Assessing Infant Gross Motor Performance With an At-Home Wearable

Assessing gross motor skills and the reaching of milestones are part of the mainstay of well-child care worldwide.^1,2 They are typically assessed several times during the first 2 years of life to ensure an age-appropriate development. The assessment usually relies on parental interviews complemented by brief observations as feasible in the given pediatric clinics.³ Gross motor skills are considered to provide efficient surrogate measures of wider neurodevelopment, with significant links to earlier adversities, environmental and other acquired effects, as well as later neurodevelopmental compromise.^3–10 

The global use of reaching gross motor milestones (GMMs) is based on the implicit assumptions that (1) they would be universally present and (2) that their recognition would be unequivocal to lay observers such as the child’s parent. However, the everyday experience and several studies challenge both assumptions. First, there are notable cultural, environmental, and historical variations in the developmental expression of milestone skills.^3,11–13 Second, GMM assessment relies heavily on parental information that is unavoidably subjective and potentially challenging in multicultural and/or multilinguistic environments.¹⁴ In addition, using discrete milestones for longitudinal tracking of development is often challenged by the considerable intra- and interindividual variation over time.^15–17 These considerations together underscore the scarcity of, and the unmet need for, genuinely objective, quantifiable, and ecologically valid measures for longitudinal neurodevelopmental follow-up.

Intuitively, infants’ performance is optimally assessed at home during spontaneous daily activity.^18,19 At-home assessments became possible recently with advanced multisensor wearable suits coupled with machine learning-based analyses,^20–23 which together may provide accurate and explainable analyses of infants’ motor performance during early development.^21,22 Here, we studied whether wearable methodology with at-home measurements could be used for the clinically important aspects of early motor assessment, including (1) detection of reaching the key GMMs,^4,24 (2) quantification of times spent in different postures, and (3) longitudinal tracking of individual’s gross motor development.

The overall study design (Figure 1A) included repeated at-home wearable measurements by the parents,^21,25 which were used to develop and validate automated algorithms for the following purposes: (1) to detect the reaching of GMMs, (2) to quantify times spent in different postures, and (3) to longitudinally track an individual’s gross motor development. The study was reviewed by the Ethics Committee and approved by the New Children’s Hospital at the Helsinki University Hospital, Helsinki, Finland (HUS/80/2021), and written informed consent was obtained from the legal guardians.

Two infant cohorts were prospectively and independently recruited for a longitudinal follow-up between ages 4 and 22 months (Table 1; Supplementary Figure 1). Cohort 1, with typically developing infants, was used for the classifier development and cross-validation. Cohort 2 was used for external validation, consisting of infants attending pediatric neurology follow-up due to an identified or suspected developmental risk. This cohort was considered external for the purpose of testing model generalizability because it was recruited separately from a different clinical follow-up program. The recruitment is described in detail in Supplementary Methods 1 and 2. By the time of concluding the present work, the infants in cohort 2 presented with the following clinical characteristics: 8 infants had a spastic hemiplegia, 3 infants had motor and cognitive delay, and 2 infants had motor delay demanding ongoing follow-up. The remaining 24 infants of cohort 2 had transient motor asymmetry, extensor tone, or mild motor delay, which normalized later.

The wearable suit (Motor Assessment of Infants with a JUmpsuit [MAIJU]; Figure 1) is a commonplace infant garment with movement sensors attached to standard locations in each limb.^21,22,25 In this study, the sensors collected accelerometer and gyroscope data at 52 Hz (1024 samples per second) over Bluetooth connection to a nearby mobile device. We used local courier service for delivering the suit to the homes and back. The parents dressed the infant in the suit and encouraged the infant to be spontaneously active (ie, “free play”) for the rest of the day or for at least 1 hour.²⁵ Upon returning to the laboratory, the data were uploaded from the mobile device to a computational server “Babacloud” for a fully automated analysis pipeline^21,25 (Supplementary Methods 1 and 2).

The parental questionnaires were collected at each measurement session to serve as the reference for the algorithm development. The parents were requested to assess the infant’s typical performance during the previous week. Multioption scales²⁶ were used to mitigate ambiguities in the verbal GMM definitions⁴ and to enable alternative definitions during analyses (Supplementary Figure 2). The parents also estimated the typical amounts of time spent in each posture from 1 (ie, least time) to 9 (ie, most time), mimicking real-life health care visits (Supplementary Methods 3).

