Unconditional and conditional standards for fetal abdominal circumference and estimated fetal weight in an ethnic Chinese population: a birth cohort study

Singapore is a city-state in South-East Asia, with 74.1 % of its population being ethnic Chinese in 2010 [26]. The Growing Up in Singapore Towards healthy Outcomes (GUSTO) study is a mother-offspring cohort study. Details of the study have been reported previously [27]. Briefly, pregnant women aged 18 years or above who had their first trimester antenatal dating ultrasound scan at the maternal units of Singapore’s two major public hospitals (KK Women’s and Children’s Hospital and National University Hospital) between June 2009 and September 2010 were recruited. This study was approved by both the National Healthcare Group Domain Specific Review Board and the SingHealth Centralized Institutional Review Board. Written informed consent was obtained from each participating pregnant woman in early pregnancy.

Singletons, conceived naturally, whose mother and father were both ethnic Chinese were included in the present analyses. For the purpose of developing standards, we selected a “healthy” group of participants. Mothers with the following characteristics were excluded: pre-pregnancy diabetes, pre-pregnancy hypertension, previous miscarriage, previous still birth, smoking during pregnancy, alcoholic consumption during pregnancy, self-reported pre-pregnancy body mass index (BMI) <17 kg/m² or ≥27 kg/m², pre-eclampsia and/or pregnancy-induced hypertension, and/or diabetes diagnosed by oral glucose tolerance test between 26 and 28 weeks of gestation according to the WHO criterion [28] or fasting plasma glucose levels above 5.1 mmol according to the International Association of the Diabetes and Pregnancy Study Groups (IADPSG) criteria [29]. Furthermore, fetuses/neonates with the following characteristics were excluded: fetal abnormality detected on antenatal karyotype or ultrasound scans, delivery before 27.0 or later than 42.0 weeks, birth weight below 2.3 kg or larger than 4.5 kg, neonatal hypoglycaemia, fetal death, still birth, and/or neonatal death. We used the 2.3 kg cut-off instead of 2.5 kg because the WHO Multicentre Growth Reference Study Group [30] showed that the first percentile in the boys and girls in the healthy cohort was 2.3 kg. The choice of pre-pregnancy BMI cut-offs took Asian pattern of BMI and metabolic diseases into account [31].

In addition to their first trimester antenatal dating ultrasound scan, participants returned to the hospitals at 19–21, 26–28 and 32–34 weeks of gestation for ultrasound scans, among other antenatal and post-natal assessments. All the ultrasound measurements were measured based on the standard views used in Fetal Medicine Foundation [32]. Fetal head circumference was measured on the transventricular view, which was obtained at the level of the thalami, with visualisation of the falx cerebri and cavum septum pellucidum. Symmetrical appearance of both hemispheres was ensured in the measured section. Abdominal circumference was measured at the level of the stomach and where the umbilical vein was at the level of the portal sinus. For femur diaphysis length measurement, the longest axis of the ossified diaphysis was measured. Machines used in the study were GE Voluson 730 Expert, transabdominal probe (AB2-7, 2-7 MHz broadband curved array transducer) and GE Voluson 730 PRO, transabdominal probe (4CA, broadband curved array transducer).

All the sonographers who participated in the GUSTO study were accredited. They were trained according to standard operating procedures and assessed by the lead sonographer at each site. All the images obtained were standardized with all the necessary landmarks identified as afore-mentioned. Re-assessments were conducted on a regular basis to ensure standardization between the sonographers and across sites. Quality control monitoring of randomly selected images by the sonographers ensured adherence to the protocol and reproducibility. The equipment was calibrated with quality assurance checks by service engineers every six months.

Gestational age (GA) was in exact weeks unless otherwise specified (e.g. 27 weeks and 5 days = 27.7 weeks). GA at first antenatal ultrasound dating scan was estimated by crown-rump length (CRL) if it was not larger than 84 mm [33, 34]. Otherwise, GA at the first scan was estimated by BPD [34, 35]. The estimated fetal weight (in grams) was based on AC, BPD, HC and FL measurements using the Hadlock algorithm [36], log (EFW) = 1.3596 − 0.00386 × AC × FL + 0.0064 × HC + 0.00061 × BPD × AC + 0.0424 × AC + 0.174 × FL. We chose this particular Hadlock formula a priori over the others, because it was based on four biometric parameters which may provide a more accurate estimate for the EFW than its counterparts that based on less than four parameters. For gender-specific analyses, gender was based on neonatal assessment.

We used the mixed-effect linear regression models for longitudinal data to develop the unconditional and conditional standards [37, 38]. This method accounts for non-normal distribution in the biometric measurements by Box-Cox transformation, captures nonlinear relationship between biometry and GA via a linearizing function of GA obtained by fractional polynomials, and generates z-scores and percentiles by mixed-effects linear modelling of the transformed variables. The method has been used in the development of EFW references for an African [14] and Norwegian population [39]. Details of the methods are available in the aforementioned publications. To facilitate the application of our findings, we provide electronic spreadsheets as Online Digital Supplements, which also provide the formula details in the program codes.

