BACKGROUND: Longitudinal growth curves, based on repeated measurements from the same group of infants, exist for preterm infant weight and length but not for BMI. Our existing BMI (weight divided by length squared) curves are based on cross-sectional birth data obtained from a different group of infants at each gestational age (GA).
METHODS: We calculated BMI over time for 68 693 preterm infants between 24 and 36 weeks GA. Stratifying infants by sex, GA at birth, and quintiles based on birth BMI, we created longitudinal median curves using R and validated the resulting curves for empirical fit, proper classification, and normality of z scores.
RESULTS: We created 2 sets of BMI growth charts. The first set displays fitted median curves for all 5 percentile groups in each GA group by sex. The second set displays fitted median curves with their corresponding third and 97th percentiles by percentile group, GA, and sex. In the validation analysis, percentage of daily observations below the median curve approximated the expected 50th percentile after the initial 3 days. Unlike the cross-sectional curves, the longitudinal curves reveal the pattern of change corresponding to nadir; postnadir, these curves remained consistently below the cross-sectional curves and varied by GA and sex as expected. The percentage of observations falling below the 50th percentile for cross-sectional curves (revealing optimal growth) was generally much higher than for longitudinal curves (revealing actual growth).
CONCLUSIONS: These new longitudinal curves provide clinicians data on how premature infants’ body proportionality changes over time.
- GA —
- gestational age
- GAMLSS —
- Generalized Additive Models for Location, Scale, and Shape
What’s Known on This Subject:
In preterm infants, the rate of weight growth is faster than length growth; thus, weight is often disproportionate to length at NICU discharge. To our knowledge, no longitudinal curves quantifying body proportionality have been published for preterm infants.
What This Study Adds:
We provide longitudinal curves to illustrate how body proportionality changes with age in prematurely born infants.
Preterm infants in the NICU have greater risk for abnormal growth relative to healthy term infants. Growth restriction, common in preterm infants, and excess body mass can both result in subsequent health consequences.1–3 Thus, accurately tracking preterm growth is important.4
Preterm infant growth is assessed by measures of weight, length, and head circumference, most commonly weight.5 Emphasis on individual growth measures does not take into consideration the potential for disproportionate growth occurring when an infant’s weight differs from that expected for a given length. Such disproportionate growth may require adjustment in nutritional care.
There is disagreement regarding the ideal measure for determining the weight and length relationship in preterm infants.6 Our research group compared 6 weight-for-length ratios in preterm infants and found BMI to be the most appropriate.7 Authors of recent studies of body composition using dual energy x-ray absorptiometry found that no proportionality index was a truly accurate measure of fat mass or fat-free mass regardless of gestational age (GA) at birth, but BMI was most correlated with fat-free mass.8 Thus, we continue to focus on BMI measures.
There are 2 approaches for tracking growth after birth in preterm infants. One is to compare observed measurements with cross-sectional curves (also called intrauterine growth curves) created by using birth growth parameters of a different group of infants at each GA. When a preterm infant’s growth is tracked by using this type of curve, it is analogous to comparing postnatal growth to intrauterine growth as recommended by the American Academy of Pediatrics.9 A second approach is to compare observed measurements with longitudinal growth curves (also called postnatal or extrauterine growth curves) on the basis of measurement of the same group repeatedly over time. These 2 types of curves allow us to evaluate an infant’s postnatal growth in the NICU compared with optimal growth by using cross-sectional intrauterine curves and actual growth of infants of the same GA by using longitudinal postnatal curves.10
Our research group previously published cross-sectional BMI-for-age preterm growth curves,7 but existing longitudinal curves are used to track only weight and height independently5,11,12; to our knowledge, there are no preterm longitudinal growth curves for BMI. Thus, our purpose with this study was to evaluate longitudinal changes in BMI in a large, racially diverse sample of preterm infants in US NICUs.
The data consist of daily observations for 189 782 preterm infants 24 to 36 weeks GA (a neonatologist estimated this using obstetric history, obstetric examinations, prenatal ultrasound, and postnatal physical examinations) who were admitted to a NICU on the day of birth. The data were collected by Pediatrix Medical Group, Inc from infants born in 1 of 248 hospitals within 33 US states between 2009 and 2013. After implementing the exclusionary criteria, 68 693 infants, representing 36% of the total cohort, were included in the study (see Fig 1 for details). Each infant contributed 1 to 61 daily observations. The sex and race breakdown of our participants was similar to data from 2012 for infants admitted to a representative sample of NICUs in the United States.13 Because our data set was deidentified, the study was considered exempt from review by the Institutional Review Board at Kennesaw State University.