To maximize global validity, we adopted GMMs from the WHO’s multicenter reference study,^4,24,27 which provides 2 benchmarking measures: (1) the cohort-level age range of reaching the GMMs²⁷ and (2) detection accuracy estimated by expert’s interrater agreement levels.²⁴ We focused primarily on the GMMs performed without support, ie, sitting, 4-limb crawling, standing, and walking. Additionally, “prone crawling” (ie, commando crawling) was taken to reflect early independent moving. The supported types of standing and walking are reported in Supplementary Figure 3 for a full parallel to the WHO’s reference study.^4,24 

For assessing the reaching of each GMM from the wearable data, we developed GMM-specific support vector machine classifiers using parental questionnaires as the target. These GMM classifiers were based on the second-by-second-level detections of infants’ postures and movements (“MAIJU features”; Figure 1B) during the measured play time, which all come automatically from MAIJU’s analysis pipeline.^21,28 Here, we used those postures and movements that were behaviorally relevant for each specific GMM (Figure 1B; Supplementary Tables 1–2). This top-down approach supports maximal transparency, interpretability, and a full backward tracking of the results in each individual infant. Notably, it follows the well-known rationale in other areas of medicine, such as cardiac monitoring, in which underlying R-peak detections are used for higher-level presentations of the cardiorespiratory status (arrythmias, heart rate variability, etc).²⁹ 

The GMM classifiers for 4-limb crawling, sitting, and prone crawling were binary (ie, yes/no), whereas the GMM classifiers for walking and standing consisted of 3 alternative categories (ie, no/supported/unsupported). Because it is possible that an infant does not spontaneously show all of the learned milestone skills during each measurement session, we devised the common clinical rationale as a decision tree algorithm (Supplementary Table 3): reaching a more advanced GMM implies reaching the less advanced GMMs as well. For instance, detection of walking would indicate that the child has also reached the ability to crawl even if it was not directly measured during the given session. All classifiers were developed using MATLAB software (MathWorks, Inc.).

We recently showed²¹ that the combination of all of an infant’s postures and movements (Figure 1A) during a wearable measurement session can provide a holistic estimate of motor maturity (Supplementary Table 2). Such quantified measure of a holistic gross motor maturation was conceptualized as a normalized index BABA infant motor score (BIMS). In the BIMS scale, 0 denotes motor performance with supine-dominant preference without rolling (corresponding to an infant aged approximately 4 months), and 100 denotes motor performance with dominant fluent walking (corresponding to a an infant aged approximately 16 months). Therefore, the BIMS of a typically developing infant is expected to correlate strongly with the chronological age (therefore also called “developmental age prediction”).²² 

Leave-one-subject-out cross-validation (cohort 1 data) was used to assess performance of the GMM detection algorithms. This means that all measurements from a single subject were used as the test data at a time, and all other measurements were used for model development. The external validation (cohort 2) was done with algorithms developed and validated with the whole cohort 1.

A two-sample Kolmogorov-Smirnov test (K-S, two-tailed P values) was used to compare empirical cumulative distribution functions (Figure 2) with the null hypothesis that they are drawn from the same underlying normal distribution.

A bootstrapping method with 10 000 randomly drawn samples was used to obtain the 95% CI for the GMM detection metrics (Figure 3).

Spearman’s ρ was used to assess correlations between wearable-derived quantitations of infants’ time spent in each posture and the parental assessments of typical postural prevalences (Figure 4), as well as between BIMS and infants’ age at the time of measurement (Figure 5C). Because individual recordings are statistically independent against the comparison targets (ie, parental assessment, age), the correlations and two-tailed P values are reported without controlling for infant identities (IDs).

Linear mixed-effects modeling (LME) modeling was used to account for constant ID-specific random effects in the age-dependent BIMS scores (Figure 5). In practice, this approach accounts for the different levels of individual growth trajectories that remained strikingly stable over time. The LME model implementation was obtained from MATLAB Statistics & Machine Learning Toolbox (MathWorks, Inc.) with the following formula: Age ∼ BIMS + (1|ID).

Cohort 1 (used for the development and cross-validation of the algorithm) consisted of 97 typically developing infants (47 boys) who were measured repeatedly (1–8 times per infant; median, 5) at age 4 to 19 months for a total number of N = 463 measurements. The external validation cohort (cohort 2) consisted of 37 infants (23 boys) measured repeatedly (1–7 times per infant; median, 6) at age 4 to 22 months for a total number of N = 157 measurements. Detailed comparison of the cohorts is presented in Supplementary Table 4 and Supplementary Figure 1.

