Introduction

A double burden of childhood malnutrition, undernutrition and obesity, is emerging in low- and middle-income countries (LMICs). In 2020, 23% of Southern African children under five years were stunted, 3% were wasted and 12% were overweight [1]. Despite this high burden of malnutrition, data on early growth trajectories are limited in LMICs.

Growth patterns established early in life may have long term consequences for health. Stunting, an indicator of chronic malnutrition [2,3] has been associated with infections in childhood and obesity in adulthood [4,5]. Childhood obesity is associated with development of several non-communicable diseases including cardiometabolic impairment, diabetes, and asthma [6,7]. Low birth weight, prevalent in LMICs, is associated with an increased risk of negative health consequences throughout childhood [8] and rapid catch-up growth from birth and 2 years may be associated with an increased risk of obesity.

Thus, defining growth patterns during childhood is key to identifying future health risk and vulnerable periods for potential interventions. The aim of this study was to identify growth profiles in children from birth to five years in a South African birth cohort, the Drakenstein Child Health Study (DCHS) using a Latent Class Mixed Modelling (LCMM) approach [9].

Additionally, the use of multivariate LCMM to identify latent growth classes based on joint height and weight measurements is proposed as this approach will not impose a restrictive, predefined relationship between height and weight as the use of zBMI or zWFH would when identifying distinct growth trajectories. No publications have identified latent classes using a multivariate response of height and weight or the comparison of latent growth class allocation across various growth responses.

Methods

Growth profiles from birth through 5 years were derived for children in the DCHS a South African birth cohort. Pregnant women in Paarl, a poor peri-urban area in South Africa, were enrolled during their second trimester of pregnancy, from March 2012 - March 2015. Mothers were recruited at an antenatal clinic, and hence pregnancy was already established by the primary healthcare providers. Mothers were recruited at an antenatal clinic, and all who were over the age of 18 and intended to reside in the Paarl area for at least one year were eligible for this study. Gestational age was calculated using second trimester ultrasound. If this was unavailable, last menstrual period was used to calculate gestational age at the time of enrolment. The most common mode of delivery was natural birth, while 20.12% of the children within the study were delivered by caesarean-section. Mothers provided informed consent at enrolment and were re-consented annually; mother-child pairs were followed from birth through childhood as previously described [10]. Comprehensive growth measurements were taken by trained study staff at birth 6, 10, 14 weeks, 6, 9 and 12 months and then 6 monthly until 5 years (Fig 1 and S1 Fig). The most rapid period of growth through childhood is during infancy, specifically throughout the first 2 years, hence measurements were captured more frequently through this period.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. The DCHS Cohort of children from birth until 5 years of age.

https://doi.org/10.1371/journal.pone.0319237.g001

This paper focuses on a model-based approach to identify latent growth profiles for standardized height and weight measurements, as well as the composite measures of these responses, namely standardised Body Mass Index (zBMI), standardised weight-for-height (zWFH) and a multivariate response of zHeight and zWeight (zHeight + zWeight).

Results

Amongst 1137 mothers enrolled there were 1143 children live born, with 981 (85.8%) followed through 5 years of age with detailed measurements as shown in Fig 1. Twenty-three (2%) children died in infancy and 139 (12.2%) children were lost to follow up, the majority of which occurred before 6 months of age. All growth measurements were recorded at each visit, however some measurements were excluded due to instrumental or measurement error.

Study cohort

Six-hundred and ninety-five (60.8%) of the mothers did not complete secondary schooling education, while 988 (86.4%) were members of households where the average monthly income was below R5000 (approximately USD276). One hundred and ninety-one (16.7%) children were born prematurely, the majority late preterm (born between 34-37 weeks gestational age). There were 247 (21.6%) HIV-exposed but uninfected children, with only 2 HIV-infected. Characteristics of included children were similar to those excluded; S2 Table summarises the total, included and excluded cohort with respect to maternal, child and socio-economic characteristics.

Average trajectory of growth responses.