We conducted model diagnostics in three ways. Firstly, we used the detrended Q-Q plot, also known as the worm plot, for visual assessment of the normality of z-scores [40]. Secondly, for each pair of measurements from two consecutive visits, the Pearson’s correlation coefficients between the conditional z-score at the present assessment and the z-score at the previous assessment were calculated. Properly developed standards should give approximately zero correlation. Thirdly, we examined the proportions of observations below the 10th and above the 90th percentiles. The associated 95 % confidence interval for each proportion was obtained via a general linear model for binary outcomes with an identity link. Logistic regression was also used to check that the proportions below (or above) the specified percentiles were independent of predictor(s). For the unconditional standards, the predictor examined was GA. For the conditional standards, the predictors examined were GA and the previous AC or EFW. Properly developed standards should show no association with these predictors. The regression models used the Huber-White robust standard error estimator for statistical inference to adjust for multiple ultrasound scans per participant [38, 41]. All statistical analyses were performed using the Stata statistical software version 12.1 [42].

There were 626 Chinese women who conceived naturally with singleton pregnancies in the GUSTO study. A total of 312 women were excluded as per the exclusion criteria afore-described. One woman was excluded due to implausible AC measurements during data inspection. The final analysis sample for constructing AC standards consisted of 920 measurements from 313 women, of whom 294 had three and 19 had two ultrasound scans other than the GA dating scan. The analysis sample included 168 male and 145 female foetuses. Inspection of the other biometric measurements was also conducted to identify implausible measurements. This led to exclusion of two more women from the EFW analysis. The final analysis sample for constructing EFW standards consisted of 901 measurements from 311 women, of whom 280 had three and 30 had two EFW measures. The analysis sample included 167 male and 144 female foetuses.

The mean (SD) age, years of formal education, height and pre-pregnancy weight of the women was 31.2 (4.8) years, 13.8 (3.8) years, 159.2 (5.8) cm and 53.0 (6.8) kg, respectively. Mean (SD) of paternal height was 171.3 (6.2) cm. Mean (SD) of GA at dating scan was 12.5 (0.8) weeks. Forty seven per cent were nulliparous. Ninety eight per cent were married or co-habiting. Fifty three per cent had spontaneous labor. The measurements of hemoglobin levels before 15 completed weeks of GA were only available in 176 out of the 313 eligible women, and the mean (SD) hemoglobin levels was 12.5 (1.1) g/dL. Twenty three per cent had caesarian section. All babies were born at full term, with four admitted to neonatal ICU. There was no maternal ICU admission after delivery. Table 1 summarizes the AC and EFW measurements between 18 and 35 completed weeks of GA.

For AC, the Box-Cox transformation controlling for GA gave a power transformation parameter estimate \( \hat{\lambda} \)=0.09 (95 % CI: -0.04 to 0.22) in the analysis pooling both genders, \( \hat{\lambda} \)=0.03 (95 % CI: -0.16 to 0.22) in males and \( \hat{\lambda} \)=0.08 (95 % CI: -0.10 to 0.26) in females. The null hypothesis of normal distribution (equivalent to λ = 1) was rejected (P < 0.001). All three parameter estimates were near zero with their respective 95 % confidence intervals including zero, suggesting that the natural (base e) logarithmic transformation (equivalent to λ = 0) was suitable for transforming the distribution towards normality [37]. Thus we used log(AC) in the analysis (base e).

The Box-Cox transformation found that a power transformation with a power term -0.1 was appropriate for normalizing EFW: \( \hat{\lambda} \)=-0.10 (95 % CI: -0.14 to -0.06) for analysis pooling both genders, \( \hat{\lambda} \)=-0.11 (95 % CI: -0.18 to -0.05) for males and \( \hat{\lambda} \)=-0.10 (95 % CI: -0.16 to -0.04) for females. Thus we used EFW^− 0.1 in the analysis.

Results for the linearizing function and the subsequent linear mixed-effect model estimation for transformed AC and EFW are shown in Table 2. In the analysis pooling both genders, the nonlinear relation between log(AC) and GA can be described by a second-degree fractional polynomial model with powers -2 and 1. The systematic component of the model was

where f(GA) = GA^− 2 − 0.0297/168.4982 × GA was the linearizing function. The transformed GA variable, f(GA), linearized the non-linear relation between log(AC) and GA. Upon fitting a mixed-effects regression of log(AC) on the transformed GA variable f(GA), the estimated mean of log(AC) is given by (model details in Table 2):

The analysis results for the (power-transformed) EFW measurements are also given in Table 2. They can be interpreted in the same way as above for log(AC).

