Introduction
FBW is an important indicator for the optimal growth, survival, and future well-being of newborns. A normal size infant is one weighing greater than 2500 g and less than 4000 g. Low birth weight is between 1500 and 2500 g. A birth weight between 1000 and 1500 g is considered a very low birth weight. Below 1000 g is an extremely low birth weight and more than 4000 g is a high birth weight or macrosomia. So, FBW estimation is to estimate these values before birth when the infant is inside his/her mother’s womb. Maternal ethnicity, infant sex, plurality, nutrition, altitude, education, and smoking affects the entire birthweight distribution in a country [
1‐
7].
Globally 2.4 million children died in the first month of life in 2020 [
8]. There are approximately 6700 newborn deaths every day, amounting to 47% of all child deaths under the age of 5 years [
9]. In Africa, 1.12 million newborn deaths occur annually [
10]. Preterm birth, intrapartum-related complications, infections, and birth defects cause most neonatal death, and our country Ethiopia is among the top 10 countries having the highest number (97 per 1000 live births) of newborn deaths, 2020 [
10].
Birth weight estimation is an input for labor and delivery management plan which is used to determine the procedure taken during this period, so it is so important to know the birth weight of fetal before his/her birth date to overcome intrapartum-related complications associated with both giving birth to an infant having large weight [
11] and small weight [
12] which are greater than or equal to 4500 g and less than 2500 g respectively. Both extremely large and small fetal birth weights may lead to complications that cause lifetime impairment of body parts or death of the infant and mother. Regular and reliable birth weight estimation throughout the pregnancy period is vital to avoid those complications as early as possible.
There are two common methods to estimate FBW; clinical method and sonographic method. In the clinical method physicians measure the fundal height of the pregnant women then calculate FBW by using formula which is used to estimate FBW or perform abdominal palpation procedure to determine the fetal birth weight. In ultrasound machine there is a built-in software which calculate FBW. Estimation of FBW using ultrasound requires predefined formulae (model) which describes birth weight as dependent variable and some other variables like fetal biometry parameters as an independent variable. Several formulae [
13‐
26] have been developed for estimating fetal weight by ultrasound. The most popular formulae are Shepard [
21], Campbell [
18], and Hadlock’s [
17]. These formulae are included in most ultrasound equipment software packages. These formulae involve different types of fetal biometric parameters obtained by sonographic measurements. The measurement is taken by physicians during ultrasound examination. The techniques outlined for assessing FBW typically yield a reasonable margin of error. However, inaccuracies may arise due to factors such as insufficient expertise, subjectivity in assessing fetal biometry, fundal height, and abdominal palpation. It is worth noting that the mathematical models employed for birth weight estimation are derived from populations in other countries, thereby resulting in an estimation error of over 10% for Ethiopian births when utilizing such models [
27]. A birth weight estimate with an error margin of 10% or less is deemed acceptable [
28,
29].
Various mathematical models have been suggested for the estimation of FBW, as documented in scholarly research [
28‐
32]. For example, a model was developed in Pakistan by S. Munim et al. [
30] using the Regressions with Leaps and Bounds method based on population data. This model reported systematic and random errors of 10 and 250 g, respectively. Another study was conducted in India by S. Hiwale et al. [
32] where multiple stepwise regression (MSR) and lasso regression methods were utilized to create population-based models with adjusted R
2 values of 0.656 and 0.633, respectively. The accuracy of both models was determined to be 81% and 82% for estimating within ± 10% of the actual birth weight (ABW). Furthermore, C. Li et al. [
33] proposed a gestational age stage-based birth weight prediction model for the Chinese population. The model employed multiple linear regression (MLR), fractional polynomial regression (FPR), and volume-based models (VM) to achieve systematic errors of 6.97%, 0.26%, and 0.36%, respectively.
A linear regression model was developed using obstetric factors (such as gravidity, gestational age, SFH, body mass index of the mother, membrane status, sex of the neonate, and actual birth weight) to estimate fetal weight by A. Yiheyis et al. [
34]. Johnson’s formula was also evaluated to determine its suitability as a model for south western Ethiopia. R. Ramya et al. [
31] utilized image processing algorithms on fetal ultrasound images to automatically measure fetal biometry, thereby increasing the accuracy of FBW estimation. The study involved measuring four major fetal biometrics (AC, HC, BPD, and FL) through different image processing steps. However, these techniques are found to be less accurate in estimation of fetal birth weights.
Deep learning techniques has been also employed in literatures for automatic estimation of fetal biometry [
35‐
38]using ultrasound image or video data. These and the above techniques are not effective in estimating fetal births of the Ethiopian population. In order to address these issues, it is necessary to conduct population-specific measurements of parameters related to FBW. This will enable the development of effective and accurate models for estimating FBW, which will facilitate proper planning and management of labor and delivery. In this paper we propose the use of an automatic image processing algorithm for measuring fetal biometry to develop a mathematical model that accurately estimates FBW based on our Ethiopian population.
