Predicting Neonatal Hypoglycemia Using AI Neural Networks in Infants from Mothers with Gestational Diabetes Mellitus

Introduction

Gestational diabetes mellitus (GDM) is a common complication during pregnancy that causes maternal hyperglycemia [1]. Because glucose can easily cross the placenta, maternal hyperglycemia resulting from GDM causes elevated glucose levels in the fetus. This increase in glucose stimulates the fetus to produce excessive amounts of insulin, a condition known as hyperinsulinism [2]. After birth, when the supply of maternal glucose is discontinued, the condition of hyperinsulinemia in the newborn can continue for 1 to 2 days after delivery [2,3]. Studies show that up to 48% of newborns from mothers with GDM can experience hypoglycemia, potentially causing severe complications, such as brain damage, seizures, and respiratory issues [4,5]. American Diabetes Association guidelines recommend early screening and intervention to manage and prevent these risks effectively [6]. Current practices by the American College of Obstetricians and Gynecologists underscore the importance of maternal health monitoring throughout pregnancy and postpartum blood glucose testing for neonates as standard care, to mitigate the complications, including hypoglycemia [7].

In recent years, several studies have focused on identifying the risk factors for neonatal hypoglycemia and developing effective tools to predict its occurrence. Yan et al suggested that in mothers with normal weight and type 2 diabetes mellitus, reducing weight gain during pregnancy could lower the risk of neonatal hypoglycemia [8]. Additionally, Wang et al found that the blood glucose level in umbilical artery cord blood can accurately predict neonatal hypoglycemia within 2 h after birth [9]. Cao et al reported that factors such as maternal overweight or obesity before pregnancy and poor blood glucose control in pregnant women with GDM are all risk factors for neonatal hypoglycemia [10]. Another study involving pregnant mothers with and without GDM showed that advanced maternal age, extended fasting periods before delivery, GDM, fetal distress diagnosis, and an increase in neonatal body mass index (BMI) all increase the likelihood of developing neonatal hypoglycemia [11]. Moreover, some studies have identified gestational age and preterm delivery as independent predictors of neonatal hypoglycemia [12,13]. However, there remains a lack of effective tools to predict a neonate’s blood glucose level based on clinical variables [14].

Recent studies have shown that machine learning outperforms traditional analysis methods, which often depend on feature engineering [15–17]. A key benefit of machine learning is its ability to learn optimal features autonomously, enabling the creation of personalized models for each individual [15]. In recent years, there have been reports of studies using deep learning methods to predict blood glucose concentrations, including recursive neural networks (RNNs) [18] and artificial neural networks (ANNs) [19]. For instance, a study by Li et al introduced a deep learning algorithm for glucose prediction using a multi-layer convolutional recurrent neural network, demonstrating high accuracy in real patient scenarios [18].

In the present study, we used a neural network to develop a new forecasting model to predict the blood glucose level of neonates from mothers with GDM during the first 24 h of life, and then evaluated its effectiveness on real patients’ data. In this study, we aimed to develop a predictive model for the association between maternal and neonatal anthropometric data and neonatal hypoglycemia based on data from 106 mothers with GDM and their neonates.

Material and Methods

ETHICS STATEMENT:

The Ethics Committee of Shijiazhuang Obstetrics and Gynecology Hospital reviewed and approved this study (No. 20230165). The research was conducted in compliance with the Declaration of Helsinki version 2008 and the laws of China. Prior to participation, all the involved mothers and the parents of all neonates reviewed and signed the consent forms. The consent covered aspects such as medical procedures and the publication of study findings in various formats, including the use of personal patient data, without any limitations regarding time or language. The clinical trial associated with the study has been registered under No. ChiCTR1800018488.

This single-center prospective study was conducted at our center between February 2023 and December 2023 (Figure 1). All the enrolled pregnant women had GDM and had delivered a baby.

