How do babies roll? Identifying the coordinated movements of infant rolling through video compared to laboratory techniques

Abstract

BACKGROUND:

Rolling is an important developmental milestone for infants where identifying the coordinated movement patterns could facilitate the early identification of motor development delays. Current methods for identifying coordinated movements of rolling are limited to a laboratory setting and not feasible for clinicians.

OBJECTIVE:

To develop video-based methods in which six coordinated movements, previously defined through motion capture, can be identified through video alone.

METHODS:

Forty-five videos of sixteen healthy infants achieving a roll were used to develop the video-based methodology and twenty-four videos had corresponding motion capture data used for validation. Four raters comprised of researchers and a clinician identified rolling coordination using the new video-based methods. A Fleiss’ Kappa statistical test determined the inter- and intra-rater reliability of agreement for the new methodology and compared it to motion capture.

RESULTS:

The comparison of the motion capture and video-based methods resulted in substantial agreement. The video-based methods inter- and intra-rater reliability were substantial and almost perfect, respectively.

CONCLUSIONS:

We developed reliable methodology to accurately identify the coordinated movements of infant rolling using only 2D video. This methodology will allow researchers to reliably define coordinated movements of infants through video alone and may assist clinicians in identifying possible motor development delays and disorders.

1. Introduction

Achieving a roll is a key motor skill and an important developmental milestone for infants. To initiate and complete a roll, infants must use whole-body, goal-oriented movements that take them from a supine to prone, or prone to supine position [1]. Motor abilities, like rolling, play an important part in the cognitive and motor development of infants [2]. Rolling specifically allows infants to interact with their external environment in daily life and encourages muscle development and postural control needed to achieve more advanced movements like sitting, crawling, and walking [3]. The assessment of infant rolling techniques are also an important component in in the early identification of motor development delay [4]. The early identification of these motor delays can improve a child’s ability to learn new skills and increase their success in school and life [5]. Currently, identifying motor delays is most often done by parents or guardians completing a developmental screening survey, where the results are reviewed by clinicians [6]. These developmental screening surveys are broad, focusing solely on when larger milestones, like picking up toys or rolling, are achieved. Because of the broad nature of these surveys, some children have a delayed diagnosis or do not receive one at all [7]. By increasing the knowledge of the specific motor milestone of rolling, the number of infants receiving an early diagnosis can also increase. To do this, previous studies have identified different rolling patterns correlating with age [8,9]. However, these studies have been completed in a laboratory setting, or the methods and instrumentation have since been improved upon, making them inaccessible to parents or clinicians. Therefore, there is a substantial need for methodology that can be used by clinicians to easily identify specific infant rolling techniques without the use of complex laboratory equipment to assist in the early identification of motor development delays in infants.

Six qualitative categories of coordinated rolling movements have been identified based on limb coordination [8]. These movements were identified using motion capture techniques, which are considered the gold-standard for tracking human motion due to their high levels of precision [10]. The six different coordinated movements are described in Fig. 1 where the limbs are codified as ipsilateral arm (IA) and leg (IL) and contralateral arm (CA) and leg (CL) in the direction of the roll. However, using motion capture techniques to describe the coordinated movements of infant rolling requires advanced biomechanics knowledge, a costly motion capture system, and is time-consuming. Furthermore, it is not an accessible method for clinicians to identify an infant’s rolling patterns over time for a potential motor delay diagnosis. For clinicians to determine an infant’s specific coordinated rolling patterns, it is crucial to develop methodology that can be completed both quickly and accurately.

Fig. 1.

Six previously established coordinated movements of rolling showing the sequence throughout the movement [8]. The limbs are codified as ipsilateral arm (IA) and leg (IL) and contralateral arm (CA) and leg (CL) in the direction of the roll. Illustrated rolls are to the infant’s left. The limbs marked with an asterisk (*) are stationary throughout the roll, the dark gray limbs initiate the roll, and the light gray limbs follow after the roll has already been initiated. Figures and captions adapted and reproduced with permission.