Median Curve Creation
Weight was measured to the nearest gram by using an electronic scale of unknown type; length was measured to the nearest centimeter by using a tape measure or length board. We calculated gestational day–specific z scores by sex using our cross-sectional weight and length curves.10 On the basis of these z scores, we excluded daily observations that were outliers (|z| >5) for either weight or length (0.06% of observations).
Weight was measured daily, and length was measured approximately weekly, so we interpolated missing lengths using linear interpolation, linear regression, and last observation carried forward. On the basis of visual inspection of the resulting curves, we found that linear interpolation was best, which is consistent with the finding of Bishop et al14 that postnatal length growth is near linear during the first week in appropriately sized infants. We then calculated day-specific average z scores and excluded outliers with a z score ±3.5 from this average for weight, length, and BMI. If either weight or length was excluded as an outlier, the corresponding BMI was also excluded (0.31% of observations).
On the basis of age categories defined by the World Health Organization15 for extremely preterm (24–27 weeks GA), very preterm (28–31 weeks GA), and moderate-to-late preterm (32–36 weeks GA) infants, we stratified the remaining data by sex and GA. Using random selection, we split the data into creation (75%) and validation (25%) sets. We further stratified the data into quintiles on the basis of birth BMI and created median curves for each quintile.
We used the Generalized Additive Models for Location, Scale, and Shape (GAMLSS) package16,17 in R to create sex- and GA-specific BMI growth curves. The lms (lambda mu sigma) function in GAMLSS was used to determine the appropriate distribution of the daily BMI data on the basis of deviance score. The final set of fitted median curves used the Box–Cox Cole and Green distribution.18 The smoothing of the fitted median curves used penalized box splines with degrees of freedom obtained from the LMS procedure. We created the corresponding third and 97th percentile curves for each quintile using this same procedure.
We validated results using several methods, including examining normality, assessing goodness of fit, and determining prediction performance. To examine normality, we calculated z scores using Box–Cox Cole and Green parameter estimates from the LMS procedure. We plotted mean, SD, skewness, and kurtosis for z scores for each day. We assessed goodness of fit using 2 approaches. First, we compared the original fitted median curves with the unsmoothed, empirical median curves created by using the validation data set and examined alignment between the curves. Second, we calculated the daily percentage of validation data observations that fell below the fitted median curves for every quintile in each GA group and examined plots to determine if this percentage aligned with the expected 50th percentile. To determine prediction performance, we compared individual growth curves for a subset of 9 randomly chosen infants from the validation data set with the original fitted median curves and their corresponding third and 97th percentile curves to determine if these individuals’ growth trajectories followed the expected growth patterns.
Comparison of Longitudinal and Cross-sectional BMI Curves
We compared our longitudinal BMI curves with our previous cross-sectional BMI curves7 using 2 approaches. First, we created a plot with longitudinal curves for the 3 GA groups along with the cross-sectional median curves. We used the longitudinal median curves for the middle quintile group for each GA group. Because we grouped infants into 3 groups according to GA, the longitudinal curves we created are based on days since birth, whereas the cross-sectional curves are based on postmenstrual age, so a direct comparison is not possible. Thus, to determine the beginning position of the longitudinal curves, we determined where BMI on the day of birth on the longitudinal curves for each GA group coincided with the cross-sectional median curve. Second, we compared the daily percentage of validation data observations falling below the cross-sectional median curves with the daily percentage of validation data observations that fell below the longitudinal median curves.
Average length of stay decreased with increasing GA, and overall number of daily observations decreased as infants were discharged from the NICU. Fewer observations decreased the reliability of the curves, so we terminated the curves at 60 days (24–27 weeks GA), 45 days (28–31 weeks GA), and 30 days (32–36 weeks GA). See Table 1 for detailed descriptive statistics. We considered each BMI value to be independent, which is reasonable given that our purpose was constructing median curves rather than testing a hypothesis or creating confidence intervals.
We created 2 sets of BMI growth charts to evaluate postnatal changes in BMI. The first set of charts reveals the fitted median curves for all 5 percentile groups in each GA group by sex. Figure 2 reveals an example of these fitted median curves for girls at 24 to 27 weeks GA. (Fitted median curves for both sexes and all GA groups are available in Supplemental Fig 8; corresponding median BMIs per day are available in Supplemental Tables 3–5.) The second set of charts reveals fitted third and 97th percentile curves along with median curves for each percentile group in each GA group by sex. Figure 3 reveals an example of these curves for girls at 24 to 27 weeks GA for percentile group 1. (Additional curves for both sexes and all GA and percentile groups are available in Supplemental Figs 9–11; corresponding BMIs per day for third, 50th, and 97th percentiles are available in Supplemental Tables 6–11.)
