Machine Learning-Based Prediction of Abdominal Subcutaneous Fat Thickness During Pregnancy

Abstract

Objective:

Current evidence regarding the safety of abdominal subcutaneous injections in pregnant women is limited. In this study, we developed a predictive model for abdominal skin–subcutaneous fat thickness (S-ScFT) by gestational periods (GP) in pregnant women.

Methods:

A total of 354 cases were measured for S-ScFT. Three machine learning algorithms, namely deep learning, random forest, and support vector machine, were used for S-ScFT predictive modeling and factor analysis for each abdominal site. Data analysis was performed using SPSS and RapidMiner softwares.

Results:

The deep learning algorithm best predicted the abdominal S-ScFT. The common important variables in all three algorithms for the prediction of abdominal S-ScFT were menarcheal age, prepregnancy weight, prepregnancy body mass index (categorized), large fetus for gestational age, and alcohol consumption.

Conclusion:

Predicting the safety of subcutaneous injections during pregnancy could be beneficial for managing gestational diabetes mellitus in pregnant women.

Introduction

Gestational diabetes mellitus (GDM) is the most common medical complication during pregnancy, characterized by the development or diagnosis of diabetes for the first time during pregnancy.¹ The prevalence of GDM is increasing worldwide, and was reported to be 8.0% in Korea, based on data from the Health Insurance Corporation in 2012.² GDM can give rise to various maternal–fetal complications, including preeclampsia, abortion, fetal abnormalities, fetal death, macrosomia (large baby), neonatal hypoglycemia, hyperbilirubinemia, and neonatal respiratory distress.³ In the long term, it may also increase the risk of the offspring developing obesity and type 2 diabetes.⁴

Pregnant women with GDM often fail to maintain proper blood glucose because the disease duration is as short as 3–4 months, resulting in them having trouble acquiring essential knowledge for diet, exercise, and self-care to change their lifestyle.⁵ Approximately 20%–50% of pregnant women with GDM require a hypoglycemic agent to maintain proper levels of blood glucose⁶; however, the long-term impact of these agents, in terms of fetal stability, has not been verified. Pregnant women with GDM may eventually require insulin therapy if they fail to achieve target blood sugar levels through diet and exercise.²

Insulin is typically administered through subcutaneous injection into the abdominal fat; however, many pregnant women have concerns about potential harm to the fetus.⁷ Recently developed 4–5-mm needles offer pregnant women the option to inject insulin at a 90° angle in specific locations on the abdomen. These sites include the areas 1 cm above the pubis, 1 cm below the ribs, and 1 cm away from the umbilicus and flank.⁸ However, it is important to note that this guideline primarily pertains to Western populations, where obesity is defined as having a body mass index (BMI) of 30 kg/m² or higher.⁹

For Asians, obesity is defined as a BMI of 25 kg/m² or higher.¹⁰ Moreover, there is insufficient evidence to support a universal recommendation for abdominal subcutaneous injections for pregnant women due to racial diversity. Taking this into consideration and based on a study examining skin–subcutaneous fat thickness (S-ScFT) at different abdominal sites in pregnant Korean women, Hwang¹¹ suggested that injections should be administered at a 90° angle while folding the skin at the abdominal site, in accordance with the recommendation by Lori-Berard et al.⁸

However, it is necessary to refine this recommendation further through predictive modeling and factor analysis of abdominal subcutaneous injection by applying an individual-tailored algorithm. Previous studies revealed that the factors affecting ScFT in pregnant women are prepregnancy obesity, weight gain during pregnancy, fetal size, and gestational periods (GP).^12,13 However, in most of these results, the predictors were calculated by measuring the ScFT at one site of the abdomen. To ensure the safety of abdominal subcutaneous injections in pregnant women, it is important to investigate how ScFT is affected at each abdominal site according to GP.

Therefore, this study thus was conducted (1) to identify the differences in S-ScFT by general characteristics, (2) to identify the S-ScFT of each abdominal site according to prepregnancy BMI and GP, (3) to compare the performance of predictive models developed using three machine learning algorithms to determine the safety of abdominal subcutaneous injection, and (4) to select the best model from among the developed models and verify its clinical applicability through a pilot study.

Methods

Study design

This was a cross-sectional study.

Participants and samples

We enrolled 172 normal pregnant women with a GP of 24 weeks or more. Since S-ScFT was measured multiple times in some women throughout their pregnancies, each measurement was considered an independent case for analysis.

Machine learning is a form of artificial intelligence that has a high compatibility with regression analysis for prediction. Therefore, the number of samples was calculated assuming an algorithm suitable for logistic regression analysis. The number of samples suitable for logistic regression was calculated using the G*Power 3.1.7 program, assuming an odds ratio of 1.4, a significance level of 0.05, a power of 0.8, and a χ distribution normal, and 348 cases were derived. Subsequently, S-ScFT measurement was performed in 354 independent cases, which met the target sample size.

Inclusion criteria

The inclusion criteria were as follows: (1) A GP of 24 weeks or more, according to the screening recommendations for GDM (24 to 28 weeks of pregnancy),¹⁴ (2) married pregnant women 19 years of age or over, (3) cephalic fetal presentation and normal-range cardiac sound with no problem in fetal movement, and (4) fetal weight over 50% for the GP determined by sonography.