INCLUSION AND EXCLUSION CRITERIA:

The following inclusion criteria were established: (1) mothers with GDM according to the criteria of the World Health Organization/International Association of Diabetes and Pregnancy Study Groups; (2) term babies (≥37 completed weeks); (3) admission for delivery during the study period; (4) prenatal care was received at our institution; and (5) written consent forms were signed and obtained.

The exclusion criteria included (1) failure to provide written consent; (2) multiple pregnancies; (3) mothers who did not receive prenatal care at our clinic; (4) major congenital abnormalities (including congenital heart defects, neural tube defects, cleft lip and palate, Down syndrome, fragile X syndrome, cystic fibrosis, phenylketonuria, congenital diaphragmatic hernias, hydronephrosis); (5) required immediate Neonatal Intensive Care Unit admission for reasons other than hypoglycemia, such as neonatal sepsis; (6) beta-blocker therapy during pregnancy; (7) fetal growth restriction; and (8) stillbirths.

DATA COLLECTION:

Clinical variables, such as demographic characteristics, were collected from the recruited pregnant women and their neonates, including the demographic data of the pregnant women, such as pregnancy weight before delivery, BMI, gestational age, and mode of delivery; medical history of the pregnant women, such as glycosylated hemoglobin (HbA1c) at trimester and blood glucose concentration at labor; and clinical variables of the neonates, such as sex, firstborn, gestational age, Apgar score, birth weight, birth height, malformation, and blood glucose level. The researchers involved in this study collected data and then performed the analysis.

PRENATAL CARE:

At our institution, prenatal care for GDM was provided by a specialized endocrinologist with expertise in managing diabetes during pregnancy. The care protocol included (1) medical nutrition therapy: pregnant women with GDM received guidance and support regarding their dietary choices to maintain optimal blood glucose levels; (2) self-monitoring of capillary blood glucose: women were advised to regularly measure their capillary blood glucose levels at home using a glucose meter to monitor their blood glucose levels; (3) insulin therapy: if more than 20% of capillary blood glucose readings fell outside the target range, insulin therapy was initiated to help regulate blood glucose levels effectively; and (4) HbA1c determination: HbA1c tests were conducted to assess long-term blood glucose control during pregnancy.

During labor and delivery, maternal capillary blood glucose levels were monitored every 1 to 2 h. Peripartum glycemic control was maintained within a target range of 3.8 to 7.2 mmol/L. Intravenous insulin infusion was initiated for mothers whose capillary blood glucose levels exceeded 7.2 mmol/L. Maternal hypoglycemia is defined as a capillary blood glucose value below 3.8 mmol/L, while maternal hyperglycemia is defined as capillary blood glucose above 7.2 mmol/L.

NEONATAL CARE:

The infants had their blood glucose levels monitored during the first 24 h after birth, specifically at 0 h (from the umbilical cord), 0.5 h, 1 h, 3 h, and 24 h. Infants stayed with their mothers unless the newborns had hypoglycemia or other clinical issues, such as infant respiratory distress syndrome, for which infants were taken to the specialized neonatal care unit for intensive care. They were prescribed early feeding on demand (breast milk or formula). Capillary blood samples for glucose measurements were obtained through a heel-prick. In infants with low heel-prick blood glucose, a capillary blood test was performed to confirm the blood glucose value.

In the present study, the latest diagnostic criteria for neonatal hypoglycemia were defined as a threshold of 2.5 mmol/L, regardless of the presence of signs of neonatal hypoglycemia [20]. For neonates diagnosed with hypoglycemia, an adequate supply of human milk was ensured, along with the infant’s ability to nurse or bottle-feed. A regular feeding schedule was established, typically every 2 to 3 h, either via breastfeeding or formula. If hypoglycemia persisted, intravenous (i.v.) fluids containing dextrose and glucose were administered.

MODEL TRAINING AND TESTING:

All neural network analyses were performed using the Python Keras framework software (Python version 3.12.3, developed by the Python Software Foundation, Keras version 3.3.3), and the Neural Network Toolbox was used to develop the algorithm.

We divided our experimental database (n=85) into a learning set (80%) and a validation set (n=21, 20%). The dataset was trained using ANN and RNN. These networks were trained using different architectures, exploring various depths and widths (Figure 2).