Both GA at birth and postnatal age influenced changes in BMI. The first 7 to 14 days after birth had an initial drop followed by a near-linear increase in BMI for all GA groups and both sexes. This early drop in BMI was driven by weight loss. The extremely preterm infants showed a more rapid increase in BMI over time than the more mature infants. A comparison within GA but across the 5 percentile groups revealed that the top percentile group showed the greatest weight loss during nadir (weight loss immediately after birth), but after the initial period of nadir, the curves for all percentile groups within GA and both sexes followed a similar growth pattern.
Plots of the z scores based on the validation data set revealed no significant deviations from normality (mean = 0; SD = 1; skewness = 0; kurtosis = 3) except during the first 3 days after birth.
Goodness of Fit
A comparison of the original smoothed, fitted median curves with the unsmoothed, empirical median curves created based on the validation data set revealed good alignment overall, indicating that the curves fit well across all quintiles and GA groups for both sexes. Figure 4 reveals this comparison for girls 24 to 27 weeks GA. Fit was better for older GA groups, which can be seen in the comparison curves for both sexes and all GAs in Supplemental Fig 12. As an additional assessment of goodness of fit, we calculated the percentage of daily BMI observations in the validation data set that fell below the fitted median curves by group for each day. An example graph is shown in Fig 5 for girls 24 to 27 weeks GA. (Graphs for both sexes and all GA groups are available in Supplemental Fig 13.) The percentage below the curve approximated the expected 50th percentile after the initial 3 days. Overall, the percentage of daily observations that fell below the median curves ranged from 48% to 52% (mean absolute deviation from 50% is 0.8%). Table 2 includes detailed data for each group.
Most infants from the validation data set followed the predicted growth patterns, and their growth trajectories were largely contained within the third and 97th percentiles of the longitudinal curves. Figure 6 reveals plots of 9 randomly selected infants. The growth trajectories of 7 of these infants followed the predicted growth patterns. Only 2 infants did not follow the expected pattern: patient 1 consistently hovered around the third percentile, and patient 5 spiked above the 97th percentile during a portion of the first 10 days.
Comparison of Longitudinal and Cross-sectional BMI Curves
Figure 7 reveals the longitudinal curves for the 3 GA groups compared with the Olsen et al7 cross-sectional median curves. As seen in this figure, the longitudinal curves revealed an initial drop for nadir that did not appear in the cross-sectional curves. After nadir, the longitudinal curves remained consistently below the cross-sectional curves. For both sexes, the extremely preterm infants (24–27 weeks GA) had a more rapid increase in BMI compared with the cross-sectional curves. The infants in this group returned to the optimal growth level over their stay in the NICU. Curves for the preterm infants (32–36 weeks GA) are parallel to the cross-sectional curves, revealing a similar growth trajectory. Rate of growth for the very preterm group (28–31 weeks GA) varied by sex. For girls, the 2 curves illustrated a similar rate of growth, whereas for boys, growth shown by the longitudinal curves increased faster than in the cross-sectional curves.
Comparison of the percentage of validation data observations that fell below the longitudinal and cross-sectional curves revealed differences in optimal versus actual growth. As discussed above, the percentage of data observations below the longitudinal curves approximated the 50th percentile after the first 3 days. For the cross-sectional median curves (see Fig 5 for the extremely preterm group of girls; all other comparison graphs are available in Supplemental Fig 13), the percentage of daily observations below the curve was high initially because of nadir. Although these percentages moved closer to 50% over time, the percentages were generally much higher than for the longitudinal curves. For the extremely preterm infants, the percentage of daily observations that fell below the 50th percentile decreased over time and eventually reached that range. For the very preterm infants, the percentage also decreased but remained consistently above the 50% level. For the preterm infants, the decline was minimal during the length of the study. The pattern of the longitudinal and cross-sectional curves did not differ by sex.
Growth curves that are currently available to assess body proportionality in preterm infants are based on cross-sectional BMI data obtained on the date of birth or on longitudinal data of weight and length considered independently. We created and evaluated longitudinal BMI curves for preterm infants specific for sex, GA, and birth BMI quintile using data from a large, contemporary sample of infants in NICUs within the United States. These curves apply only to preterm infants admitted to the NICU, because those infants born between 35 and 36 weeks GA and not admitted to the NICU might have a different growth pattern.