Discussion
Proper and effective labor and delivery management plans for pregnant women in health facilities require the main input parameter of FBW. Factors that affect FBW include maternal ethnicity, infant sex, plurality, altitude, education, and smoking [
1‐
7]. Typically, a normal infant birth weight ranges from 2500 to 4000 g, and deviations from this range can result in complications for both the mother and the fetus [
11,
12].
During pregnancy, FBW can be estimated either by a clinical or sonographic method. The latter requires predefined formulae or models that describe birth weight as a combination of variables. However, these estimation methods can be unreliable due to the subjective nature of parameter measurement and the ineffectiveness of the models used for our country’s population.
The aim of this study was to develop a FBW estimation model tailored to the Ethiopian population, using a dataset of 484 singleton pregnant women who underwent sonographic assessments. The dataset included 1,452 fetal ultrasound images, with fetal biometry variables measured by physicians and image processing algorithms, as well as sociodemographic, obstetric, and date variables. The dataset was divided into modeling and testing subsets.
Multiple linear regression (MLR) analysis was used to develop the FBW estimation model via two approaches. The first approach incorporated fetal biometry variables measured by physicians in combination with other variables. Independent variables were selected using correlation analysis based on their strength and nature of relation with the dependent variable (actual fetal weight), and included abdominal circumference, gestational age, and maternal ethnicity.
In the second approach, an image processing algorithm proposed by the study was used to measure fetal biometry from the collected ultrasound images. Fetal biometry variables measured by this algorithm and other variables were analyzed to select variables that had a strong and positive relationship with the dependent variable (actual birth weight). The independent variables selected for MLR analysis to develop another FBW estimation model included abdominal circumference, biparietal diameter, femoral length, and gestational age.
The analysis showed that the model based on fetal biometry measured by image processing algorithm provided estimates with less than10% error in 78.9% of the estimated values during the model testing procedure, with a mean percentage error of 5.89%. In comparison, the model based on fetal biometry measured by physicians provided estimates with < 10% error in 70.61% of the estimated values, with a mean percentage error of 7.53%. The mean percentage errors were calculated from the entire test set. Additionally, as indicated in the model summary tables (Tables
3 and
7), the model based on fetal biometry measured by image processing algorithm had a higher level of prediction (R-value) than the second model. Therefore, the model based on fetal biometry measured by image processing algorithm was chosen as the final model.
This study compared newly developed models with pre-existing models for estimating FBW. A literature review identified 35 models that utilized only four commonly measured fetal biometrics as independent variables, and were developed for general fetal weight estimation (excluding models for low weight and macrocosmic fetuses). Selection criteria based on population and year of publication were used to select models for analysis. The accuracy of estimated fetal weights was compared to actual birth weights using the mean percentage error (MPE). The final new model had an MPE of 5.89%. Among the compared models, Jordaan et al. [
16] and Hadlock et al. [
25] had an MPE of less than 20%, while the others had an MPE between 20 and 30%. Please refer to Table
9 for more details.
Table 9
Comparative analysis of fetal birth estimation models (HC-head circumference, BPD- biparietal diameter, AC- abdominal circumference, FL-femoral length, GA-gestational age)
1. | | AC-BPD-HC-FL | USA | 18.23% |
2. | | AC-BPD | South Africa | 15.20% |
3. | | AC-BPD | China | 23.59% |
4. | | AC-FL | Italy | 22.02% |
5. | | AC-HC-FL | USA | 27.44% |
6. | | AC-BPD-FL | Japan | 21.07% |
7. | | AC-BPD-FL | Hong Kong | 22.48% |
8 | | AC-FL | Pakistan | 21.81% |
9 | | AC | UK | 25.69% |
10. | | AC-BPD | Germany | 21.23% |
11. | The current model | AC-BPD-FL-GA | Ethiopia | 5.89% |
In summary, our experimental results indicated that the model based on fetal biometry measured by the image processing algorithm outperformed the model based on fetal biometry measured by physicians in terms of mean percentage error and R and R2 values in the model summary of each model. Additionally, this model yielded better results compared to existing FBW estimation models.
The proposed approach holds the promise of reducing both infant and maternal mortality rates by providing precise fetal birth weight estimates, which is a pivotal factor that underpins effective, accurate, and appropriate obstetric planning, management, and decision-making. Additionally, the model can be integrated into portable devices such as point-of-care ultrasound machines, making it accessible and applicable in rural areas.
We acknowledge that the proposed model was built only using datasets gathered from South west of Ethiopia and the capital Addis Ababa. Although the model demonstrated good performance, its effectiveness could be enhanced by increasing the variability of the dataset through the collection of additional data from all regions of the country. Additionally, the study was restricted to pregnant women within the 30–42 gestational age range. Incorporating more study variables such as maternal body mass index before, during, and after pregnancy could enhance the model effectiveness. Utilizing machine learning techniques to develop models tailored to specific groups such as small and large for gestational age fetuses may also improve the model’s performance in estimating FBW for these groups.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.