Exclusion criteria

The following individuals were excluded from the study: (1) Pregnant women within the high-risk group, including those presenting with multiple fetal indicators of possible complications, (2) those with an opened cervix or amniotic membrane rupture, (3) those with warts, bruises, pierced wounds, and scratches on the abdomen, and (4) those with pre-existing DM or GDM requiring insulin administration.

Measurements and instruments

Gestational periods

In Korea, most pregnant women are screened for GDM at 24 weeks and receive prenatal check-ups every 4 weeks. In this study, GP was classified into 24⁺⁰–27⁺⁶ weeks, 28⁺⁰–31⁺⁶ weeks, 32⁺⁰–35⁺⁶ weeks, and 36⁺⁰–40 weeks, and used for S-ScFT analysis.

Prepregnancy BMI

Prepregnancy BMI is a reliable indicator for the prediction of ScFT. Based on the prepregnancy weight, participants were divided into four BMI groups (underweight group: <18.5 kg/m², normal weight group: 18.5–22.9 kg/m², overweight group: 23–24.9 kg/m², and obese group: ≥25 kg/m²).⁹

Criteria for the judgment of S-ScFT for an abdominal subcutaneous injection

The S-ScFT is the sum of skin thickness and subcutaneous fat thickness. Although the thickness of abdominal subcutaneous fat differs depending on age, sex, and obesity,¹⁵ skin thickness is relatively constant at 0.2 cm regardless of other variables.¹⁶ A 2-mm-long needle is ideal, considering the needle hub length and pressure on the skin during injection.¹⁷ Thus, this study predicted that insulin injection is possible when subcutaneous S-ScFT is over 0.6 cm.

Data collection and procedure

Data were collected from December 2020 to August 2021. The participants were recruited from an obstetrics and gynecology hospital during their pregnancy. Among normal pregnant women who wished to participate in this study after reading the recruitment notice for the research project, those who met the selection criteria were enrolled as research participants. The participants' general and obstetric characteristics were collected through an interview when the written consent was obtained, and the examination results were investigated by a research assistant through the review of medical records.

The S-ScFT for each abdominal site by GP was measured in centimeters using an ultrasonic device (HS60; Samsung Medison, Seoul, Korea). The S-ScFT measurement site in the abdomen was divided into 12 sites on the left side of the abdomen (three rows: from the lower margin of the lateral rib to the iliac crest; four columns: from 1 cm away leftward of the umbilicus to the flank) (Fig. 1), measured by five obstetricians–gynecologists with more than 5 years of clinical experience. Fetal body weight was also estimated in kilograms using the same ultrasonic device.

FIG. 1.

Skin–subcutaneous fat thickness on the abdomen measurement sites

Ethics considerations

This study was approved by the Institutional Review Board of Woosuk University (No: WS-2020-12). Data were collected in compliance with the Declaration of Helsinki and Good Clinical Practice guidelines. After the S-ScFT measurement, a predetermined gift was provided to the participants.

Data analysis

Classical statistical method

The participants' characteristics were analyzed by frequency and percentage for categorical variables and by the mean and standard deviation for continuous variables. Differences in S-ScFT according to general characteristics were analyzed by t-test or analysis of variance. Post hoc analysis was performed using the Scheffe's test.

Machine learning analysis (machine learning predictive modeling and factor analysis)

Overview of the analysis

This study used machine learning with deep learning, random forest, and support vector machine algorithms. Site 7, considered an ideal location for an abdominal subcutaneous injection, was selected as the prediction site. S-ScFT was predicted and tested for each abdominal site by machine learning regression modeling through each algorithm. In addition, root mean squared error, the most popular performance metric in the evaluation of machine learning algorithms,¹⁸ was used to compare the performance level of the algorithms. All data analyses were performed using Rapid Miner Studio version 9.

Used variables

This study included variables that are known factors affecting abdominal ScFT according to previous studies and are presented in Table 1.

Table 1.

Variables Used in the Analysis

Variables	Descriptions
GP	Gestational period (weeks)
GP (C; Gr 4)	Categorized gestational period (weeks) 1 = 24⁺⁰–27⁺⁶ 2 = 28⁺⁰–31⁺⁶ 3 = 32⁺⁰–35⁺⁶ 4 = 36⁺⁰–40
Age	Maternal age (years)
Age ≥35 years (y/n)	Maternal age group based on the cutoff of 35 years 1 = yes; 35 years or older 2 = no; 19 years and older and younger than 35 years
Prepregnancy weight	Maternal prepregnancy weight (kg)
Weight during pregnancy	Maternal weight during pregnancy (kg)
Weight change during pregnancy	Maternal weight change during pregnancy (kg)
Prepregnancy BMI	Maternal prepregnancy BMI (kg/m²)
Prepregnancy BMI (C; Gr 4)	Categorized maternal prepregnancy BMI (kg/m²) 1 = underweight group; < 18.5 2 = normal weight group; 18.5–22.9 3 = overweight group; 23–24.9 4 = obese group; ≥25
Alcohol consumption (y/n)	Maternal alcohol consumption 1 = yes; alcohol-consuming group 2 = no; nonalcoholic group
GDM	Gestational diabetes mellitus 1 = no; not having a gestational diabetes mellitus 2 = yes; having a gestational diabetes mellitus
Familial DM	Maternal history of familial diabetes mellitus 1 = no; not having a family history of diabetes mellitus 2 = yes; having a family history of diabetes mellitus
Menarcheal age	Maternal menarcheal age (years)
Menarcheal age ≥16 year (y/n)	Maternal menarcheal age group based on the cut-off of 16 years 1 = yes; menarche at 16 years or older 2 = no; menarche before 16 years
Parity (primi or multi)	Maternal parity 1 = primiparity; first-time mother 2 = multiparity; mother who has given birth one or more times
Menarcheal regularity (y/n)	Maternal menarcheal regularity 1 = yes; regular menarcheal 2 = no; irregular menarcheal
Estimated fetal body weight	estimated fetal body weight (kg)
LGA (y/n)	Large fetus for gestation age 1 = yes; large fetus for gestational age 2 = no; normal-sized fetus for gestational age
S-ScFT for each abdominal site	Skin subcutaneous fat thickness for each abdominal site (cm)
Mean S-ScFT	Mean skin subcutaneous fat thickness of abdominal (cm)