Two-dimensional (2D) video analysis has grown in popularity with the increased availability of smartphones and other mobile recording devices. In the world of biomechanics, 2D video analysis is frequently used when retrospectively studying the mechanisms of sports injuries or athletic techniques which cannot easily be replicated in a laboratory setting [11,12]. For example, in anterior cruciate ligament (ACL) injuries, video-based methods have been developed to identify individuals at an increased risk for ACL injury [13]. These methods can be used with simple video and have been verified with motion capture analysis. It also has the potential for use as a screening tool to determine ACL injury risk in young athletes [14]. In adults, video is being used to study patients with Parkinson’s disease [15] and for telehealth appointments [16,17]. For developmental pediatrics, video analysis has been used to study movement disorders and movement patterns in a variety of developmental diseases [18–20]. One study demonstrated similar efficacy between home videos and in-person observation of the gross motor development of infants [21]. This study also reported high parent satisfaction with the ease of recording videos. Another study also reported value in home videos for monitoring neonatal development outside of a hospital or clinic setting [22]. Thus, the natural next step would be to implement similar video techniques when monitoring and studying rolling in both a laboratory and clinical setting, as well as incorporating at-home videos. Therefore, the purpose of this study was to develop video-based methods in which six coordinated movements, previously defined through motion capture, can be identified through video alone.

2. Methods

2.1. Participants

This study is part of a larger prospective study exploring muscle activation during infant rolling. Thirty-eight healthy infants (age: 6.5 ± 0.7 months; 23M/15F) were included in this IRB-approved (#126-MED20-005, Originally Approved: 06/01/2020) in-vivo biomechanics study. Infants were between the ages of 4 and 7 months, > 37 weeks gestation, between the 5^th and 95^th percentile for birth height and weight, and had successfully achieved at least two independent supine to prone rolls at the date of testing. Infants were excluded if they had any diagnosed orthopaedic or neurological conditions that may cause motor or developmental delays. Before participating, guardians provided assent and completed an Ages and Stages Questionnaire, an at-home screening tool that monitors development [6]. Infants were excluded from the study if they scored below the gross motor category cutoff for their age range where further assessment by a professional is recommended (6-month cutoff score = 22.25).

2.2. Experimental procedures

2D video and three-dimensional (3D) motion capture data were collected simultaneously during an infant rolling movement, defined as the supine to lateral rotation of the torso [8]. A retroreflective marker-based motion capture system (Vicon, 100 Hz) tracked infant kinematics using custom 3-marker rigid body clusters (6.5 mm markers) and 9 mm individual markers. Rigid body clusters used for analysis were placed on the torso and both feet while individual markers were placed on the backs of both hands (Fig. 2). A GoPro camera recorded video of each trial and a pulse oximeter monitored heart rate and SpO₂ levels. Based on the clinical safety standard for SpO₂, if the infant’s SpO₂ reading was < 95% for more than five seconds [23,24], the testing was ended for that trial. Infants were placed in a supine position on a playmat for five minutes or until they achieved a complete supine to prone roll. During the trial, infants were encouraged to roll by guardians or researchers offering toys, food, or simply calling out to them. Infants were excluded from the study if they did not achieve a roll within the five-minute testing period.

Fig. 2.

(A) Experimental setup showing retroreflective marker placement and the pulse oximeter (SpO₂ monitor) used to monitor safety. Electromyography sensors wrapped in self-adherent bandage were also used in this experimental setup but were outside the scope of this study. (B) Retroreflective marker placement used for data analysis where 3-marker rigid body clusters (6.5 mm markers) were placed on the torso and both feet and 9 mm individual markers were placed on the backs of both hands.

2.2.1. Motion capture analysis

We developed custom MATLAB code to analyze the motion capture data and determine the coordinated movement chosen, similar to previously developed procedure [8]. The data was filtered in the direction of the rolling movement using a 4^th order, 6 Hz Butterworth filter and truncated to the first 25% of the rolling movement. The start of the rolling movement was determined based on the first limb movement initiating a roll, beginning from a neutral supine position. The rolling movement was marked as completed when an infant’s trunk reached a lateral position. The marker most visible on the rigid body clusters was used for analysis. Rolling movements were excluded from analysis if any trunk or limb markers were occluded for 25 or more consecutive frames.

The speed of each limb and the trunk was calculated and normalized to the speed of the torso. Since previous research has determined that only the limbs on the ipsilateral side could be stationary and any limb could be moving [8], if the normalized speed of the IA or IL was less than 125% the limb was considered stationary, otherwise the limb was considered moving. Once each ipsilateral limb was defined as stationary or moving, time series plots showing the normalized speed of all limbs were used to determine the limb coordination (Fig. 3). Once all coordination was determined, the rolls were defined as one of the six previously established coordinated movements [8].