We first visualized the cohort-level reaching of each GMM, allowing a convenient comparison among the GMM algorithm, parental information, and the prior multicenter reference cohort reported in the WHO study.²⁷ The cohort-level reaching of GMMs (Figure 2) was nearly identical between automated detections and the parental questionnaires in all milestones (K-S P > .1). Comparison with the WHO reference age windows²⁷ shows high concordance, except for “sitting without support,” which showed later reaching in our study. This is likely due to different definitions; in our study, the parents assessed the child’s typical spontaneous behavior, whereas the WHO study tested the child’s ability to sit when purposefully tested.⁴ Taken together, the cohort-level reaching of GMMs is comparable between the classifier and parents, and our infant population compares well with the WHO reference data.

Accuracies of the GMM detection algorithms were assessed by comparing them with the parental questionnaire information. We also benchmarked the algorithms’ GMM detection accuracy with the empirical threshold of human equivalent performance,^20,30 ie, the level of interrater agreement among trained experts in the WHO data.²⁴ Finally, algorithm accuracies were compared with the “clinical baseline,” in which reaching of GMMs is predicted from an infant’s age alone.

The automated detection of reaching the GMMs (Figure 3; Supplementary Figure 4) was very accurate for independent walking and 4-limb crawling (accuracy and κ ranging from 94.3% to 96.8% and .86 to .93, respectively, for both cohorts) and almost as high for standing (90.9%-92.4% and .82-.84, respectively), sitting (93.1%-95.8% and .79-.84, respectively), and prone crawling (94.3%-95.5% and .77-.81, respectively). Importantly, these accuracies are at a par with human-level performance, as they clearly overlap with the interrater agreement levels in the WHO reference study.²⁴ The algorithms’ accuracies were also found to generalize very well to the external validation data (cohort 2). As a positive control, GMM predictions from the infants’ age alone (“clinical baseline”) was clearly less accurate in infants attending neurological follow-up clinics (cohort 2). The effect sizes and other finding details are presented in Supplementary Table 4 and Supplementary Figure 4.

The fully automated quantitation of infants’ postures was then compared with the parental estimates of time that their child is typically spending in the given postures (numerical scale 1 to 9, with higher number indicating more time spent). We found strong and highly significant correlations between wearable-derived measures and parental assessments (Figure 4), with the strongest correlations for “prone,” “crawl,” and “standing” postures (Spearman’s ρ = .68, .76, .81, respectively; P < .001). The findings also reflect the fundamental problem in assessing quantities with questionnaires: the parental response is, by default, unitless, ordinal more than quantitative, and not normalized across postures. The actually measured and true times spent in different postures can vary widely and irrespective of their correlations with parental assessments. For instance, parents’ ratings of 1 to 9 for crawl posture correspond to 0% to 40% of the time spent in crawling posture, whereas parents’ ratings of 1 to 9 for standing correspond to 0% to 80% of the time spent in standing posture.

We then compared longitudinal tracking of infants’ gross motor development with 2 approaches: (1) By using the conventional, parentally informed reaching of GMMs (dichotomic, ie, yes/no) and (2) by computing the holistic motor maturation measured BIMS from the wearable data. In the typically developing infants (cohort 1), an ideal developmental measure should correlate with the infant’s age, ie, support predicting the age of the given infant. Such age prediction from the parentally informed GMMs was distinctly stepwise (Figure 5A). In contrast, the BIMS measure produced a strikingly smooth continuum across the full age range (Figure 5B), suggesting far richer information related to the gross motor development (see also Supplementary Figures 5–7 for further details). This finding prompted assessing further how BIMS allows longitudinal tracking of an individual’s developing gross motor performance. Pooling 456 repeated measures from 90 typically developing infants (cohort 1) showed that BIMS values from individual measurements correlates strongly with infants’ age (Spearman’s ρ_RAW = .88, P < .001). Moreover, the individuals’ BIMS trajectories appeared to proceed consistently on their respective percentile levels (“growth channels”), and the age correlations were even higher when accounting for the individual’s mean percentile level (Spearman’s ρ_LME = .93, P < .001; Figure 5C).