We applied the early stopping functions in the ANN and RNN models. To achieve the best model, we initiated the hidden layer with a single neuron and continued until the root mean square error (RMSE) reached a stable point. We also conducted a statistical test (examining the slope and intercept) for validation and took measures to prevent overfitting.

In the ANN model, the Levenberg-Marquardt algorithm was used during the learning phase. The weights and biases were adjusted, aiming to achieve a reduction in RMSE. The Levenberg-Marquardt algorithm is known for its effectiveness in reducing the RMSE and improving the overall accuracy of the neural network models. To determine the optimal number of training epochs, training epochs with the highest test accuracy for the model were selected (Figure 2A). The activation function used for the hidden layers was the rectified linear unit. During the training process, mean square error (MSE) served as the chosen loss function. The optimization algorithm was Adam, with a learning rate of 0.001. The model uses full-batch gradient descent for training, with a maximum of 100 iterations (epochs). L2 regularization was applied with an alpha value of 0.0001 to prevent overfitting.

The RNN model incorporated both convolutional and long short-term memory algorithms during the learning phase. We conducted experiments involving various configurations of both layers (hidden and dense) to explore different configurations to find the optimal combination for their glucose prediction task (Figure 2B). We found that adjustments could initially improve prediction accuracy, but after a certain point, they could lead to overfitting. This finding suggested that a depth of around 50 units in hidden layers was optimal for the RNN in its glucose prediction task while also minimizing computational cost. In terms of initialization, the kernel initializer for all dense layers was set to RandomNormal. The activation function used for the hidden layers was the rectified linear unit. The loss function was MSE. The optimizer was Adam, with a learning rate of 0.001. The model used full-batch gradient descent for training, with a maximum of 500 iterations (epochs). L2 regularization was applied with an alpha value of 0.0001 to prevent overfitting.

In addition, we used 3 different statistical models – random forest, multiple linear regression, and support vector regression – to create models based on the input variables to predict the lowest blood glucose levels of neonates in the first 24 h after birth. The data were split into training and testing sets, as 80%: 20%. All analyses were performed using the Python Scikit-Learn framework (version 1.4.2). In the random forest analysis, number of decision trees (ntree)=500, and mtry=9 (random forest model)/7 in the variable-restricted random forest model. The minimum samples per leaf was set to 1. The feature selection method was set to “auto”, which uses the square root of the number of features. Bootstrap sampling was enabled, and the random state was set to 42 for reproducibility. The support vector regression model used the Radial Basis Function kernel with a regularization parameter set to 1.0, an epsilon parameter of 0.1, and the gamma parameter set to “scale”. The data were standardized using StandardScaler to ensure each feature had a mean of 0 and a standard deviation of 1. The loss function used was Quadratic Loss. The model training used default stopping criteria, either reaching the maximum number of iterations or achieving convergence.

MODEL EVALUATION METRICS:

By using these different evaluation metrics, a comprehensive understanding of the model’s performance from various angles was obtained. Mean absolute error and RMSE provided information about the absolute prediction error; the coefficient of determination (R2 score) assessed the explanatory power of the model. To estimate prediction accuracy, Clarke error grid analysis was used. This multidimensional evaluation helps in assessing the strengths and weaknesses of the regression models in predicting blood glucose levels. To assess the accuracy of all network models in predicting blood glucose, we used the following metrics [21–24]: mean absolute error, RMSE, R2 score, error grid analysis, and mean absolute percentage error.

STATISTICAL ANALYSIS:

To assess the performance of the ANN model, linear regression analysis was conducted using experimental values and ANN-predicted values from the validation dataset. The data preparation involved obtaining real blood glucose values from the validation dataset (experimental values) and blood glucose values predicted by the ANN model for the same dataset (predicted values). The linear regression model was constructed with the experimental values as the independent variable (X) and the predicted values as the dependent variable (Y). The performance metrics, R² and RMSE, were computed to reflect the proportion of variance in the dependent variable predictable from the independent variable and to measure the average magnitude of the errors between predicted and experimental values, respectively. To validate the regression model, a scatter plot of experimental values vs predicted values was created, overlaid with the regression line to visualize the model’s fit.