Fig. 3.

Time series plots showing example limb coordination speeds for movements B (Right) and C (Left) for the first 25% of the rolling movement. The blue line represents the contralateral arm (CA), the orange line represents the ipsilateral arm (IA), the yellow line represents the contralateral leg (CL), and the purple line represents the ipsilateral leg (IL). For coordinated movement B (Right) the CA initiates the rolling movement. For coordinated movement C (Left) the CA, CL, and IL work together to initiate the rolling movement.

2.2.2. Development of video-based methods

To determine the coordinated movement type used during a roll from video alone, we created detailed methodology (Supplementary Data) to assist in the roll type characterization. Before using these methods, the video quality must be assessed to ensure the accuracy and reliability of classifying the coordinated movements. Video criteria included (1) all limbs must be visible throughout the roll, (2) the infant must not be holding any toys or objects, (3) The video must begin before or at the start of the roll, and (4) the camera must be in close proximity (~5 ft) of the infant.

The methods begin by defining the important characteristics of a roll needed to differentiate the six different coordinated movements previously established [8]. These definitions include how to identify a rolling movement in a video, how to define the limbs in reference to the rolling direction, and how to define stationary versus moving limbs. Once these concepts are understood, the methodology takes the user through the three-step procedure for determining the coordinated movement: (1) identifying the direction of the roll; (2) identifying stationary and moving limbs; and (3) determining the synchronicity of the limbs. Finally, key information is provided about each of the six previously defined coordinated movements that is specifically adapted for video analysis of rolling. A flow chart is also provided to assist in the identification of the different movements shown in Fig. 4.

Fig. 4.

Flow-chart used to guide the identification of the six previously defined coordinated movements based on the number of stationary limbs and the coordination of the moving limbs.

2.2.3. Video-based methods analysis

A total of four reviewers used the final methodology to categorize the 45 videos. Three reviewers were researchers who were familiar with the different infant rolling maneuvers before rating. One reviewer was a pediatric clinician who had no previous knowledge of these specific infant rolling coordinated movements. Each reviewer determined the coordinated movement used for each of the 45 videos in a random order. Reviewers were only allowed to use the video-based methodology documentation for reference. Once all four reviewers completed the categorization, a statistical analysis was performed using the SPSS statistical package (version 26; SPSS Inc., Chicago, IL). Using a Fleiss’ Kappa statistical test, each reviewer’s response was compared for each video to determine the inter-rater reliability. This analysis was completed to determine the reliability of agreement for each individual coordinated movement category and an overall agreement score [25]. The interpretation of the Fleiss’ Kappa agreement was slight (0.00–0.20), fair (0.21–0.40), moderate (0.41–0.60), substantial (0.61–0.80), and almost perfect (0.81–1.00) [26]. One research reviewer determined the coordinated movements an additional two times for the same set of 45 videos in different random orders, with a minimum of one month between each review. The intra-rater reliability was then found using the same Fleiss’ Kappa statistical analysis comparing responses of the different reviews for each video [25]. Based on previous studies using this statistical analysis with an infant population, three reviewers are sufficient to determine the inter-rater reliability of agreement and one reviewer is sufficient for the intra-rater reliability of agreement [27–29].

2.2.4. Statistical comparison

To determine the accuracy of our 2D video-based methods compared to the gold-standard of 3D motion capture, we performed a Fleiss’ Kappa statistical analysis using the SPSS statistical package (version 26; SPSS Inc., Chicago, IL). The results of the motion capture analysis and the corresponding video responses from each of the four reviewers were compared to determine the inter-rater reliability of agreement on each individual coordinated movement category and the overall agreement score [25].

3. Results

No infants were excluded due to their Ages and Stages Questionnaire results. From the data collected, sixteen infants (6.6 ± 0.7 months; 9M/7F) had acceptable video data where at least one rolling movement was achieved during data collection. Fourteen of these infants (6.6 ± 0.7 months; 8M/6F) had acceptable corresponding motion capture data. This resulted in 45 rolling movements for the video-based methods analysis with 24 rolling movements having corresponding usable motion capture data. Fourteen infants did not complete a rolling movement in the five-minute period and were excluded from analysis. An additional eight infants did not have adequate video data based on the video-based methodology criteria and were also excluded. Table 1 shows the demographics and developmental screening results of our included participants.