GP, gestation period; C, categorized; DM, diabetes mellitus; GDM, gestational diabetes mellitus; Gr, group; LGA, large fetus for gestation age; S-ScFT, skin subcutaneous fat thickness; y/n, yes or no.

Data pre-processing

Before data preprocessing, the ratio of missing data was examined to ensure the quality of the study data. In total, 38 values were missed out of 10,936 values, and the missing rate was 0.34%, indicating that the data quality was very good. In addition, the correlation between the variables was .9 or higher, indicating no multicollinearity.

Results

General Characteristics of the Participants and Differences in S-ScFT

The differences in S-ScFT according to general characteristics of the participants are presented in Table 2. The participants' mean age was 31.8 ± 4.1 years, and 17.4% were over 35 years of age. The mean prepregnancy BMI was 22.53 ± 3.79 kg/m² (the proportion of obese and underweight women was 19.8%, and 11.0%, respectively). There was a difference in S-ScFT according to the obesity classification (F = 8.97, P < 0.001). Moreover, women who consumed alcohol had significantly thicker S-ScFT than those who did not consume alcohol (t = 1.83, P = 0.034).

Table 2.

Difference of Skin–Subcutaneous Fat Thickness by General Characteristics (N = 172)

Variables	Division		n (%)	S-ScFT (cm) M ± SD	t/F/r	P
Age (years)	19 ≤ ∼ < 35		142 (82.6)	1.33 ± 0.87	0.55	0.291
Age (years)	≥35		30 (17.4)	1.43 ± 0.78
	Mean ± SD	31.8 ± 4.1
	IQR	5
Residential area	City		120 (69.8)	1.29 ± 0.66	1.21	0.115
Residential area	Rural		52 (30.2)	1.47 ± 1.19
Prepregnancy weight (kg)	Mean ± SD	58.95 ± 10.29		1.35 ± 0.86	0.37	<0.001
Prepregnancy BMI (kg/m²⁾	Underweight^a		19 (11.1)	1.04 ± 0.58	8.97	<0.001
	Normal weight^b		91 (52.9)	1.14 ± 0.58	a, b<c<d
	Overweight^c		30 (17.4)	1.59 ± 0.75
	Obesity^d		32 (18.6)	1.90 ± 1.34
	Mean ± SD	22.53 ± 3.79
	IQR	4.58
Alcohol consumption	No		144 (83.7)	1.40 ± 0.90	1.83	0.034
Alcohol consumption	Yes		28 (16.3)	1.08 ± 0.53
Smoking	No		168 (97.7)	1.34 ± 0.85	1.26	0.104
Smoking	Yes		4 (2.3)	1.88 ± 0.90
GDM	No		170 (98.8)		1.14	0.886
GDM	Yes		2 (1.2)		1.14	0.886
Familial DM	No		143 (83.1)	1.32 ± 0.84	1.02	0.155
Familial DM	Yes		29 (16.9)	1.50 ± 0.94
Menarcheal age (years)	<16		153 (89.0)	1.36 ± 0.90	0.60	0.276
Menarcheal age (years)	≥16		19 (11.0)	1.24 ± 0.41
	Mean ± SD	14.1 ± 1.4
Menarcheal regularity	Yes		139 (80.8)	1.34 ± 0.87	0.25	0.403
Menarcheal regularity	No		33 (19.2)	1.38 ± 0.79
Parity	Primiparity		121 (70.3)	1.37 ± 0.91	0.54	0.295
Parity	Multiparity		51 (29.7)	1.29 ± 0.73
Ratio of fetal growth	AGA (50∼ < 90%)		150 (87.7)	1.38 ± 0.90	1.05	0.148
Ratio of fetal growth	LGA (≥90%)		21 (12.3)	1.17 ± 0.42
Gestational period (weeks)	24⁺⁰–27⁺⁶		65 (37.8)	1.32 ± 0.70	1.15	0.331
	28⁺⁰–31⁺⁶		38 (22.1)	1.54 ± 1.54
	32⁺⁰–35⁺⁶		35 (20.3)	1.36 ± 0.86
	36⁺⁰–40		34 (19.8)	1.17 ± 0.53
	Mean ± SD	29⁺⁶ ± 4⁺⁵ weeks (range: 24⁺⁰–38⁺¹ weeks)
	IQR	2
Measured frequency (Total = 354)	1 case		49 (28.5)	49
	2 cases		59 (34.3)	118
	3 cases		53 (30.8)	159
	4 cases		8 (6.4)	32

AGA, appropriated fetus for gestation age; BMI, body mass index; DM, diabetes mellitus; GDM, gestational diabetes mellitus; IQR, interquartile range; LGA, large fetus for gestational age; SD, standard deviation; S-ScFT, skin subcutaneous fat thickness.