The variables are presented as follows: discrete variables are expressed in counts (percentages); normally distributed continuous variables are represented as min, max, and mean±standard deviation (SD); and skewed continuous variables are shown with min, max, and a 95% confidence interval (CI).

Results

DEMOGRAPHIC DATA AND CLINICAL VARIABLES:

From February 2023 to December 2023, a total of 114 pregnant women with GDM were included in this study; however, 8 cases were excluded from the final analysis. In these excluded 8 cases, 2 patients withdrew from the study, 1 patient did not complete perinatal healthcare, 2 patients had twins, 1 patient took beta-blockers during pregnancy, 1 neonate had congenital abnormalities, and 1 patient had a stillbirth. Therefore, our study included 106 pregnant women with GDM and their eventual neonates.

Table 1 displays the demographic characteristics of all study participants. Ninety-one mothers had obesity (85%), of which 38 mothers received medicine to treat GDM (42%). Seventy-eight percent of mothers had a vaginal delivery. One mother had pre-eclampsia.

The median gestational age at delivery was 38.49 weeks, and the average birth weight of the neonates was 3335.83 g. Ninety-three percent of babies received breastmilk for their first feeding. Blood glucose levels were measured in all 106 babies before their second feeding.

PROPOSED ANN MODEL:

In this study, an ANN model was used to predict blood glucose concentrations (Figure 2). Eighty percent of the data was allocated for training the network model, while the remaining 20% was reserved for validation. The models were constructed using maternal and neonatal variables as inputs – age at pregnancy, pre-delivery weight, maternal height, BMI, trimester HbA1c, gestational age, pregnancy weight gain, labor blood glucose concentration, parity, total labor time, and deliver mode (natural delivery, midwifery, or cesarean delivery), along with neonatal sex, Apgar score, height, weight, and Ponderal index – each of which was integrated into a general network. The output unit of the network corresponded to the prediction of the lowest neonate blood glucose concentrations within the initial 24 h of life.

In determining the optimal neural network architecture for predicting blood glucose concentrations, an iterative process was used. This process involved conducting 8400 runs with 100 iterations for each neuron in the hidden layer, ranging from 1 to 8 neurons. The goal was to identify the network topology that yielded the best performance for blood glucose concentrations.

VALIDATION OF THE ANN MODEL:

Overall, the study used an ANN model with specific architecture to predict the lowest neonate blood glucose concentrations in the first 24 h of life. Based on the experimentation, the optimal network architecture was found to be 16-8-1 for blood glucose concentrations. This architecture signifies the 16 neurons in the input layer, 8 neurons in the hidden layer, and 1 neuron in the output layer. The regression coefficients (R²) and RMSE were found to be 0.869 and 0.274, respectively, for the lowest neonate blood glucose values within the initial 24 h of life.

Figure 3 displays the comparison between the experimentally measured and predicted values of the lowest neonate blood glucose concentrations for the testing database. This provides an overview of the performance of the ANN model. To further assess the relationship between the actual and ANN-predicted values, a linear regression model was used. The regression equation used was as follows: ANN predicted values=−0.31+1.11 actual values, with an R2 of 0.726 and an RMSE of 0.348.

Furthermore, we trained various models, including multiple linear regression, random forest, and support vector regression, to predict the lowest neonatal blood glucose concentrations within the first 24 h of life. We calculated R2, MSE, and RMSE to evaluate their performance. As detailed in Table 2, the ANN model demonstrated superior performance, compared with the other models, with a high R2 score and relatively low MSE and RMSE.

ANN PERINATAL MODEL SENSITIVITY EXAMINATION:

The research entailed a methodical assessment of neural network models, focusing on the analysis of the matrix of weights and applying Garson’s formula. This approach was used to qualitatively evaluate the impact of various input factors on the predicted hypoglycemia in neonates during the initial 24 h after birth. Garson’s formula provides a measure of the relative importance of each input variable in influencing the output.

