Background
Birth weight is a primary measurement and significant indicator to ensure the optimal growth, survival, and future well-being of new-borns. Deviation from normal delivery weights (2500–3999 g), such as low birth weight (LBW) (< 2500 g) and macrosomia (> 4000 g) could lead to some negative consequences on neonatal health [
1‐
3]. While macrosomia may cause neonatal and maternal morbidity [
4], LBW is well-documented to be one of the most contributing factors to the neonatal mortality [
1]. LBW is defined as weight less than 2500 g at birth regardless of gestational age (GA) and can be caused by preterm birth or intrauterine growth restriction [
5]. In this paper, LBW includes both preterm and term new-borns of appropriate for GA.
Routine and reliable estimates of foetal weight at a given GA throughout pregnancy are vital. These estimates could create evidence-based track records/analysis to assist medical practitioners to detect the signs of potential LBW during pregnancy and provide the appropriate interventions. Although a wide range of simple and advanced multivariate weight prediction models based on clinical and ultrasonic measurements has been developed, most are only based on maternal or foetal factors [
6‐
25]. Less is known about the combinations of these characteristics to estimate foetal weight during pregnancy despite the fact that birth weight is significantly associated with characteristics of both mother and foetus [
1,
26].
Several models based on combined maternal and neonatal characteristics have been developed and reviewed, these existing models were mostly developed based on the information available at delivery time [
27,
28]. In most developing countries, the availability of foetal biometric measurements during pregnancy is low, particularly in rural areas due to limited access to ultrasound machines and skilled personnel [
29]. Westerway et al. (2000), Loughna et al. (2009), and Papageorghiou et al. (2014) have used a large number of ultrasonic measurements to develop formulas that estimate foetal biometric characteristics at a given GA [
30‐
32]. These formulas then could be used to fill the foetal database gaps during pregnancy when ultrasound facility is absent.
The present research develops foetal weight prediction models based on combined maternal and estimated foetal biometric characteristics to estimate foetal weight at any given GA. The proposed models can be simply implemented in low-resource primary health care centres where ultrasound machines and trained ultra-sonographers are not always available. The predicted foetal weight will assist in the development of foetal growth charts for Indonesia. No such charts currently exist for the Indonesian population.
Discussion
Our study highlights that the use of combined maternal and estimated foetal biometric characteristics can provide a reliable estimate of delivery weights between 35 and 41 weeks of GA. This result confirms the previous study that shows a significant association between birth weight and characteristics of mother and foetus [
1,
26].
Both clinical and estimates of ultrasonic predictors are used in our proposed models. Maternal FH measurement was selected as one of the clinical predictors as it is one of the most recommended and accessible predictors to estimate foetal weight and monitor foetal growth during pregnancy [
3,
23,
49,
50]. Although the clinical approach using FH screening had reportedly low sensitivity for detecting intergrowth and birth weight abnormalities (ranged 16–45%) [
51,
52], it is a simple and inexpensive clinical activity [
29,
53], especially true in rural areas where ultrasound machines and skilled personnel are not always available. The utility of FH remains an important first level screening tool, widely used during routine ANC in both high and low income settings [
29] even though it had high false-negative rates for small for GA [
53].
In ultrasonic settings, foetal biometric characteristics monitored during pregnancy include HC, biparietal diameter (BPD), occipitofrontal diameter (OFD), AC, and femur length (FL). These characteristics are routinely measured by ultrasound every 5 weeks after the first initial dating scan (between 8 and 14 weeks’ gestation). The standard ranges for ultrasonic measurements are (14–18), (19–23), (24–28), (29–33), (34–38), and (39–42) weeks [
54] or at least once every trimester of pregnancy, i.e. between weeks 10–14 (first trimester), 20–24 (second trimester), and 30–32 (third trimester) [
55].
Assessment of foetal biometric characteristics during ANC is vital to ensuring normal foetal size and safe delivery. In the absence of ultrasound facility, particularly in low-resource primary health care settings, the measurements of these characteristics are not always accessible. Therefore, a reliable prediction of these characteristics during pregnancy would be a proxy of foetal biometrics and vitally improve the quality of ANC services in monitoring foetal inter growth assessment which currently remain low due to the database gaps [
56‐
59].
Several ultrasonic formulas to estimate the foetal characteristics at different GA have been developed [
30‐
32]. The foetal HC and foetal AC are widely recognised as the most influential predictors for predicting foetal weight [
10,
11,
46,
60,
61]. Our results show that the best fit formulas to estimate these foetal characteristics at a given GA in our population were based on the Australian population [
30].
To the best of our knowledge, in the majority of Indonesian primary health care centres where ultrasound facility is not accessible, none of the existing ultrasonic formulas were adopted to estimate foetal HC and foetal AC. Therefore, the formulas potentially can be deployed to fill in the database gaps on the inter growth process of foetus during pregnancy. Consequently, early informed intervention could be initiated to prevent abnormal growth and delivery weights.
Several techniques have been available to reduce collinearity, such as centering, multiplying variables by various constants (scaling), the use of orthogonal polynomials, and other transformations [
62]. Currently, the use of automated machine learning, such as Genetic Algorithm rather than a conventional fractional polynomial approach has also been applied to model multiple biometric variables of foetus that are highly correlated [
54].