The unconditional z-scores and percentiles can be obtained by plugging the information in Table 2 into the formula in Royston [37] (also in Landis et al. [14], Cheung [38], Johnsen et al. [39]). Details of the calculations and details of all the unconditional charts discussed can be found in the supplied electronic spreadsheet UnconditionalChart_FetalGrowth_Supp.xls (see Additional file 1). The unconditional standard for both genders pooled is shown in Fig. 1 in terms of percentiles, after back-transforming the AC and GA values to their original scales. Each percentile increased roughly linearly with GA from 18 to 36 exact weeks. The vertical distance between percentiles increased over time, indicating increase in variability as fetuses grew. Similar patterns were observed in both gender-specific charts (Fig. 2). The median AC values for males were about 1 % to 2 % larger than those for females at the same GA. The 10th and 90th percentile values for males were about 2 % and 1 %, respectively, higher than those for females at the same GA.

Unconditional standards for EFW pooling both genders and specific to each gender were shown in Figs. 3 and 4. The gain in EFW showed acceleration as GA increased within from 18 to 36 weeks. Similar to AC, there was more gender-difference in the 10th percentile than in the 50th or 90th percentiles. The gender differences in grams were similar between the 10th, 50th and 90th percentiles. In terms of percentage, however, the differences in the 10th, 50th and 90th percentile were approximately 4 %, 3 % and 2 %, respectively.

The conditional z-scores and percentiles can be obtained by plugging the information in Table 2 into the formula in Royston [37] (also in Landis et al. [14], Cheung [38], Johnsen et al. [39]), which has also been implemented in the supplied electronic spreadsheet ConditionalChart_FetalGrowth_Supp.xlsx (see Additional file 2). To illustrate, participant number 020-66086 had EFW 1138 grams at gestational age 27.7 and 2024 grams at 33.7 weeks (Fig. 5). Using the unconditional standards, the second assessment would be considered normal as it fell above the unconditional 10th percentile (z-score = -0.46). However, using the conditional standards to take into account the previous EFW measurement, the present EFW fell below the conditional 10th percentile (z-score = -1.62).

The detrended Q-Q plot for the unconditional standard for AC, pooling both genders showed satisfactory fit (Additional file 3: Figure S1). Most of the data points were around the horizontal lines, which indicate normal distribution of the z-scores. Only six of the 920 z-scores fell outside the 95 % confidence interval (CI). The detrended Q-Q plot of the conditional z-scores also showed satisfactory fit (Additional file 4: Figure S2). Only eight of the 607 z-scores fell outside the 95 % CI. The correlation coefficients between the conditional z-scores and the previous (unconditional) z-scores for each pair of two consecutive study visits (i.e. between 19-21 and 26-28 weeks, and between 26-28 and 32-34 weeks) were both close to zero: -0.06 (P = 0.319) and 0.03 (P = 0.656), respectively. The proportions of the 920 measurements classified below the 10th percentile and above the 90th percentile of the unconditional standards pooling both genders were 10.9 % (95 % CI: 8.4 % to 13.4 %) and 9.7 % (95 % CI: 7.5 % to 11.8 %), respectively. Furthermore, logistic regression showed no association between gestational age and the classifications below the 10th or above the 90th percentile (each P > 0.1). Similarly, the proportions classified below the 10th and above the 90th percentiles of the conditional standards were 10.2 % (95 % CI 7.8-12.6 %) and 8.6 % (95 % CI 6.4-10.7 %), respectively. None of the classifications was associated with either gestational age or the previous AC measurement (each P > 0.1). Results on model diagnostics for the gender-specific unconditional and conditional standards were similar and both demonstrated satisfactory performance (details not shown).

The detrended Q-Q plots of unconditional and conditional EFW standards indicated sufficient fit, with the data points mostly scattered around the horizontal lines (Additional file 5: Figures S3; Additional file 6: Figures S4). Only ten of the 901 unconditional z-scores and seven of the 590 conditional z-scores fell outside the 95 % CIs. The correlation coefficients between the conditional z-scores and the initial (unconditional) z-scores for each pair of two consecutive study visits were both close to zero: 0.05 (P = 0.358) and -0.05 (P = 0.355). The proportions of the 901 EFW measurements classified below 10th percentile or above the 90th percentile of the unconditional standards pooling both genders were 9.2 % (95 % CI: 6.7 % to 11.7 %) and 10.4 % (95 % CI: 7.9 % to 12.9 %), respectively. Logistic regression showed no association between gestational age and the classifications (each P > 0.1). Similarly, the proportions classified below the 10th and above the 90th percentiles of the conditional standards were, respectively, 10.3 % (95 % CI: 7.9 % to 12.7 %) and 9.5 % (95 % CI: 7.2 % to 11.7 %). None of the classifications was associated with gestational age or previous EFW measurement (each P > 0.1). Results on model diagnostics for the gender-specific unconditional and conditional standards were similar and both demonstrated satisfactory performance (details not shown).