Our study shows that unsupervised at-home wearable measurements with fully automated analyses can very accurately (1) detect reaching of GMMs, (2) quantify the time spent in different postures, and (3) track individual-level gross motor development. Importantly, the accuracy in automated milestone detections were comparable with the experts’ agreement levels reported by the WHO.²⁴ The novel wearable MAIJU accurately reflects the natural continuum of maturing gross motor skills and supports constructing individual-level growth charts of gross motor development.^21,22 All of these individual-level data could be used for an objective and effective clinical support in developmental pediatrics, to benchmark therapeutic interventions or other counselling,³¹ to advance health care equity,³² and to support a wide range of developmental studies across disciplines.^3,5–7 

Prior wearable studies on infants’ motor development have only quantified the total amount of gross movements^33,34 or selected postures.^23,35 The present MAIJU method introduces a clear conceptual advance: Here, all gross motor assessments are based on the fully automated and validated algorithmic detections of the infants’ canonical postures and movements.^21,22,25 Such a “bottom-up” approach renders the results fully transparent and explainable, even at the individual level. Meanwhile, the practical ease of the multisensor wearable measurements makes them widely scalable across different health care, societal, and research settings.

Wearable measurements combined to intuitively understandable and transparent analytics hold promise for overcoming many daily challenges in the early neurodevelopmental assessments, such as infants’ stress-related behavior, denial to show skills in a new environment at an outpatient clinic, or inappropriate time schedule overlapping with the infant’s nap or mealtime. The reliability and accuracy of parental information is often unknown,¹⁴ whereas the limited hospital or laboratory observations are always ecologically deficient with respect to infants’ natural behavior at home.³⁶ The wearable methods may provide complementary tools by supporting fully unsupervised, at-home measurements that can last long enough to reliably capture children’s typical gross motor abilities during their natural activity (here referred to as “free playtime”).

Introducing novel electronic health technologies to clinical use calls for a careful consideration of the aimed use case and the biological and medical validity of the newly proposed measures.^32,37,38 We chose the GMMs as the primary output measure due to their worldwide recognition and routine use as proxies of motor development. The routine dichotomic reporting of reaching the milestones ignores the natural and concurrently evolving spectrum of all the acquired motor skills.^3,8,12,15 We devised 2 further measures: First, the time spent in each posture, which has wide implications in developmental^3,7,8,39 and clinical studies,⁹ because recent studies have indicated strong posture-dependence in the early language, social, and other perceptual development.^{6–8,16,40,41} Second, the holistic measure of motor maturation (BIMS; Figure 5B-C), which combines the situational variability in infants’ motor activity^12,15 in a way that mimics the intuitive experts’ assessment. Moreover, such measures can even support building motor growth charts with strikingly stable individual trajectories.

Our work has potential limitations. First, parental assessment of time spent in postures (scale 1–9) is unitless and nonnormalized across postures, which may deflate the reported correlations. Second, the infant cohorts used for the algorithm development and external validation came from distinct populations, but they may not represent all of the variability related to cohorts recruited elsewhere or with other criteria. Our fully automated and fixed analysis pipeline removes operator-related variability, and our experience to date from several ongoing studies in other countries and settings supports strong generalizability. Nevertheless, it is essential to validate the utility of these methods in different use cases and environments.

This is the first study, to our knowledge, to demonstrate that at-home wearable measurements provide objective, reliable, quantified, scalable, and ecologically valid assessments of infants’ gross motor skills. The measures presented are biologically and behaviorally reasoned, transparent, and intuitively explainable, and they hold promise for a wide range of health and research applications such as the following: (1) detecting developmental delays earlier among infants at risk for neurodevelopmental delays, (2) studying the elusive links between gross motor delay and delays in other developmental areas, (3) supporting comparative studies across different neurodevelopmental entities and disease levels,^42,43 (4) using as a clinical support system in well-child care, especially in ambiguous situations, (5) providing surrogate outcome measures to benchmark any observational and interventional studies, including global health projects,⁴⁴ (6) providing actionable, individual-level targets and efficacy metrics for early (physio-)therapy, (7) devising evidence-based and even globally applicable guidelines for the early developmental care,^1,13 (8) providing culture- and language-free infant assessments that directly facilitate international studies and very large-scale (big) data pooling, and, finally, (9) improving equity in infant health care by providing the same affordable electronic health solutions. With respect to developmental science in general, the ability to track an individual’s gross motor development with growth charts shows that the overall gross motor development follows predictable trajectories,^3,8,22,39 with precision that goes far beyond the conventionally assessed, discrete reaching of milestones.

BIMS

BABA infant motor score

GMM

gross motor milestone

ID

identity

K-S

Kolmogorov-Smirnov test

LME

linear mixed-effects model

MAIJU

Motor Ability Assessment on Infants With a JUmpsuit

WHO

World Health Organization