizes the development of reaching (Thelen et al., 1996); (ii) their shape through the curvature of the movements; (iii) the space explored through the amplitudes and their variability for each axis, as well as for the entire trajectory; and (iv) the overall activity during the recorded sequence (through different measures of the pauses) (Figure 2).

The HM in the three environmental settings were different for all descriptors, except for the number of pauses, for which the mean value of approximately four pauses per recorded sequence was observed overall. In all cases, the sequence in which the infant played with the cuddly giraffe was different from the other two when the sequence effect was significant. The infants' HM in this sequence were characterized by larger amplitudes, higher velocity and accelerations, shorter pauses, and more overall time spent in movement. Only three descriptors differed (xSd, ySd, and PausePerc) between the free play and nursery rhyme sequences. The amplitudes of the movements (xSd, ySd) and the time spent in movement (PausePerc) were slightly lower for the sequence with the nursery rhyme. The strong similarity of the movements in the latter two sequences probably reflects the similarity of the communication between the infant and the mother in these two settings, as opposed to playing with the cuddly giraffe.

There was a significant age effect for all descriptors, except for those related to pauses (p = 0.28). Spatial, curvature, acceleration, and velocity descriptors were highly dependent on age, with the strongest associations for the spatial descriptors, more specifically vertical amplitude (Table 3). The 40% greater amplitude (yRange p-value of 6e-25) reflects, in part, the growth of the infants' arms during the study. The increase in movement velocity with age was also highly pronounced: from a 25% increase for vMean to a 50% increase for the acceleration metrics.

The HM were also different by context (Table 3). We tested the relationship between the HM, the three experimental interactive settings, and the infants' age in the interaction between age and sequence in a linear regression model for the various descriptors. The interaction between age and sequence was significant (p < 0.05) for descriptors associated with vertical amplitude (ySd, yRange), velocity (vMean, vSd), and time spent in motion (pauseMean, pausePerc), but not acceleration, suggesting that the evolution of the movements differed depending on the situation. The cuddly giraffe sequence was found to be different from at least one other sequence.

We compared infants in the control cohort at three different ages with the three sequence settings to better characterize typical development of the infants' HM in these different interactive settings (Figure 4). The infants' HM were very similar at early stages of development (three to 4 months), irrespective of the descriptor (non-significant p-values) in the three interaction sequences. By 5–6 months, HM characterizing the interaction sequence with the cuddly giraffe were significantly different from those of the other two sequences for nine of the 16 descriptors. The largest difference was seen for descriptors related to movement acceleration (accSd, accMax; p < 0.001) and the surface explored (ampMax, yRange; p < 0.002). The differences in the infants' HM in the cuddly giraffe sequence, with respect to those of the other two sequences, further increased after the seventh month, when most of the descriptors became statistically different. At this point, the biggest difference was for surface exploration, especially vertical amplitude (yRange, p < 0.0001), reflecting reaching gestures (Figure 4, top panel). The difference observed in the overall time spent in movement for this sequence (Figure 4, bottom panels) became statistically significant (p = 0.01) only after the seventh month.

We estimated the interaction between age and cohort in a linear regression model to compare the developmental trajectories in the clinical cohorts, taking into account all the three sequences. Representative results of the models are shown in Figure 5. The movements of the infants with the VIM cohort evolved differently from those of the control cohort for several descriptors related to velocity, acceleration, amplitude, and extreme curvature of the HM, with an observed increase that was slower than that of the control cohort (sometimes null and even the reverse) (Figures 5C,D, blue line). In the PB cohort, the evolution of the variability of acceleration and velocity was significantly lower than that of the controls (Figure 5D, red line). The OD cohort showed a stronger increase than the controls for descriptors associated with the vertical amplitudes (ySd and yRange) of the movements (Figure 5C, gray line). The WS cohort differed significantly from the controls for the pause and curvature descriptors (pauseNb, curvMean, and curvMax), with an increase in the number of pauses (Figure 5A, green line) and a decrease in the average curvature of the movements (curvMean) between 3 and 10 months, whereas the control cohort presented no significant evolution (Figure 5B, green line) for either of these descriptors.

We performed a specific analysis of the WS cohort, using both chronological age (CA) and DA, to evaluate whether the difference in evolution of the HM in the WS cohort was associated with the infants' development. Among the 13 descriptors showing an age effect in the population of all cohorts, six (vMean, vSd, accSd, accMeanMov, decMaxMov, and ampMax) exhibited no significant evolution with CA for the WS cohort, but showed a significant evolution when the global DA was considered.

We tested whether interactions remained significant for descriptors that showed an interaction between age and cohort and differentiated the WS cohort from the control cohort (pauseNb, curvMean, and curvMax), when the DA was used instead of the CA for the WS cohort (Figure 6). There was no interaction for the number of pauses (not shown) nor average curvature (Figure 6B) when DA was used for the WS cohort relative to the control cohort, indicating that their evolution was similar. The difference in the evolution between the WS and control cohort was still present despite this age correction for the descriptor maximum attained curvature (curvMax), which reflects greater jerkiness of the movements.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The reviewer, DC, declared a past co-authorship and a shared affiliation with one of the authors, KB, to the handling Editor.