Fig 2 illustrates the standardised growth responses from birth until five years. As these observations represent standardized growth scores, the average trajectory should follow the line of y = 0. On average, lower zHeight and zWeight scores from what were anticipated according to WHO references were observed. Larger than expected zBMI and zWFH scores were observed, indicating that while children were below expected weight, their height deviated even further below expectation, resulting in above expected zBMI and zWFH. The trajectories of the unstandardised, observed, growth responses are illustrated within S3 Fig and S1 Table in the supplementary materials.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Standardized growth measurements from Birth until age 5 with a smoothed average trajectory.

https://doi.org/10.1371/journal.pone.0319237.g002

Growth trajectories

Univariate trajectories.

Each response, described using the broken stick model previously indicated, fit with the latent class mixed model given the optimal number of k is illustrated in Fig 3, thus illustrating the trajectories of each latent growth class identified within the respective growth responses. The model estimates from the fixed effect components can be found within S4–S9 Tables in the Supplementary Materials.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The average trajectories of latent classes identified within A) zHeight, B) zWeight, C) zWFH and D) zBMI from birth until age of five.

https://doi.org/10.1371/journal.pone.0319237.g003

zHeight.

The four latent classes for zHeight are illustrated in Fig 3 (A). The following features were observed: Firstly, zHeight profiles below the expected are observed (this was also reflected in zWeight profiles). Secondly, much of the differences in structure of trajectories occur before 1.5 years, after which the four trajectories settle at different levels. This change in profiles prior to and after 1.5 years is reflected in the early changes in slope and later levels naming of the four identified classes (for example, sharp increase to expected levels). Two classes that settled slightly closer to the expected height are seen, one class of 321 (labelled “Gradual Increase, Gradual Decrease to Below Expected”) children who started at expected height, increased to slightly higher heights than expected and then settled at a level of less than one standard deviation below the expected level. Another class (labelled “Sharp Increase, Sharp Decrease to Below Expected”) includes 187 children who started with lower heights than expected but by the age of 1 year have caught up to a level less than one standard deviation below the expected height. A further two classes show children who were taller than expected at birth, but who then attained a level up to two standard deviations below the expected height, one group of 516 children (labelled “Sharp Decrease to Low”) earlier than the other group of 112 children (labelled “Gradual Decrease to Low”).

z Weight .

Four latent classes were identified within zWeight. Similar to zHeight, the heterogeneity in structure is seen in the profiles prior to 1.5 years. Two classes (labelled “Gradual Increase, Gradual Decrease to Expected” and “Sharp Decrease, Sharp Increase to Expected”) initialize close to the expected zWeight at birth and diverge within the first 1.5 years, one class slightly above expected zWeight and another dipping well below expected, after which they meet and show subjects that follow expected weight for age from 1.5 to 5 years old (Fig 3 (B)). Two classes with birthweights as expected who then deviated from expected weight profiles were observed, resulting in a class with above expected weights (labelled “Sharp Increase to High”) and a class with lower-than-expected weights (labelled “Gradual Decrease to Low”).

z WFH.

Three classes were identified for zWFH (Fig 3 (C)). A class that almost follows the expected weight-for-height from birth until five years (labelled “Gradual Decrease to Expected”), and two classes with consistently greater weights given their height and sex. The group with the greatest weight for height (labelled “Sharp Increase to High”), showed the smallest weight-for-height at birth.

zBMI.

Three classes were identified for zBMI, all of which were characterised by early increasing zBMI measurements (Fig 3 (D)). A class that closely follows the expected zBMI from birth until five years, except for a brief fluctuation above the expected BMI before 6 months of age (labelled “Sharp Increase, Sharp Decrease to Expected”). Two classes with smaller than expected zBMI at birth which both settle above the expected zBMI. One class settles earlier on, at 3 months of age (labelled “Sharp Increase to High”) while the second settles around one year of age (labelled “Gradual Increase to High”).

Fig 3 interestingly also show that those with the most extreme lower than expected measurements at or soon after birth, settled at the highest levels, and vice versa, indicating an over-correction of sorts.