Figure 4 displays the comparative significance of the determined input factors for neonatal hypoglycemia. The graph indicates that every variable significantly influenced hypoglycemia in neonates during the initial 24 h of life. Specifically, the sensitivity analysis revealed that neonatal birth weight had the highest influence, accounting for 18.6% of the variability in hypoglycemia. Maternal BMI followed, with a contribution of 15.9%, while maternal weight gain during pregnancy had an influence of 14.8%. Additionally, maternal weight before delivery had an influence of 11.6%, maternal HbA1c during the third trimester contributed 9.0%, and neonatal Ponderal index contributed 8.8%. On the other hand, factors such as maternal morbidity (gestational age, age at pregnancy, maternal height, total labor time) and neonatal clinical data (sex, Apgar score) had relatively lesser importance, accounting for 7.4% to 1.0%, respectively. These findings highlight the significant role of maternal and neonatal factors, particularly neonatal birth weight, maternal BMI, maternal weight gain during pregnancy, and maternal weight before delivery, in determining the levels of neonatal blood glucose concentration.

Overall, the ANN model for perinatal studies effectively forecasted the experimental outcomes, specifically the predicted hypoglycemia and blood glucose concentrations in neonates within the first 24 h of life, using neonatal and maternal anthropometric data. These results demonstrated the importance of various maternal parameters in predicting neonate blood glucose concentrations.

RNN MODEL PERFORMANCE AND VALIDATION:

To solve the problem of accurately predicting neonate blood glucose concentrations during the initial 24 h of life, we used an RNN model with a multitask learning approach. This model predicts neonate blood glucose concentrations based on various maternal and neonate parameters, including maternal age at pregnancy, pre-delivery weight, maternal height, BMI, trimester HbA1c, gestational age, pregnancy weight gain, labor blood glucose concentration, parity, total labor time, and deliver mode (natural delivery, midwifery, or cesarean delivery), along with neonatal sex, Apgar score, height, weight, and Ponderal index, and the varying levels of glucose in the neonate’s blood at different times throughout the first day of life (from immediately after birth to 24 h after birth).

Our RNN model’s forecasts consistently fell within an acceptable margin, indicating the model’s resilience in handling the intersubject variability of glucose dynamics across both datasets. Our model was particularly adept at identifying blood glucose concentration trend changes by learning the complex interactions and dynamics within the dataset. The RNN structure used memory cells to analyze past time series data. It demonstrated performance, with an RMSE of 0.63 mmol/L and an mean absolute error of 0.53 mmol/L. The mean absolute percentage error was computed as 12.9%. The accuracy can be further optimized by training with a larger dataset. The case with high-frequency fluctuation had poor performance, with an RMSE of 1.38 mmol/L, and the optimal case, characterized by low-frequency fluctuation, achieved the highest performance, exhibiting an RMSE of 0.02 mmol/L.

We conducted an examination of the errors in the forecasted values, as depicted in Figure 5. The analysis of errors adhered to the guidelines of the Clarke error grid lines. This figure validates the excellent clinical correlation between the observed and estimated glucose values. Most of the predicted values fell within zone A. Some pairs of points were in zone B, indicating minor overestimations or underestimations. Very few points were in the erroneous zone D. Notably, no points were in zone E, which represents a zone of significant error.

Discussion

In the present study, an ANN model was created to predict the lowest level of neonatal blood glucose and hypoglycemia during the initial 24 h of life. The ANN model, based on inputs from maternal and neonatal clinical data, predicted blood glucose levels with a high degree of accuracy. Neonatal birth weight and maternal BMI were the 2 most significant factors in predicting hypoglycemia. Moreover, we used an RNN model to predict the multiple time points of neonatal blood glucose levels during the initial 24 h of life. The RNN model forecasted glucose levels effectively, addressing the dynamic changes in blood glucose. Our model investigated the correlations between various variables, learned their causal effects, and displayed superior performance in forecasting neonatal blood glucose concentrations.