Table 1.

Demographics of the motion capture and video-based methods participants, where the motion capture participants are a subset of the video-based methods participants

Variable category	Motion capture N = 14	Video-based methods N = 16
Age in months, mean (SD)	6.6 (0.7)	6.6 (0.7)
Gestational age at birth in weeks, mean (SD)	39.3 (1.3)	39.3 (1.3)
Height in cm, mean (SD)	65.0 (2.9)	64.8 (3.0)
Weight (kg)	7.8 (1.0)	7.7 (1.0)
Ages and stages gross motor score, mean (SD)	41.1 (11.5)	43.4 (11.9)
Sex, n (%)
Male	8 (57.1)	9 (56.3)
Female	6 (42.9)	7 (43.7)
Ethnicity, n (%)
White	11 (78.6)	12 (75.0)
Asian	2 (14.3)	3 (18.8)
Hispanic	1 (7.1)	1 (6.2)
Rolling movements
Achieved, n	24	45

3.1. Motion capture analysis

Comparing the motion capture rolling movements with their corresponding video responses of the four reviewers resulted in an overall Fleiss’ Kappa reliability score of 0.764, making the agreement between methods substantial (Table 2). Coordinated movement A had almost perfect agreement (κ = 0.830), while coordinated movements C and D had substantial agreement (κ = 0.776 and κ = 0.678, respectively). Coordinated movements B and F had moderate agreement (κ = 0.603 and κ = 0.548, respectively) and coordinated movement E had only slight agreement (κ = 0.147).

Table 2.

Comparison of motion capture and video-based methodology with Fleiss’ Kappa statistical analysis

Variable category	Reviewer 1a	Reviewer 2a	Reviewer 3a	Reviewer 4a	Motion capturea	Inter-Kappab	Agreementc
Overall	–	–	–	–	–	0.764	Substantial
A	1	2	1	1	1	0.830	Almost perfect
B	5	4	6	5	5	0.603	Moderate
C	6	5	9	8	5	0.776	Substantial
D	5	6	4	5	2	0.678	Substantial
E	2	1	3	0	5	0.147	Slight
F	5	6	1	5	6	0.548	Moderate

^aNumber of each coordinated movement identified where Reviewer 1 (first review), 2, and 3 are researchers and Reviewer 4 is a clinician.

^bThe overall and individual coordinated movement inter-rater reliability of agreement score determined from a Fleiss’ Kappa statistical analysis.

^cAgreement score corresponding to the Fleiss’ Kappa statsitcal analysis.

3.2. Video-based methods analysis

The overall inter-rater reliability Fleiss’ Kappa score comparing the results of our video-based methods from our four reviewers using 45 videos was 0.660, making the agreement substantial (Table 3). Coordinated movement types A through D exhibited substantial agreement of roll type identification between all four reviewers ranging from κ = 0.676 to κ = 0.795. Coordinated movements E and F had the lowest reliability scores, with only fair (κ = 0.295) and moderate agreement (κ = 0.470), respectively. An overall intra-reliability score of the Fleiss’ Kappa statistical analysis was 0.848, making the agreement almost perfect (Table 3). Each individual coordinated movement also had an almost perfect agreement score ranging from κ = 0.815 to κ = 1.000.

Table 3.

Video based methodology Fleiss’ Kappa statistical analysis comparing the inter- and intra-rater reliability of agreement

Variable category	Reviewer 1.1a	Reviewer 1.2a	Reviewer 1.3a	Reviewer 2a	Reviewer 3a	Reviewer 4a	Inter-Kappab	Agreementc	Intra-Kappad	Agreementc
Overall	–	–	–	–	–	–	0.660	Substantial	0.848	Almost perfect
A	1	1	1	2	1	1	0.794	Substantial	1.000	Almost perfect
B	8	11	9	8	10	8	0.795	Substantial	0.865	Almost perfect
C	13	10	13	12	20	17	0.676	Substantial	0.848	Almost perfect
D	10	8	9	10	9	9	0.757	Substantial	0.815	Almost perfect
E	3	2	3	2	4	1	0.295	Fair	0.867	Almost perfect
F	10	13	10	11	1	9	0.470	Moderate	0.840	Almost perfect

^aNumber of each coordinated movement identified where Reviewer 1 (1.1 is the first review, 1.2 is the second review, and 1.3 is the third review), 2, and 3 are researchers and Reviewer 4 is a clinician.