Multivariate trajectories.

Fig 4 illustrates the zHeight and zWeight profiles for the five classes determined for the joint zHeight + zWeight measurements. Three hundred and eighteen children were allocated to a class with low zHeight and zWeight from birth until five years (labelled “i”). Both zHeight and zWeight was above expected at birth but slowly decreased to below expected at birth for 205 children (labelled “ii”). A small group of children (n = 75) showed a delay in both zHeight and zWeight in the first few months of growth, however this did not appear to have long lasting impacts on zHeight or zWeight trajectories (labelled “iii”). A class of 296 children illustrate weight and height measurements equal to what was expected prior to age 1 however in subsequent observations these children were shorter than expected given their age and sex, while maintaining an expected weight (labelled “iv”). The last identified class shows individuals with a low zWeight at birth which steadily increased within their first year and remained above what was expected (labelled “v”). This same group was also shorter than expected at birth, with a slight delay in height growth during the first year of life, after which they followed expected height for age and sex.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. The average trajectories of four classes identified within zHeight

 + zWeight from birth until age of five, illustrated using A) zHeight and B) zWeight.

https://doi.org/10.1371/journal.pone.0319237.g004

The accordance or dissimilarity of zHeight and zWeight trajectories within the latent classes is of interest. Class (iv) shows different profiles in particular before 1.5 years where an increasing zWeight and decreasing zHeight can be seen. Additionally, class (v) also shows an increasing zWeight with decreasing zHeight within the first few months of life.

To determine whether the choice of knot location impacted the subsequently identified latent profiles, the latent class identification process was repeated with additional knots placed at later time points. The extended knot locations used were (0.25,0.75,1,1.5,2,3,4) years for all responses considered. The results from this longitudinal specification are presented in the supplementary materials (S12–S17 Figs). The results did not differ, and thus the knot locations specified in S3 Table were used.

Comparison of latent class allocations

The correspondence between latent class allocations for individual children based on different growth measurements and on allocations based on univariate versus multivariate models are shown in Tables 1 and S10–S12. Table 1 illustrates the correspondence between allocations based on zHeight and zWeight individually and zHeight and zWeight within a joint model in a three-way contingency table. S10 and S11 Tables summarise the agreement between composite measures of height and weight (zBMI and zWFH) and the class allocations based on models for joint zWeight and zHeight. Finally, S12 Table summarises the agreement between the latent classes identified given zWFH or zBMI.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Tabulation of multivariate group allocation versus.

https://doi.org/10.1371/journal.pone.0319237.t001

Association with abnormal growth

To classify Rapid Weight Gainers (RWG), children were identified who experienced an increase in weight between birth and 9 months that was greater than 0.67 standardized units. The proportion of RWG children within each identified growth trajectory for weight is shown in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Proportion of children that experienced rapid weight gain within the first 9 months of life within the four zWeight growth classes.

https://doi.org/10.1371/journal.pone.0319237.t002

Most children identified as RWG were allocated to the “Gradual Increase; Gradual Decrease to Expected” and “Sharp Increase to High” zWeight classes, while very few were allocated to the “Gradual Decrease to Low” and “Sharp Decrease, Sharp Increase to Expected” classes (Table 2). 93% of children allocated to the “Sharp Increase to High” class experienced RWG, showing an association with RWG and the high zWeight trajectory. Within the “Gradual Increase; Gradual Decrease to Expected” class, 56% experienced RWG. Thus, this class is almost equally represented by children who experienced RWG, and those who did not, who have settled at a normal weight.

The proportion of those that experienced RWG within other growth trajectories is illustrated within the supplementary materials (S13–S15 Tables). Within zBMI, both “Increase to High” trajectories comprised over 50% RWG individuals, this is also seen within the zWFH “Sharp Increase to High” trajectory. Otherwise, the zBMI and zWFH classes do not show a clear association with RWG, which is to be expected as zBMI and zWFH measures consider both zHeight and zWeight profiles. Within the multivariate zHeight + zWeight classes a large percentage of RWG within those allocated to the multivariate classes (iv) and (v) (61% and 81% respectively) occurred, in line with results seen in zWeight.