Our data generated from the ANN model demonstrated strong concordance between actual values and ANN-predicted values, displaying a strong correlation and minimal error. Furthermore, by inputting data from mothers and newborns into the ANN model, we were able to effectively predict hypoglycemia in newborns during the first 24 h after birth. Therefore, the ANN model provided a means to predict hypoglycemia and the lowest blood glucose levels during the initial 24 h of life, solely based on anthropometric data, offering a practical approach for clinical applications. The ANN model not only established mathematical models for estimating neonatal blood glucose concentrations based on anthropometric data but also identified key maternal and neonatal variables that were biologically relevant for predicting these values. Sensitivity analysis of the ANN perinatal model revealed that neonatal birth weight and maternal BMI were the most influential parameters in predicting hypoglycemia in the first 24 h of life, while maternal metabolic health played a significant role in simulating neonatal blood glucose levels during the initial 24 h.

Previous studies have explored the risk factors for hypoglycemia in neonates born to mothers with GDM, indicating that multiple maternal and neonatal factors, such as gestational age, weight gain during pregnancy, and maternal overweight, increase the risk of neonatal hypoglycemia. These findings are consistent with our data. Maternal BMI and weight gain are also critical factors, as identified in our analysis and other studies [8,10]. Previous research primarily used traditional statistical methods, such as logistic regression analysis by Yan et al [8] and Cao et al [10], to identify risk factors. However, our present study included comprehensive variable inclusion, advanced predictive modeling, iterative optimization, performance comparison, and quantitative sensitivity analysis. Unlike previous studies focusing on specific factors, we incorporated a broader range of maternal and neonatal variables for a more comprehensive analysis. Neonatal birth weight is a significant contributor to hypoglycemia, aligning with the study by Wu et al, which links increased neonatal BMI to a higher probability of hypoglycemia [11]. Newborns with a high BMI often have higher glycogen reserves, but due to elevated insulin levels, these stores can deplete rapidly, causing blood glucose levels to drop.

The use of ANN allowed us to capture complex, non-linear relationships between variables, enhancing predictive accuracy. The iterative process to determine the optimal network architecture ensured the best performance in predicting blood glucose concentrations. By comparing the ANN with multiple linear regression, random forest, and support vector regression models, we validated the superior performance of the ANN, providing a robust predictive tool. The application of Garson’s formula offered a detailed understanding of each input variable’s relative importance, valuable for clinical decision making. Thus, this ANN model, incorporating a wide range of variables and using sophisticated modeling techniques, resulted in superior predictive performance, compared with traditional models. The ANN was chosen because it efficiently handles independent and identically distributed data, capturing static relationships between different features to provide accurate risk predictions for hypoglycemia. The ANNs are suitable for static feature extraction and classification problems. Predicting hypoglycemia risk involves multiple independent risk factors, and the ANN, with its multilayer structure, can effectively learn and integrate these factors, to make accurate classifications. The ANN’s characteristics make it superior in handling independent variables and performing classification tasks, which is ideal for static data analysis and risk assessment.

In addition, both our study and a study by Wang et al [9] used the blood glucose level of umbilical artery cord blood to predict neonatal hypoglycemia. However, while Wang et al used only correlation analysis to investigate the relationship between umbilical artery cord blood glucose levels and neonatal hypoglycemia within 2 h after birth, we used an RNN model to predict blood glucose level changes at multiple time points during the first 24 h after birth, based on maternal and neonatal clinical variables. Our RNN model forecasted dynamic changes in neonatal blood glucose within the first 24 h. Using an RNN model with multitask learning captures complex, non-linear relationships between variables, enhancing predictive accuracy over an extended period (24 h). Moreover, the RNN model demonstrated robust performance with low RMSE and mean absolute error, and detailed error analysis confirmed its reliability in clinical settings. A significant challenge for clinicians is predicting the specific times at which hypoglycemia might occur within those 24 h [25]. To address this, we introduced the RNN model to predict fluctuations in neonatal blood glucose levels during the first 24 h of life. This approach allows clinicians to predict changes in blood glucose over the subsequent 24 h, using clinical indicators from mothers and newborns. Our research has shown the superior performance of our RNN model in predicting changes in blood glucose concentration within the first 24 h of life. The advantages of these models are due to an intricate architecture composed of convolutional and long short-term memory layer stacks, facilitating superior learning of complex features in multivariate datasets.