S-ScFT by prepregnancy BMI group and GP

The mean S-ScFT of the total abdomen was 1.24 ± 0.77 cm. The higher the prepregnancy BMI, the thicker the S-ScFT. Additionally, there was a decreasing tendency for S-ScFT as the GP increased (Table 3). The percentage of S-ScFT less than 0.6 cm for each abdominal site ranged from 8.5% (site 05) to 23.4% (site 12).

Table 3.

Skin–Subcutaneous Fat Thickness on Each Abdominal Site by Prepregnancy Body Mass Index and Gestational Period (Cases = 354)

Division	Prepregnancy body mass index				Gestational period (weeks)					S-ScFT <0.6 cm n (%)
Division	Underweight (n = 39)	Normal (n = 183)	Overweight (n = 65)	Obesity (n = 67)	24⁺⁰–27⁺⁶ (n = 70)	28⁺⁰–31⁺⁶ (n = 92)	32⁺⁰–35⁺⁶ (n = 105)	36⁺⁰–40 (n = 87)	Total mean	S-ScFT <0.6 cm n (%)
Site 01	0.99 ± 0.53	1.07 ± 0.55	1.33 ± 0.70	1.54 ± 1.02	1.41 ± 0.72	1.33 ± 0.98	1.11 ± 0.52	1.00 ± 0.46	1.20 ± 0.71	47 (13.3)
Site 02	1.11 ± 0.58	1.15 ± 0.56	1.44 ± 0.73	1.78 ± 1.25	1.48 ± 0.76	1.48 ± 1.13	1.23 ± 0.65	1.13 ± 0.49	1.31 ± 0.81	36 (10.2)
Site 03	0.97 ± 0.55	1.08 ± 0.53	1.41 ± 0.72	1.69 ± 1.04	1.39 ± 0.76	1.36 ± 0.92	1.17 ± 0.66	1.09 ± 0.53	1.25 ± 0.73	44 (12.4)
Site 04	0.87 ± 0.60	0.95 ± 0.48	1.21 ± 0.58	1.52 ± 1.06	1.19 ± 0.66	1.23 ± 0.89	1.05 ± 0.66	0.94 ± 0.46	1.12 ± 0.83	60 (16.9)
Site 05	1.03 ± 0.60	1.20 ± 0.53	1.51 ± 0.82	1.83 ± 1.23	1.50 ± 0.79	1.45 ± 1.06	1.31 ± 0.71	1.21 ± 0.59	1.36 ± 0.82	30 (8.5)
Site 06	1.10 ± 0.59	1.21 ± 0.59	1.59 ± 0.83	1.91 ± 1.25	1.50 ± 0.87	1.47 ± 1.06	1.37 ± 0.78	1.29 ± 0.62	1.40 ± 0.85	31 (8.8)
Site 07	0.96 ± 0.55	1.09 ± 0.58	1.45 ± 0.67	1.76 ± 1.43	1.36 ± 0.74	1.36 ± 1.20	1.21 ± 0.78	1.17 ± 0.59	1.26 ± 0.87	52 (14.7)
Site 08	0.87 ± 0.57	1.00 ± 0.51	1.30 ± 0.61	1.70 ± 1.43	1.24 ± 0.69	1.29 ± 1.16	1.13 ± 0.78	1.05 ± 0.53	1.17 ± 0.83	62 (17.5)
Site 09	1.00 ± 0.62	1.10 ± 0.57	1.38 ± 0.58	1.86 ± 1.61	1.26 ± 0.68	1.38 ± 1.41	1.25 ± 0.70	1.24 ± 0.61	1.28 ± 0.92	58 (16.4)
Site 10	1.19 ± 1.72	1.07 ± 0.65	1.38 ± 0.68	1.79 ± 1.56	1.22 ± 0.77	1.37 ± 1.42	1.29 ± 1.18	1.21 ± 0.63	1.27 ± 1.07	54 (15.3)
Site 11	0.93 ± 0.63	1.02 ± 0.59	1.32 ± 0.61	1.70 ± 1.28	1.20 ± 0.71	1.29 ± 1.14	1.15 ± 0.69	1.14 ± 0.63	1.19 ± 0.82	60 (16.9)
Site 12	0.80 ± 0.63	0.88 ± 0.55	1.13 ± 0.48	1.49 ± 1.01	1.01 ± 0.63	1.10 ± 0.92	1.01 ± 0.64	1.00 ± 0.55	1.05 ± 0.76	83 (23.4)
Total mean	0.99 ± 0.57	1.07 ± 0.52	1.37 ± 0.63	1.71 ± 1.23	1.31 ± 0.70	1.34 ± 1.08	1.19 ± 0.66	1.12 ± 0.52	1.24 ± 0.77

Predictive modeling and factor analysis of the S-ScFT of the abdomen based on machine learning

Deep learning, random forest, and support vector machine algorithms were applied for S-ScFT predictive modeling and factor analysis for subcutaneous injection in the abdomen. After the prediction model was developed, the performance and suitability of the algorithms were evaluated. The importance of the variables acting on the performance and prediction of each algorithm in the final model is presented in Table 4. Deep learning had the best prediction performance, followed by random forest and support vector machine.