The number of children that ever experienced some form of abnormal growth, namely stunting (zHeight Score < -2), overweight (zWeight Score >  2) or underweight (zWeight Score < -2), at one or more time points between birth and 5 years were identified (Tables 3–4). Both “Decrease to Low” (Sharp and Gradual) classes, based on zHeight, show over 50% of individuals classified as stunted at one or more time points (Table 3). Within the “Gradual Increase, Gradual Decrease to Below Expected” class 34% of subjects were classified as stunted at one or more time points, and within the “Sharp Decrease, Sharp Increase to Below Expected” class the largest percentage of 73% of individuals classified as stunted at one or more time points. The largest percentage of children identified as underweight at one or more time points, 50%, were allocated to the “Sharp Decrease, Sharp Increase to Expected” class when considering zWeight (Table 4). Meanwhile, the largest percentage of children identified as overweight at one or more time points, 41%, were allocated to the “Sharp Increase to High” class when considering zWeight (Table 4). These identified latent profiles thus indicate children with abnormal growth features across the 5-year period rather than at isolated ages.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Proportion of ever stunted children between birth and 5 years as defined by zHeight scores.

https://doi.org/10.1371/journal.pone.0319237.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Proportion of children ever underweight or overweight between birth and 5 years as defined by zWeight scores.

https://doi.org/10.1371/journal.pone.0319237.t004

Discussion

The structure and heterogeneity of growth profiles for children in a LMIC African birth cohort characterised using LCMM identified distinct growth classes from birth through 5 years. Through these latent classes children were identified who deviated from expected growth patterns and their progression through early childhood was tracked. We identified a clear association between univariate zWeight and zHeight classes and zHeight + zWeight classes and abnormal growth patterns of stunting, underweight, overweight and rapid weight gain.

This is the first study to identify latent growth classes using a joint growth response made up of zHeight and zWeight during childhood. Additionally, the trajectories identified using this joint response were contrasted against those identified using zBMI, zWeight, zHeight and zWFH. The results were internally validated using two specifications of the longitudinal process, through the use of standardised and unstandardised growth responses as well as through repeating the LCMM using internal cross-validation using randomly sampled subsets. Consequently, there is confidence in the latent profiles identified within this report.

Growth responses may be analysed within their observed or standardised formats. The observed, unstandardised approach does not control for prematurity, age or gender. Thus, a key advantage of using standardised growth responses is the generalisability of trajectories, as the impact of gender or gestational age on growth has been removed from the process of latent class identification.

A Cluster Analysis approach could have been used to identify latent groups with respect to growth instead of the LCMM approach. This would have required the definition and analysis of features of a given growth profiles as opposed to the longitudinal growth measurements. Gough et al., used K-means clustering approaches to identify latent trajectories within standardised height (zHeight) scores and identified four classes within this growth response [15]. Similarly, Mebrahtu et al., identified three classes within standardised weight (zWeight) scores using the GMM approach [16]. Here, within zHeight and zWeight four classes were identified, while zHeight + zWeight lead to the identification of five latent profiles. The classes identified using the multivariate approach were not identical to those found using the univariate approach, and additionally allocation to these classes was not identical. This shows the benefit of considering both the uni- and multivariate approaches when identifying latent growth trajectories. The multivariate approach considers both the change of zHeight and zWeight responses over time that are adjusted for age and sex, with an unrestricted relationship unlike zBMI and zWFH. This also reiterates that the multivariate approach allows the identification of latent trajectories that consider the change of both height and weight independently over time. The univariate and multivariate approaches consistently identified a low zHeight class, which may serve as a stunted, or at high risk of stunting class, a low zWeight class (those at greater risk of underweight) and a high zWeight class (those at greater risk of overweight).