The RNN was chosen because it can model temporal dependencies in time series data, considering previous time points to provide more accurate future blood glucose level predictions. The RNN is designed for handling and predicting time series data. Since predicting future blood glucose levels is a time series prediction problem, the RNN can leverage its recurrent neurons and feedback mechanism to capture complex patterns and dependencies in the time series, providing reliable predictions. The RNN’s design makes it excellent for time series prediction tasks, effectively capturing temporal dependencies and being suitable for dynamic data analysis and future value prediction.

These findings align with the proposed mechanism of blood glucose regulation during pregnancy. Previous studies have revealed that the metabolism-related health of pregnant women, such as weight, BMI, and GDM, is crucial in determining fetal insulin levels [8–10,26]. Neonates exposed to GDM have higher insulin levels than do non-exposed neonates, which could then result in hypoglycemia [27,28]. Fetal insulin is produced by the fetus’s own pancreas. Insulin from the mother does not traverse the blood-placental barrier. The high insulin levels observed in neonates of mothers with GDM could be an adaptive response to increased levels of blood glucose in the mother [29,30].

Currently, in clinical practice, there is no widely recognized and universally applicable “formula” or model for predicting neonatal hypoglycemia risks in infants of mothers with GDM, due to the complexity and multiple variable factors. However, if basic predictors, such as neonatal birth weight, maternal BMI, maternal HbA1C, and maternal gestational age, can be integrated to develop a proven and effective model, it could provide a quick and effective method for initial assessments at birth. Therefore, we recognize the importance of simplicity and practicality in clinical applications. We plan to collect more data, optimize our algorithms, refine our complex models, and explore the development of a simplified model that maintains reasonable accuracy while being user-friendly. This approach will allow us to cater to a broader range of clinical needs and settings, ensuring that our research benefits the widest possible audience.

The study has several limitations that should be acknowledged. One limitation is the small sample size, which can affect the generalizability of the findings. Although adaptive learning algorithms such as ANNs and RNNs can partially overcome this limitation by using a portion of the database for training, increasing the sample size in future simulations could help improve performance. ANNs and RNNs require large amounts of labeled data and significant computational resources for training. They can also struggle with interpretability, making it difficult to understand the reasoning behind their decisions. RNNs face their own set of challenges, such as vanishing and exploding gradients, which can complicate training, especially for long sequences. This often necessitates using advanced variants, such as long short-term memory networks or gated recurrent units, to mitigate these issues. Furthermore, RNNs can be computationally intensive and slower to train, due to their sequential nature, making them less efficient for large-scale applications than architectures like Transformers. In the RNN model, to predict neonatal blood glucose concentrations, the prediction of glucose was anticipated to be reactive to hypoglycemia (low levels), yet our model showed a lack of sensitivity toward hyperglycemia (high levels). Therefore, this algorithm could be too sensitive to hypoglycemia in practical settings, potentially neglecting high glucose values when applied. In our subsequent research, we intend to explore solutions for potential inaccuracies in hyperglycemia detection. This model needs to be updated to identify frequent high variations. Additionally, we highlight the need to collect more information regarding insulin doses and food intake during pregnancy in future research, as these factors could significantly impact glucose levels.

Conclusions

In summary, neonatal birth weight and maternal BMI were critical predictors of hypoglycemia of neonates in the first 24 h of life. The ANN and RNN models developed in this study demonstrated promising accuracy and low error percentages. They also effectively predicted neonatal hypoglycemia in neonates of mothers with GDM, making them practical for monitoring newborn blood glucose levels and assessing the risk of hypoglycemia in neonates.