Table 4.

Performance Evaluation and Critical Variables of Predictive Model (Cases = 354)

Model		Deep learning	Random forest	Support vector machine
Performance evaluation		RMSE (SD)0.2416 (0.1526)	RMSE (SD)0.4163 (0.3700)	RMSE (SD)0.5688 (0.5361)

	Rank feature (Variable)	Weight feature (Variable)	Weight feature (Variable)	Weight
Critical variables	1 Weight during pregnancy	0.2097 Menarcheal age	0.1392 Prepregnancy BMI (C; Gr 4)	0.1333
	2 Menarcheal age	0.1189 Prepregnancy weight	0.1181 GP (C; Gr 4)	0.0991
	3 Prepregnancy weight	0.0952 EFBw	0.1001 Menarcheal age	0.0835
	4 Age ≥35 years (y/n)	0.0849 Weight change during pregnancy	0.0901 Weight during pregnancy	0.0754
	5 Primiparity	0.0584 LGA (y/n)	0.0478 Prepregnancy BMI	0.0578
	6 LGA (y/n)	0.0553 Familial DM (y/n)	0.0337 Alcohol consumption (y/n)	0.0265
	7 Alcohol consumption (y/n)	0.0494 Alcohol consumption (y/n)	0.0274 EFBw	0.0204
	8 Familial DM (y/n)	0.0360 Age ≥35 years ((y/n)	0.0248 LGA (y/n)	0.0169
	9 Prepregnancy BMI (C; Gr 4)	0.0358 Prepregnancy BMI (C; Gr 4)	0.0146 Prepregnancy weight	0.0115
	10 Age	0.0302 GP (C; Gr 4)	0.0105 GP	0.0087

BMI, body mass index; C, Category; DM, diabetes mellitus; EFBw, estimated fetal body weight; Gr, group; GP, gestation period; LGA, large fetus for gestation age; RMSE, root mean squared error; SD, standard deviation; y/n = yes or no.

Prediction model for S-ScFT of the abdomen through deep learning algorithm

Deep learning calculates the value of the next layer by applying an expression consisting of a linear combination of input layers to an activation function and calculates the final output value (output layer) by constructing several layers of such hidden layers. Subsequently, the model is optimized through error backpropagation, which sets weights corresponding to each input flow and adjusts each weight in a direction to reduce the difference between the final output value and the actual value (correct answer).

This study employed a method wherein the learning process of creating a deep learning model is halted if the error does not improve for 10 epochs and 10 iterations. The final model was determined to have two hidden layers. For the prediction of S-ScFT in the abdomen using the deep learning algorithm, weight change during pregnancy was found to have the highest importance. The variables were then ranked in terms of importance, with menarcheal age, prepregnancy weight, age ≥35 years (yes or no), primiparity, large fetus for gestation age (LGA) (y/n), alcohol consumption (y/n), familial diabetes (y/n), and prepregnancy BMI (C; Gr 4), showing high importance, in that order (Table 4).

Prediction model for S-ScFT of the abdomen through random forest algorithm

Random forest is a decision tree-based algorithm that classifies data through a process of multiple questions to obtain a regression result and derives the result using the average value of data corresponding to the final leaf node. In this study, after tuning for the optimized model, the best prediction performance was provided by the case where the number of decision trees was 100, and the maximum depth of the model was seven. The random forest algorithm utilized for predicting S-ScFT in the abdomen, revealed the variable of menarcheal age as having the highest importance. Subsequently, the variables of prepregnancy weight, estimated fetal body weight, weight change during pregnancy, LGA (y/n), familial diabetes (y/n), alcohol consumption (y/n), age ≥35 years (y/n), prepregnancy BMI (C; Gr 4), and GP (C; Gr 4) were found to possess significant importance, in the mentioned order (Table 4).

Prediction model for S-ScFT of the abdomen through support vector machine

Support vector machine is mainly used for data classification but can be similarly applied to regression problems.¹⁹ We applied 210 as the kernel's total number of support vectors and 0.005 as the gamma value of the kernel in the study. In the support vector machine algorithm for S-ScFT prediction of the abdomen, the prepregnancy BMI (C; Gr 4) variable showed the highest importance, and the importance of variables was found to be high in the order of GP (C; Gr 4), menarcheal age, weight during pregnancy, prepregnancy BMI, alcohol consumption (y/n), estimated fetal body weight, LGA (y/n), prepregnancy weight, and GP (Table 4).

Verification for evaluation of clinical applicability of the model

To evaluate the clinical applicability of the abdominal S-ScFT prediction model, the deep learning model, which had the best performance among machine learning algorithms developed in this study, was verified using actual patient data. The data concordance rate, comparing the actual data and predicted value, was 96.7% (Table 5).

Table 5.