Three growth classes were identified within zBMI and zWFH scores between birth and five years. In contrast, Rickman et al. identified four latent classes within zWFH, however they considered zWFH from birth until two years within a Western Kenyan cohort. While they identified an additional class, there is strong agreement between the trajectories identified by Rickman et al. and those within this study; the difference being with an additional low zWFH class identified by Rickman et al. [17]. It is possible that cohort differences may explain the difference in profile numbers identified here as Rickman et al., observed an increased risk of suboptimal growth - considering 8 distinct growth measures - within HIV-exposed uninfected children.

Stein et al., (2010) compared the patterns of height in early childhood (birth to age 5) in five birth cohorts from distinct populations around the globe (Brazil, Guatemala, India, Philippines and South Africa) [18]. Considering the same cohorts, Poveda Rey et al., (2021) looked at weight throughout early childhood [19]. These studies illustrated diversity in growth profiles across multiple LMIC nations. The Brazilian cohort mostly followed expected standardised height and expected standardised weight within the first five years of life. While the Guatemalan, South African and Indian cohorts were born smaller (with respect to length) than the expected; the South African cohort recovered by the age of five while the Guatemalan and Indian cohorts did not. The Guatemalan, Filipino and South African cohorts were born weighing slightly less than expected. The South African cohort did not recover, nor did they show further reduction in expected weight. However, the Guatemalan and Filipino cohorts did show further reduction in standardised weight scores, to a greater extent for the Filipino cohort. The Filipino cohort was born with an expected height but decelerated to much smaller than expected by one year of age. The Indian cohort was born with the smallest standardised weight which did not improve through early childhood. Interestingly, all cohorts showed a decreased standardised height at two years compared to birth; while some were able to recover from this deceleration (Brazil, South Africa), other cohorts did not (Guatemala, India, Philippines).

Previous studies that make use of latent class analysis to identify groups of children with similar growth trajectories primarily focus on BMI as an indicator of early onset obesity [20,21]. Robinson et al., reviewed all publications post 2000 that investigated latent class identification within BMI responses [21]. Of the eight studies found, six identified four clear trajectories within BMI during childhood [21]. Accordingly, identification of latent classes within BMI or zBMI is seen more frequently in the literature than zWFH. Most commonly, three or four latent classes are identified within zBMI or BMI, however these studies often focus on growth from the age of four onwards [21–24]. Wang et al., found dynamic BMI growth patterns (such as BMI catch-up and stable overweight) were more predictive, and thus more informative, than static BMI measurements of cardiovascular structure and function in early childhood [25]. Rapid weight gain in early childhood (or catch-up growth) has been associated with an increased risk of overweight, obesity and other chronic diseases during later childhood and adulthood [26–29]. Additionally, catch-up growth has been associated with various factors such as altered insulin metabolism [30,31] and hence an increased risk of Type II diabetes [32], increased systolic blood pressure [33], increased risk of coronary heart disease [34] and an increased risk of childhood asthma [35] Various factors may modify the risk of rapid weight gain or obesity, including bottle feeding, shorter gestation age and being firstborn [36,37].

Conventionally weight-for-height/length (zWFH) is used to assess over/under-weight status in children below the age of 2 years, while Body Mass Index (BMI) is used from 2 years onwards [38]. Weight-for-height and standardised BMI (zBMI) scores are calculated using WHO charts; both represent the relationship between height and weight. However, zBMI scores are standardised by age while zWFH scores are not [11]. Instead, these are created using charts that represent the expected relationship between height and weight between birth and two years or two years and five years of age [39]. Aris et al., found that the use of zBMI or zWFH before 2 years did not impact the ability to predict future adiposity or cardiometabolic outcomes in children, suggesting these scores may be equally beneficial.