^bThe overall and individual coordinated movement inter-rater reliability of agreement score determined from a Fleiss’ Kappa statistical analysis.

^cAgreement score corresponding to the Fleiss’ Kappa statistical analysis.

^dThe overall and individual coordinated movement intra-rater reliability of agreement score determined from a Fleiss’ Kappa Statstical Analysis.

4. Discussion

The purpose of this study was to develop video-based methods in which six coordinated movements previously defined through motion capture, can be identified through video alone. We hypothesized that (1) when compared to the gold-standard of motion capture techniques, our video-based methods would have substantial agreement, and (2) that the inter- and intra-rater reliability of the video-based methods alone would have substantial agreement.

4.1. Motion capture analysis

Previously, the six coordinated movements have been identified using motion capture techniques. To determine the validity of our video-based methods, we compared the results of the motion capture techniques, the gold-standard for infant roll identification, to our new video-based methods. An overall agreement score between the four reviewers and the motion capture methods was substantial (κ = 0.764), meaning that our video identification methods match the motion capture gold-standard with significant agreement and confirming our first hypothesis. Previously, researchers created video-based methods as a screening tool for ACL injuries comparing to motion capture as the gold-standard [13,14], similar to the development of our methods for infant rolling. In this study, they found good-to-excellent inter- and intra-rater reliability, confirming that video-based methods can provide accurate identification in the same way as motion capture.

Coordinated movement A had the highest agreement between the video-based methodology and the motion capture techniques with almost perfect agreement (κ = 0.830). This is explained by the distinct nature of limb coordination when compared to the other movements as well as the limited number of rolling movements found in this category. Coordinated movements C and D have substantial agreement (κ = 0.776 and κ = 0.678, respectively). These coordinated movements are the only movements with a single stationary limb, the ipsilateral arm, allowing for the easy identification of limb coordination once the stationary limbs have been determined. Both coordinated movements B and F had moderate agreement (κ = 0.603 and κ = 0.548, respectively). These movements may be misidentified due to their similarities to other rolling movements, B to E and F to C, where the only difference is the ipsilateral arm. The coordinated movement with the lowest score was movement E with only slight agreement (κ = 0.147) found between motion capture and video-based methods. This score indicates that coordinated movement E is the most difficult to identify with our current video-based methodology.

4.2. Video analysis

For this study, we developed video-based methodology in which each of the six previously defined coordinated movements for roll initiation could be accurately identified through video alone. This methodology is similar to other video techniques that monitor and study infant development and motor disorders [18–20,22]. Additionally, similar techniques have been used for adults to diagnose diseases like Parkinson’s [15], conduct patient examinations [16,17], and analyze injury risk for ACL tears [14]. In our study, we found that coordinated movements could be identified with almost perfect agreement (κ = 0.848) between a single reviewer, meaning that a single person can repeat these methods with high accuracy. Comparing multiple reviewers, we found that the overall agreement score was substantial (κ = 0.660), confirming the second part of our hypothesis. However, based on the similarities between different movements, key distinctions had to be made.

Coordinated movements B and D are similar, both featuring axial rotation. For coordinated movement B, both the ipsilateral arm and leg are stationary, while the contralateral arm initiates the axial rotation of the roll and the contralateral leg follows. In roll type D, the movement is only changed by the ipsilateral leg synchronously initiating the roll with the contralateral arm. To differentiate these movements through video, reviewers must observe the torso and shoulders during the roll. If the contralateral shoulder and part of the torso have lifted off the floor before the ipsilateral leg moves, this roll is considered coordinated movement B. Otherwise, it is considered movement D.

Through video, coordinated movement A is unique when compared to the other six movements. Foi coordinated movement A, the ipsilateral side is stationary throughout the movement and the contralateral side initiates the roll, differentiating this movement type from the other five coordinated movements. Although coordinated movement A had the fewest number of occurrences, due to the unique movements described, it could easily be identified, resulting in substantial agreement (κ = 0.794) between reviewers.