The Field Applicability Evaluation of Model for Verification

Classification		Abdominal S-ScFT predictive value (cm)	Abdominal S-ScFT actual value (cm)	Weight during pregnancy (kg)	Menarcheal age (years)	Prepregnancy weight (kg)	≥35 years (y/n) ^a	Parity ^b	LGA (y/n) ^a	Alcohol drinking (y/n) ^a	Familial DM (y/n) ^a	Prepregnancy BMI (C; Gr 4) ^c
Underweight	1	0.96	0.93	54	14	47	1	1	1	2	1	1
	2	0.35	0.31	55	14	45	2	2	1	1	1	1
	3	0.51	0.47	51	14	42	1	1	1	2	1	1
Normal	1	0.59	0.86	67	13	54	1	1	1	1	1	2
	2	0.47	0.62	77	12	68	1	1	1	2	1	2
	3	0.87	1.16	84	16	57	2	1	2	1	1	2
Overweight	1	0.84	0.83	76	13	67	1	1	1	1	2	3
	2	1.07	1.1	70	16	65	2	1	1	1	2	3
	3	1.61	1.54	69	15	60	1	1	1	1	1	3
Obesity	1	1.49	1.24	78	11	72	1	2	1	2	1	4
	2	2.41	2.51	68.5	14	67	1	1	1	1	1	4
	3	0.57	0.57	82	13	81	1	2	1	1	1	4
The concordance rate			96.7%

Classified as “1” for “yes” and “2” for “no”.

Classified as “1” for “primiparity” and “2” for “multiparity”.

Classified as “1” for “underweight,” “2” for “normal weight,” “3” for “overweight,” and “4” for “obese group”.

C, category; DM, diabetes mellitus; Gr, group; LGA, large fetus for gestation age; S-ScFT, skin subcutaneous fat thickness; y/n, yes or no.

Discussion

In this study, we developed the best predictive model for abdominal subcutaneous injection and evaluated its clinical applicability by applying three methods of machine learning with S-ScFT measurement data of normal pregnant women. For predictive performance, the best results were obtained with deep learning, which has been shown to have superior performance than existing machine learning methods in various fields.²⁰

The primary aim of this study was to predict abdominal S-ScFT during pregnancy. The secondary objective was to assess the safety of abdominal subcutaneous injections during pregnancy by comparing the results obtained with different S-ScFT measurements with those obtained with a standard S-ScFT measurement of 0.6 cm and to propose safety guidelines for the administration of these injections. In the case of insulin administration, injecting the needle deeper than the subcutaneous fat leads to a rapid absorption of the drug into the muscle, resulting in hypoglycemia.⁵ Conversely, injecting the needle too shallowly may cause drug leakage, leading to hyperglycemia and other complications.¹⁷

Therefore, the present study was designed on the assumption that the S-ScFT of the abdomen needs to be thicker than 0.6 cm when insulin is injected into the abdomen of pregnant women with a 4–5-mm needle. In the three machine learning methods employed, all derived predictive factors can be judged to be robust variables. Therefore, we discuss these variables by selecting them, focusing on the variables commonly derived from the three types of machine learning: menarcheal age, prepregnancy weight, prepregnancy BMI, LGA (y/n), and alcohol consumption (y/n).

The variable with the highest importance was menarcheal age, indicating that the later the menarcheal age, the thinner the S-ScFT. Therefore, we conducted a multivariate analysis to analyze the different variables that affect S-ScFT. According to a study by Ramezani Tehrani et al.,²¹ in case of high prepregnancy BMI, the menarcheal age was early, and in case of low prepregnancy BMI, the menarcheal age was late. Due to the high correlation between pregnant women's abdominal ScFT and prepregnancy BMI,¹³ the S-ScFT is thinner in case of low prepregnancy BMI. Therefore, it can be inferred that a later menarcheal age is associated with thinner S-ScFT, considering the relationship among menarcheal age, prepregnancy BMI, and S-ScFT.

The variables with the next highest importance on S-ScFT were prepregnancy weight and prepregnancy BMI (C; Gr 4). Eley et al.¹³ previously reported that the correlation between abdominal S-ScFT and prepregnancy BMI of pregnant women was very high. In this study, the mean S-ScFT of the entire abdomen was 1.24 ± 0.77 cm. Lori-Berard et al.⁸ recommended the lateral part of the abdomen as the insulin injection site because the fetus is located in the abdomen's central part. However, Hwang¹¹ reported that the S-ScFT in the lateral part of the abdomen was thinner than that in the central part. Because the obesity standard of Asians is lower compared with Westerners, Western standards for abdominal subcutaneous injections cannot be applied to Asian pregnant women.

This study revealed that the percentage of abdominal S-ScFT measuring below 0.6 cm, which serves as the safety criterion for subcutaneous injections proposed in this research, varied from 8.5% (site 05) to 23.4% (site 12) across the 12 different sites examined. Notably, site 7, commonly employed for abdominal subcutaneous injections during pregnancy, exhibited a 14.7% incidence of S-ScFT below 0.6 cm. These findings underscore the importance of exercising additional caution when administering abdominal subcutaneous injections during pregnancy. Therefore, we strongly recommend that insulin injections be administered at a 90° angle, with a double-fold of abdominal skin, or alternatively, at the arm or thigh as typically practiced, regardless of the needle's length.

The prepregnancy weight ranks the top five variables in both deep learning and random forest models, whereas it falls outside the top five rankings in the support vector machine model. Conversely, prepregnancy BMI (C; Gr 4) holds the first rank in the support vector machine but does not appear within the top five rankings in the deep learning and random forest models. These results show that prepregnancy weight, rather than prepregnancy BMI (C; Gr 4), better reflects the variation in S-ScFT during pregnancy. It is important to note that individuals with the same BMI may exhibit different distributions of S-ScFT.²² Therefore, some researchers prefer to rely on weight alone.²³ Since height remains constant after reaching full growth, prepregnancy weight may serve as a more dependable predictor of S-ScFT compared with BMI. To validate these findings, we recommend conducting future studies in this area.