A clear limitation for the use of zWFH is the lack of age-dependent standardisation, which is resolved when one makes use of zBMI [11,39]. However, a shortcoming of the use of BMI to study future risk of obesity, is the fact that BMI does not consistently reflect body composition [40]. It has even been suggested that BMI and BMI increase during early childhood is more predictive of lean mass than adiposity during adulthood [41]. Body composition, fat mass and fat-free mass may be measured, and subsequent latent classes have been identified within the observations of fat mass and fat-free mass independently [42]. However, further research is needed to identify whether this may be a better predictor of obesity in children. Additionally, this requires dedicated machinery that may not be available at many care facilities in Low to Middle Income (LMIC) countries. Here, the trajectories identified within zBMI and zWFH as well as allocation to these classes across these two responses were in some agreement, indicating that there may not be a clear benefit to using zWFH over zBMI. Similarly, Aris et al., considered the association of overweight based on zBMI or zWFH and cardiometabolic risk during the first two years of life and found no clear difference using zBMI or zWFH [39].

Both groups of latent trajectories identified using zBMI and zWFH did not show a strong agreement in allocations when contrasted with multivariate zHeight + zWeight. This illustrates additional information the multivariate zHeight + zWeight approach may add to such analyses. The multivariate zHeight + zWeight approach was able to identify abnormal growth with respect to zHeight, zWeight or both while the zWFH and zBMI approaches focus on zWeight with zHeight as a reference regardless of whether zHeight is classified as normal or abnormal. Thus, the use of multivariate zHeight + zWeight has provided a different perspective to our understanding of growth during childhood.

Remarkably, 31% of overweight children under five live in LMICs, while 72% and 59% of wasted and stunted children under five, respectively, live in LMICs. While the prevalence of stunting and wasting in children is decreasing in Southern Africa, the prevalence of obesity is increasing [1]. Fifty percent of subjects within this cohort were identified as stunted at one or more timepoints, in particular the “Sharp Decrease, Sharp Increase to Below Expected” zHeight class. Forty one percent of those allocated to the “Sharp Increase to High” zWeight class were classified as overweight at one or more timepoints. Half of those allocated to the “Sharp Decrease, Sharp Increase to Expected” zWeight class were identified as underweight at one or more timepoints. This illustrates the ability of using identified latent classes as a proxy for those at high risk of stunting, overweight or wasting in a longitudinal setting.

The “Sharp Increase to High” zWeight class identified 93% of RWG children, who are at risk of obesity, and development of non-communicable diseases. Those that experienced RWG make up the majority of children identified within the BMI and WFH “Sharp Increase to High” classes, indicating a higher risk of RWG related adverse effects within these classes. Interestingly the RWG children were allocated to similar classes when considering either zHeight + zWeight or zWeight. Within zHeight + zWeight class (v) 81% of experienced RWG, and hence subjects allocated to this class are at greater risk of RWG associated adverse effects.

Limitations include the use of a single cohort. However, this cohort comprised of almost 1000 children with numerous repeated growth measurements and high cohort retention. Furthermore, this cohort is representative of LMIC child populations with high rates of poverty and infectious diseases, which is a consistently underrepresented population group within latent growth analysis. observations were acquired more frequently during the first year but this is the period of most accelerated growth [43], with growth patterns setting a developmental trajectory for life [25,44]. This may have impacted the shapes of growth curves identified as there was more information to describe these curves during the first year but use of a greater degree of smoothing at early ages may account for this.

To describe the longitudinal growth responses a broken-stick model was used. While more flexible growth models such as a cubic-spline, polynomial or more conventional growth models such as the SITAR could have been used none of these approaches would produce as interpretable results as identified here. The broken-stick model allows direct comparison of changes in growth between respective intervals across latent classes.

Future work will include investigating the association between growth trajectories and childhood or future illness, hence identifying possible areas of intervention to promote optimal growth in children. Here the understanding of how weight and height may change throughout childhood has been broadened. With the identification of these distinct growth trajectories, one may be able to explore whether these specific profiles may serve as an indicator of early onset illness.

Supporting information

Acknowledgments

The authors thank the study and clinical staff at Paarl Hospital, Mbekweni and TC Newman clinics, as well as the CEO of Paarl Hospital, Dr. Kruger, and the Western Cape Health Department for their support of the study. The authors thank the families and children who participated in this study.