In this paper, we used the best subset selection algorithm to prevent the inclusion of highly correlated variables and select the best subset of predictors to be included in the models. It has been emphasized that a formula for estimating foetal weight should be simple and straightforward to be used by doctors and midwives and be easily understood by patients [
63]. This would improve the quality of communication, information, and education as part of routine ANC service in low-resource primary health care centres.
Based on our comparison analysis, the proposed Models (1), (2), (3), and (4) produced the least mean prediction errors (between − 0.2 and − 2.4 g), the MAPEs (between 5.01 and 5.10%), and the MEDAPEs (between 4.10 and 4.22%). The mean percentage prediction error (MPE) steadily tended towards zero as the time interval between the last scan and birth decreased [
42]. Our MPEs were ranged between − 0.1 and − 0.3% in those born within 0 day (
n = 38) which are lower than the previous research [
42] reported by − 0.8% in those born within 1 day (
n = 198).
Our proposed models were unbiased for predicting weight between 35 and 41 weeks of GA. In the group born within 0 day of the last measurements, the MAPEs were ranged between 5.0 and 5.10% with 89–92% of predicted weights falling within 10% of the true birth weights which are smaller than those reported in previous study [
42]. This was particularly for Model (1) which was simply developed based on FH only.
The comparison between the proposed Model (1) and the widely used Johnson-Toshach model shows that Model (1) (developed based on the Indonesian data) was more accurate in predicting the estimated foetal weight than the Johnson-Toshach model (developed based on the United States data). Furthermore, the Johnson-Toshach model requires the knowledge of FS. The results presented in Table
3 also shows that the inclusion of FS in the model has not reduced the prediction errors in foetal weight estimations yet raise a subjectivity issue unless there is a standard protocol to determine FS with less error [
20]. Therefore, we recommend the proposed Model (1) be deployed in Indonesia and other countries with similar health systems and challenges for weight prediction.
Our comparison study confirms that the proposed Models (3) (based on FH and estimates of foetal HC) and (4) (based on FH and estimates of foetal AC) perform better than the ultrasonic models: the Jordaan, the Weiner II, the Hadlocks, the Stirnemann, and the Sotiriadis models. The incorporation of estimated foetal HC or estimated foetal AC has increased
R2 slightly (provided in Additional file
5: Table S5), but it did not improve the predicting accuracy (Table
3). However, access to these values will enable the practitioners to monitor foetal growth during pregnancy where advanced equipment, such as ultrasound, is not always available. Consequently, detecting foetal growth abnormality, such as small for GA, prematurity, intrauterine growth retardation, and LBW during pregnancy will be possible.
Strengths and limitations
Our retrospective study has investigated the utilisation of some commonly used foetal weight prediction models in Indonesia. Particularly, the combination between maternal and estimated foetal biometric characteristics was proposed. The aim of this combination was whether it could improve the prediction accuracy of foetal weight at any given GA in the absence of ultrasound machines and trained ultra-sonographers.
The retrospective cohort study was undertaken to provide baseline data on the selected primary health care centre. It is possible that women have used different health services than that reviewed in this study. Although this may result in underestimation in data records, it is unlikely to impact on the validity of the analyses. This study also encountered limitations associated with the accuracy of the information recorded on the manual pregnancy register or inaccurate data transfer to the electronic database. However, monitoring and controlling the process of data transfer was conducted to reduce potential error. Further study should be conducted to assess the efficacy of the proposed models using prospective data [
64].
The proposed prediction models are linear regressions. However, the authors have investigated non-linear models. The non-linear models did not improve the estimation accuracy. Therefore, complex models do not guarantee significant improvement in the prediction accuracy. Furthermore, due to the fact that the objective of the study is to provide simple yet reliable foetal weight estimating models for low-resource areas, we are recommending the proposed models. We believe that the findings can be applied in other low-resource settings to improve ANC services.
Acknowledgements
We are grateful to the Australian Agency for International Development (AusAID) for funding DA’s PhD scholarship in Mathematical Sciences at the School of Science, RMIT University, Melbourne, Australia.
The authors would like to thank Feri Anita Wijayanti, M.Mid, Bd. for the provision of antenatal care references in the Indonesian context. We would also thank the dedicated midwives for their roles in supervising the data measuring and recording task in the primary health care centre.
The authors are greatly indebted to the Higher Degree Research (HDR) Language and Learning Advisor of RMIT University, Dr. Ken Manson, for his roles in providing language help and proofreading the article.
DA: PhD candidate in the Mathematical Sciences (Applied Statistics), School of Science (Mathematical and Geospatial Sciences), College of Science, Engineering, and Health, RMIT University, Melbourne, Australia and Junior Lecturer at Study Program of Statistics, Faculty of Mathematics and Natural Sciences, University of Lambung Mangkurat (ULM), South Kalimantan, Indonesia.
MA: Senior Lecturer of Statistical Quality Control and its applications in: manufacturing industry, air pollution control, software quality, univariate and multivariate processes, health industry, and the banking system, School of Science (Mathematical and Geospatial Sciences), College of Science, Engineering, and Health, RMIT University, Melbourne, Australia.
KM: Senior Lecturer of Applied Statistics and Mathematics, Market Research, and Numerical Analysis in aerospace engineering, clinical sciences, geomatic engineering, and oncology and carcinogenesis, College of Science, Engineering, and Health, RMIT University, Melbourne, Australia.