The LGA (y/n) and GP (C; Gr 4) were also identified as major variables affecting S-ScFT within the top 10 rankings in three types of machine learning. In this study, when the fetuses were larger than GP and when the GP increased, the S-ScFT became thinner. As the fetus comes out of the pelvis from the 12th week of the GP and grows in the abdomen, the circumference of the abdomen increases as much as the fetal size. It is known that, as the fetal size increases, it acts on the tension of the abdomen, and the S-ScFT becomes thinner.¹⁰ In the case of prepregnancy BMI in the underweight group where the fetuses are bigger than GP, special attention is required in injecting insulin, especially in pregnant women who have GDM.

Alcohol consumption (y/n) was also in the top 10 factors in the three types of machine learning. In this study, the S-ScFT in the alcohol-consuming group was very thin compared with the nonalcoholic group. According to a previous study, a nonalcoholic group had a higher prepregnancy BMI and weight gain than an alcohol-consuming group during pregnancy.²⁴ Since alcohol interferes with the absorption of carbohydrates in the body, it is assumed that the prepregnancy BMI of the alcohol-consuming group is lowered, and the S-ScFT is thinner. In addition to S-ScFT, because chronic alcohol consumption during pregnancy causes fetal alcohol syndromes, such as growth and mental retardation, facial deformities, and nervous system malformations in newborns,²⁵ prenatal education emphasizing the importance of alcohol abstinence is necessary.

This study verified the clinical applicability of the deep learning model, which was evaluated as the best predictive model. As a result, the concordance rate between the actual and predicted value was shown to be very good. The deep learning model developed in this study is adequate for predicting S-ScFT. If these variables can be modified and supplemented to develop deep learning-based applications, they could provide significant benefits for managing GDM in pregnant women.

However, this study had some limitations. First, this study did not investigate factors that could affect the measurements such as the amount of amniotic fluid, fetal position, food and activity, hydration, and edema. Additionally, only the left portion of the umbilical line was used to measure S-ScFT. Therefore, the results of this study should be interpreted considering these points. Second, data, such as prepregnancy weight, was not directly measured by the researchers or medical staff but rather obtained through patient self-report during enrollment. Therefore, the reliability of the data cannot be guaranteed.

Third, pregnant women with GDM or pre-existing diabetes who require insulin injections were excluded from this study. Therefore, further research is recommended to include pregnant women who require practical abdominal subcutaneous injections during the gestational period, including those with GDM or pre-existing diabetes. Finally, as this study was conducted on Asian women, specifically in Korea, it may not be applicable to other ethnicities. Therefore, guidelines for safe abdominal subcutaneous injections during pregnancy should be developed by conducting research targeting diverse population groups, considering ethnic diversity.

Conclusions

This study used three different machine learning algorithms to develop a predictive model for the safety of abdominal subcutaneous injection during pregnancy. The algorithms showed excellent performance in the order of deep learning, random forest, and support vector machine. Prepregnancy weight, prepregnancy BMI, LGA, menarcheal age, and alcohol consumption were identified as predicting variables by all three algorithms. The concordance rate of the developed model was 96.7%, indicating its validity. A short needle may be used for overweight pregnant women to inject insulin throughout the abdomen; however, insulin injection by creating a skin fold in the abdomen is recommended in normal and low BMI pregnant women. Since the S-ScFT of the abdomen may vary depending on the condition of the fetus, information on the site in the abdomen where insulin injection is possible should be provided through prenatal ultrasonography.

Thus, this study provides valuable information for the personalized care of pregnant women with GDM and contributes to improving nursing care quality.

Ethical Conduct of Research

This study was conducted in accordance with ethical principles and guidelines for human subjects' research. This study was approved by the Institutional Review Board (IRB) at Woosuk University (IRB number: WS-2020-12). Informed consent was obtained from all participants before their participation in the study. All participants' rights, welfare, and confidentiality were protected throughout the study.

Footnotes

Acknowledgment

The authors would like to thank Dr. Seo, Sung Jin, and Dr. Jung, and Nam Ok of the Hanbyal Women's Hospital in Jeonju, for their support and contributions, as well as for the pregnant women who participated in this study.

Authors' Contributions

M.S.H. conceptualized the study, was involved in data collection, investigation, analysis, validation, visualization, writing original draft, funding acquisition, and project administration. E.S. participated in data collection, investigation, analysis, software, and writing original draft. J.A. participated in validation, visualization, and writing original draft and review. S.P. participated in investigation, validation, visualization, and writing original draft and review.

Author Disclosure Statement

The authors have no conflicts of interest to report.

Funding Information

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2020R1F1A105SS7511).

References

Saeedi

, Petersohn

, Salpea

, et al. Global and regional diabetes prevalence estimates for 2019 and projections for 2030 and 2045: Results from the International Diabetes Federation Diabetes Atlas, 9th edition. Diabetes Res Clin Pract, 2019; 157:107843; doi: 10.1016/j.diabres.2019.107843

Kim

. Diagnosis and medical management of gestational diabetes. J Korean Diabetes, 2015; 16(3):174–181; doi: 10.4093/jkd.2015.16.3.174

Korean Society of Obstetrics and Gynecology. Obstetrics. 6th ed. Koonja Publishing, Inc.: Paju; 2019.