Coordinated movements C and F are another pair of movements that are similar with only one key difference. For coordinated movement F, all limbs are working together to initiate the roll. Coordinated movement C is the same with the exception of the ipsilateral arm being stationary. With only the ipsilateral arm differentiating these two movements, it was critical to define the difference between moving and stationary limbs. Coordinated movement C exhibited substantial agreement (κ = 0.676) between reviewers, however, coordinated movement F only had moderate agreement (κ = 0.470) between reviewers. This indicates that reviewers may determine the ipsilateral arm as stationary when it is actually moving.

The coordinated movement types that had the lowest inter-rater agreement were F and E with moderate (κ = 0.470) and fair (κ = 0.295) agreement between reviewers, respectfully. These two movements are the only coordinated roll types where both the arms are moving synchronously together to initiate the roll, again relying heavily on the distinction between moving and stationary limbs. Both movements also have similar coordinated movement pairs, F to C and E to B. In their corresponding pairs, the ipsilateral arm is now stationary instead of moving. This indicates that if the movement of the ipsilateral arm is misidentified, the coordinated rolling movement would also be incorrect, resulting in the lower agreement for these movement types. Coordinated movement E, also had a small number of videos with this roll type, however, unlike coordinated movement A, the movements used to achieve this roll type are not as distinct, resulting in only fair agreement (κ = 0.295) for the inter-rater reliability score.

Creating video-based methods to instruct raters on how to consistently determine whether a limb was stationary or moving during the roll was a challenge. While some limbs are defined as stationary, they are not necessarily immobile during the roll. Stationary limbs provide support and balance during the roll, so some movement is expected. To differentiate between stationary and moving limbs, our methodology states that moving limbs are usually above or across the torso during the roll while stationary limbs usually stay lower to the surface and closer to the body. However, this definition did not always account for the ipsilateral arm moving as well, where the arm does not cross the body and may not move upwards when initiating the roll. The limitations in the moving limb definition for the ipsilateral arm explain the lower agreement level for both E and F movements. In the future, improvements to these definitions may improve the distinction and agreement between these coordinated movements. It is also possible that infants will exhibit rolling coordination that cannot be categorized as one of these six previously defined techniques. One developmental rolling study describes a pre-rolling movement that features head/neck and trunk extension [9] that may not easily be categorized into the six movements previously defined [8]. Although all of the rolling movements analyzed in our study were categorized as one of the six coordinated movements, for unusual techniques, we suggest that raters consider them as an “other” category.

The addition of video-based methodologies will allow clinicians another monitoring technique for infant motor development which can be especially important for premature infants [30]. Monitoring the motor milestone of rolling is a crucial component in the early identification of motor delays and currently many children do not receive the timely diagnosis they need [4,7]. With this new video-based methodology that increases the accessibility of infant monitoring, the number of infants receiving an early diagnosis may also increase.

4.3. Considerations

One limitation of our video-based methodology is that rolling was only analyzed on a flat, firm surface and does not incorporate movements that may be used in different mechanical environments or commercial infant products. In addition, only six different coordinated movements have been fully defined, and other movements may exist to achieve a roll. Future studies should expand the mechanical environments in which rolling is taking place to determine if the techniques of rolling change and determine if the video-based methodology should be expanded upon to incorporate those different techniques.

5. Conclusion

We developed reliable methodology to accurately identify the coordinated movements of infant rolling using only 2D video. Using motion capture analysis, we were able to determine limb coordination based on the speed of the limb in relation to the torso. This allowed us to identify the six different coordinated movements using motion capture and thus validating our video-based methods. These methods will provide clinicians with another resource in identifying infant rolling patterns and techniques. This study is the first that we know of that distinguishes coordinated movement patterns of infant rolling through video alone. While other studies have demonstrated rolling categorization through in-person observations [9] and advanced laboratory equipment [8], these methods can be hard to accomplish in a clinical setting and too complex for laboratories completing long-term studies on infant rolling development. The new video-based methods developed in this study will allow researchers to reliably define coordinated movements of infants through video alone. In the future, adding a correlation of age and coordinated movement to out video-based methodology may also further assist clinicians in the diagnosis of possible motor delays and disorders.