The study protocol was approved by the University of Malaya Medical Center, Medical Research Ethical Committee (MREC) with MECID.No: 2021329-9997. All the data involved in the current research project originated from Pusat Perubatan Universiti Malaya (PPUM), a government-funded medical institution in Kuala Lumpur, Malaysia. Prenatal data from May 2011 to Apr 2021 were extracted from the system. Scan records with missing values, pregnancy with multiparity, stillbirth, and consist of values that fall outside the range of three times the interquartile range were removed. Only one ultrasound measurement was used for each fetus. We have a total of 68,897 scans from May 2011 to Apr 2021. A total of seven features were extracted from the scan records and are described in Table 1. All ultrasound measurements were performed by sonographers certified by the Fetal Medicine Foundation.

Our primary outcome was to predict SGA at birth. We defined SGA at birth when the birth weight is less than the 10th centile, based on the INTERGROWTH-21st preterm and term birth weight chart (17, 18). If birth weight is above the 10th centile, it is defined as appropriate gestational age (AGA) at birth.

We adopted the first nine-year datasets from May 2011 to Apr 2020 (n = 67,063 scans) to generate the fetal growth chart using the Lambda-Mu-Sigma (LMS) statistical method (19). The LMS method is an established method in creating reference charts (20–22). The LMS method summarizes the distribution of fetal biometrics by gestational age in three aspects, which are Lambda (L) which indicates the skewness of the distribution of fetal biometrics by Box-Cox transformation power, Mu (M) which indicates the Median of the fetal biometric, and lastly the Sigma (S) that indicates the coefficient of the variation of the fetal biometric. Nature smoothing spline function was applied to obtain the smoothed value of Lambda, Mu, and Sigma for each gestational age, these values were then fed into the equation as followed to calculate the percentile value in a particular gestational age:(1)C=M(1+(L⋅S⋅Z))1Lwhere C is the unit value at a particular percentile level to be calculated; M, L, and S are the Mu (median), Lambda (skewness of distribution), and Sigma (coefficient of variation) as described previously; Z is the corresponding Z-Score of the percentile in a normalized distribution (e.g. for the value of percentile 2.5th, 5th, 10th, 25th, 50th, 75th, 90th, 95th, and 97.5th, Z will be substituted as −1.960, −1.881, −1.645, −1.282, −0.675, 0, 0.675, and 1.282, 1.645, 1.881, 1.960). The generated unit value at a particular percentile using the LMS method was aggregated and presented as a fetal growth reference curved with intervals of one week by gestational age.

The difference between our LMS fetal growth chart and the growth chart from WHO (8), INTERGROWTH-21st (6, 23), and Hadlock (5) was compared using relative percentage difference (Equation 2).(2)Relative%Difference=(InternationalGrowthChartCentileValue−LMSCentileValueLMSCentile)×100%

The 10th-year datasets (May 2020 to Apr 2021, n = 1,834 scans) were used to evaluate the accuracy of our fetal growth chart in predicting SGA at birth. Fetuses with EFW that fall below the 10th centile were predicted as SGA at birth while fetuses with EFWs above the 10th centile were predicted as AGA at birth.

The evaluation was divided into three parts. First, the performance of the local growth chart, generated using local data with the LMS method, in predicting SGA at birth was evaluated. Second, we analyzed the performance of the local growth chart in predicting SGA at birth in preterm and term infants. A preterm infant is defined as when an infant is born before the 37th week while a term infant is defined as when an infant is born after the 37th week (24). Third, we further analyze the performance of the chart in predicting SGA at birth using second and third-trimester data for term and preterm infants. Any data with gestational age between the 13th week and 27th week was identified as “second trimester” and data with gestational age more than or equal to 27th was identified as “third trimester”.

The performance of the local growth chart and WHO (8), INTERGROWTH-21st (6, 23), and Hadlock (5) fetal growth charts in predicting SGA at birth were compared. The dataset is imbalanced as the number of SGA is much lesser than the number of AGA cases, balanced accuracy is used instead of accuracy (25). Balanced accuracy is the mean of sensitivity and specificity. Balanced accuracy, sensitivity, and specificity were reported.

Statistical analysis was performed to test if there were any significant differences between the AGA and SGA. For continuous variables, normality test was performed to check for data distribution. If data is normally distributed, Student T-test is used else non-parametric Mann-Whitney test is used for analysis. For categorical variables, chi-square test was performed to determine if there is a significant difference between AGA and SGA. The data were deemed significantly different if p < 0.05.

Figure 1 shows the data distribution used to generate our fetal growth chart curves. In 2011, UMMC had just started using a proper electronic system for keeping patient records and thus the datasets in 2011 were limited and all of them were in 2nd trimester. Table 2 tabulates the patients' characteristics from May 2020 to Apr 2021, which was used for evaluating the accuracy of various growth charts in identifying SGA at birth. There are 781 newborns with 706 AGA and 75 SGA, defined using the birth weight of the infants. There were significant differences between AGA and SGA fetuses, where the occurrence of maternal hypertension, pre-eclampsia, gender, as well as the birth anthropometric measurements of birth weight, length, and head circumference differ. No significant differences were observed for maternal age, anemia, gestational age at birth, and APGAR score for 1 min and 5 min.