Farahvar

, Walfisch

, Sheiner

. Gestational diabetes risk factors and long-term consequences for both mother and offspring: A literature review. Expert Rev Endocrinol Metab, 2019; 14(1):63–74; doi: 10.1080/17446651.2018.1476135

Han

, Lim

, Oh

, et al. Needs for development of IT-based nutritional management program for women with gestational diabetes mellitus. Korean J Community Nutr, 2017; 22(3):207–217; doi: 10.5720/kjcn.2017.22.3.207

Brawerman

, Dolinsky

. Therapies for gestational diabetes and their implications for maternal and offspring health: Evidence from human and animal studies. Pharmacol Res, 2018; 130:52–73; doi: 10.1016/j.phrs.2018.02.002

Carson

, Henderson

, King

, et al. Perceptions and concerns regarding diabetes mellitus during pregnancy among American Indian women. Care Manag J, 2014; 15(4):160–169; doi: 10.1891/1521-0987.15.4.160

Lori-Berard

, Donna

, Gail

, et al. Recommendations for Best Practice in Injection Technique. FIT Forum for Injection Technique: Canada; 2017:27.

World Health Organization. Obesity: Preventing and Managing the Global Epidemic. Report of a WHO Consultation. World Health Organ Tech Rep Ser, 2000; 894:i-xii, 1–253.

10.

Seo

, Lee

, Kim

, et al. 2018 Korean society for the study of obesity guideline for the management of obesity in Korea. J Obes Metab Syndr, 2019; 28(1):40–45; doi: 10.7570/jomes.2019.28.1.40

11.

Hwang

. Abdominal skin subcutaneous fat thickness over the gestational period in Korean pregnant women: A descriptive observational study. Korean J Women Health Nurs, 2021; 27(4):318–325; doi: 10.4069/kjwhn.2021.12.12

12.

Kennedy

, Quinton

, Brown

, et al. Changes in maternal abdominal subcutaneous fat layers using ultrasound: A longitudinal study. Obes Res Clin Pract, 2017; 11(6):655–664; doi: 10.1016/j.orcp.2017.10.003

13.

Eley

, Sekar

, Chin

, et al. Increased maternal abdominal subcutaneous fat thickness and body mass index are associated with increased cesarean delivery: A prospective cohort study. Acta Obstet Gynecol Scand, 2019; 98(2):196–204; doi: 10.1111/aogs.13486

14.

Li-Zhen

, Yun

, Xiao-Dong

, et al. Evaluation of guidelines on the screening and diagnosis of gestational diabetes mellitus: systematic review. BMJ Open, 2019; 9(5):e023014; doi: 10.1136/bmjopen-2018-023014

15.

Gibney

, Arce

, Byron

, et al. Skin and subcutaneous adipose layer thickness in adults with diabetes at sites used for insulin injections: Implications for needle length recommendations. Curr Med Res Opin, 2010; 26(6):1519–1530; doi: 10.1185/03007995.2010.481203

16.

Derraik

, Gusso

, Peart

, et al. Preliminary data on dermis and subcutis thickness in adults with type 1 and 2 diabetes mellitus. Australas J Dermatol, 2015; 56(4):e93–e95; doi: 10.1111/ajd.12177

17.

Sim

, Hwang

, Kim

, et al. The appropriateness of the length of insulin needles based on determination of skin and subcutaneous fat thickness in the abdomen and upper arm in patients with type 2 diabetes. Diabetes Metab J, 2014; 38(2):120–133; doi: 10.4093/dmj.2014.38.2.120

18.

Woldaregay

, Årsand

, Walderhaug

, et al. Data-driven modeling and prediction of blood glucose dynamics: Machine learning applications in type 1 diabetes. Artif Intell Med, 2019; 98:109–134; doi: 10.1016/j.artmed.2019.07.007

19.

Smola

, Schölkopf

. A tutorial on support vector regression. Stat Comput, 2004; 14(3):199–222; doi: 10.1023/B:STCO.0000035301.49549.88

20.

Kamilaris

, Prenafeta-Boldú

. Deep learning in agriculture: A survey. Comput Electron Agric, 2018; 147:70–90; doi: 10.1016/j.compag.2018.02.016

21.

Ramezani Tehrani

, Mirmiran

, Gholami

, et al. Factors influencing menarcheal age: Results from the cohort of Tehran lipid and glucose study. Int J Endocrinol Metab, 2014; 12(3):e16130; doi: 10.5812/ijem.16130

22.

Greco

, Mallio

. Artificial intelligence and abdominal adipose tissue analysis: A literature review. Quant Imaging Med Surg, 2021; 11(10):4461–4474; doi: 10.2103/7/qims-21-370

23.

Sengeis

, Müller

, Störchle

, et al. Body weight and subcutaneous fat patterning in elite judokas. Scand J Med Sci Sports, 2019; 29(11):1774–1788; doi: 10.1111/sms.13508

24.

Kim

, Lee

, Shin

, et al. How drinking during pregnancy affects women's pregnancy weight gain and neonatal birth weights. J Korean Acad Addict Psychiatry, 2011; 15(1):16–21.

25.

Seoul National University Hospital. Fetal alcohol syndrome. SNUH; 2017. Available from: https://www.snuh.org/health/nMedInfo/nView.do?category=DIS&medid=AA000477 [Last accessed: December 02, 2022 ].