Curves for all GAs, percentile groups, and both sexes revealed a postnadir linear increase in BMI over time, with the most premature infants showing the greatest increase. Within GA group, all percentile groups for both sexes revealed similar rates of growth after nadir. Unlike the cross-sectional BMI curves, these longitudinal curves are the first to reveal the weight loss that commonly occurs within the first few days after birth (see Fig 7).
Figure 2 (fitted median curves for all percentile groups), Fig 3 (fitted median, third, and 97th percentile curves for 1 percentile group), or their analogs found in the Supplemental Information can be used in clinical practice to evaluate an infant’s growth compared with his or her peers. Using information about an infant’s sex, GA, and birth BMI, one can determine the infant’s percentile group on the basis of the birth BMI ranges provided in the legend of Fig 2. To track this infant’s growth using Fig 2, a clinician can compare the infant’s calculated BMI with the curve for others with similar birth BMIs. Examining the infant’s BMI growth across time allows the clinician to determine how the infant is developing relative to all percentile groups specific to the sex and GA of the infant. For example, if the infant is placed in the second percentile group on the basis of birth BMI, her growth in the NICU can be tracked relative to other NICU infants of similar GA and size at birth, as well as to smaller or larger infants (in lower or higher percentiles). An infant’s growth can also be tracked by using Fig 3. Using the growth curve specific for an infant’s percentile group derived from Fig 2, the clinician can monitor changes in BMI compared with infants in the third, 50th, and 97th percentiles.
The best method for creating growth curves has been debated, with authors of recent studies using quantile regression19,20 or GAMLSS.16,21,22 We created curves using both approaches for comparison purposes. On the basis of this comparison, we opted to use GAMLSS because it produced better fit to the sample median than quantile regression.
Infants born prematurely grow differently from infants born at term. Understanding how BMI changes after birth as compared with fetal growth is important in assessing the quality of growth. Comparing what we observe in clinical practice by using longitudinally derived growth charts represents what we are currently able to achieve. As observed in our study, the actual growth represented in the longitudinal curves falls below the optimal growth illustrated by our cross-sectional intrauterine BMI curves. Cross-sectional intrauterine curves remain our best available tools for targeting optimal growth in preterm infants during their stay in the NICU. However, the longitudinal BMI curves presented in this study are appropriate for tracking an individual infant’s growth and provide additional useful information for clinicians tracking proportionality of growth in preterm infants in the NICU. It is our hope that together these curves will be used to identify and guide practice for infants growing disproportionately in weight relative to length, who currently are overlooked by growth assessment that is focused on size for age.
Potential limitations are that data for this study were collected during routine clinical care, which may introduce error. Inaccurate length measurements in newborns in clinical practice are well documented.23 However, given that overestimating and underestimating actual length are equally likely, this error would be random and unbiased.23,24 Additionally, data collected during routine care are often incomplete. Because of incomplete data and our goal to model the growth of healthy premature infants, we excluded 64% of the infants in our initial sample.
The longitudinal BMI curves created and validated here provide clinicians an additional tool for following growth and identifying disproportionate preterm growth. What actions to take will be clearer as these curves are tested in the clinical setting. Determining the impact of different growth patterns on short- and long-term outcomes will be essential, including how well these longitudinal curves predict outcomes for preterm infants and whether these provide more meaningful clinical information beyond longitudinal weight and length independently.
- Accepted June 13, 2018.
- Address correspondence to M. Louise Lawson, PhD, Department of Statistics and Analytical Sciences, Kennesaw State University, 275 Kennesaw State University Rd, MD 1103, Kennesaw, GA 30144. E-mail:
FINANCIAL DISCLOSURE: The authors have indicated they have no financial relationships relevant to this article to disclose.
FUNDING: No external funding.
POTENTIAL CONFLICT OF INTEREST: The authors have indicated they have no potential conflicts of interest to disclose.
- Cole TJ,
- Statnikov Y,
- Santhakumaran S,
- Pan H,
- Modi N; Neonatal Data Analysis Unit and the Preterm Growth Investigator Group
- American Academy of Pediatrics, Committee on Nutrition
- Ehrenkranz RA,
- Younes N,
- Lemons JA, et al
- Bishop NJ,
- King FJ,
- Lucas A
- March of Dimes
- Save the Children
- Paul IM,
- Schaefer EW,
- Miller JR, et al
- Stasinopoulos MD,
- Rigby RA,
- Heller GZ,
- Voudouris V,
- de Bastiani F
- World Health Organization
- Department of Nutrition for Health Development
- Copyright © 2018 by the American Academy of Pediatrics