Figures 2, 3 shows the reference charts for biparietal diameter, head circumference, abdominal circumference, femur length, and estimated fetal weight generated using the LMS method. Supplementary Tables S1–S5 show the centile estimations for completed weeks of gestation.

Figure 4 shows the relative percentage difference of the EFW centile curve between the local generated curve and the WHO, INTERGROWH-21st^, and Hadlock centile curves. A positive percentage error indicates the percentile value of growth reference (WHO/ INTERGROWH-21st/Hadlock) is larger than the percentile value of growth reference developed in our study. In other words, the positive and negative percentage error representing fetal growth are over- and under-estimated, respectively, if international growth references (WHO/ INTERGROWH-21st/Hadlock) are adopted for fetal growth assessment in the local population.

In our study, we found that the discrepancy between our EFW chart and the WHO growth chart ranged from +18% to −10% across gestation, indicating an 18% of overestimation at early gestation and a 10% underestimation of fetus growth at late gestation. From Figure 4A, the WHO growth chart overestimated the FGR fetus growth (light grey line) and the overestimation of FGR fetus growth was 15% in the 15th week and dropped gradually to below 5% after the 23rd week. The percentage error for evaluating EFW below the 10th centile between our growth chart and WHO curve was small, approximately 0.56%, between 25th to 37th week.

The Hadlock curve exhibited a similar reducing trend of percentage error when compared to our local EFW curves. The percentage error was 16% in the 15th week and dropped to below 5^% after the 25th week. The percentage error for evaluating EFW below the 10th centile from 25th to the 37th week between our growth chart and Hadlock curve was approximately 2.03%.

As there were no information available before the 21st week in the INTEGROWTH-21st chart, the comparisons were only made between the 22nd to 40th week. From Figure 4C, the INTEGROWTH-21st chart exhibited a negative discrepancy between the 22nd to 25th week, indicating that some of the fetuses may be misdiagnosed as AGA. The discrepancy increased to positive after the 25th week and declined around the 32nd week.

Table 3 shows the results of each fetal growth reference chart in predicting SGA at birth. Based on the result, we noticed that our fetal growth reference chart generated using the LMS method achieved similar balanced accuracy (69%), as the WHO chart. The recall for both local and WHO charts were similar, achieving 42% and 43%, respectively. The INTERGROWTH-21st chart had the poorest performance with a balanced accuracy of 57% and 19% recall. All charts had similar specificity of 95%–96%.

To further understand the performance of each chart in predicting SGA, we performed another analysis by segregating SGA into preterm and term SGA (Table 4). Based on the result, we observed a similar pattern as observed in Table 3, where the WHO and our local chart achieved the highest accuracy in predicting SGA for both preterm and term infants, followed by the Hadlock and INTERGROWTH-21st charts. For preterm infants, via the LMS method, the WHO and local growth chart achieved balanced accuracy of 76% and recall of 65% in predicting SGA at birth. The Hadlock chart had slightly lower accuracy and recall as compared to the WHO and local charts. The INTERGROWTH-21st chart showed the lowest balanced accuracy of 63% and recall of 31% for preterm SGA. For term infants, the WHO chart depicted the highest balanced accuracy, with 66% and recall of 35% for SGA, followed by our local growth chart with 65% balanced accuracy and 34% recall. The INTERGROWTH-21st chart, again, showed the lowest accuracy of 55% and recall of 14%. For preterm SGA, the INTERGROWTH-21st chart showed the highest specificity of 94%, followed by the Hadlock chart, with 90% specificity. The WHO and our local charts had slightly lower specificities, 88%. For the term SGA, all four charts showed similar specificity of approximately 96%.

Table 5 shows the results of various fetal growth charts in predicting term and preterm SGA in the 2nd and 3rd trimesters. Interestingly, although the INTERGROWTH-21st chart had the poorest performance in Tables 3, 4, it had the highest balanced accuracy when predicting SGA in 2nd trimester but dropped greatly in 3rd trimester (Balanced Accuracy: Preterm SGA: 75% drop to 59% and Term SGA: 64% drop to 55%). The WHO chart had poor performance in 2nd trimester but improved substantially in 3rd trimester (Balanced Accuracy: Preterm SGA: 66% to 79% and Term SGA: 56% to 67%). The Hadlock chart achieved an average balanced accuracy of 52%–76%.

As compared to the WHO and INTERGROWTH-21st charts, our local growth chart showed a more consistent trend in the 2nd and 3rd trimesters in predicting SGA at birth. The discrepancy of balanced accuracies between the 2nd and 3rd trimesters was as large as compared to the WHO and INTERGROWTH-21st charts. From Table 5 for preterm SGA, our local growth chart achieved 70% balanced accuracy in the 2nd trimester and increased to 78% in the 3rd trimester while for term SGA, our local growth chart shows 58% balanced accuracy in the 2nd trimester and improved to 66% in the 3rd trimester.