Development of a machine learning model to identify the predictors of the neonatal intensive care unit admission

Abstract

Scientists aim to create a system that can predict the likelihood of newborns being admitted to the neonatal intensive care unit (NICU) by combining various statistical methods. This prediction could potentially reduce the negative health outcomes, deaths, and medical costs associated with NICU stays by detecting potential cases early on. This study utilized a retrospective cohort design. The primary outcome of the research focused on admissions to the NICU. The real-time data of pregnant women with a cephalic presentation who gave birth between January 2020 and December 2022 were extracted from the electronic health records of Khaleej-e-Fars Hospital in Bandar Abbas, Iran. The first step of the analysis involved comparing healthy babies to those admitted to the NICU. Variables that had a significant p-value (less than 0.05) were selected as features for the machine learning approach. The input data were utilized to train nine different machine learning models. In our assessment, we used the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1- score to evaluate the effectiveness. During the study period, the rate of NICU admission at our center was 477 out of 3,062 deliveries (15.5%). In comparison to other models, the random forest classification had the highest accuracy (0.87) and AUC (0.87) for predicting NICU admission. According to our findings, the most significant predictors of NICU admission among several maternal and clinical factors were gestational age, maternal age, parity, a history of neonatal death, onset of labor, multiple pregnancy, fetal distress, meconium-stained amniotic fluid, method of childbirth, neonatal weight, and sex. We have identified several important factors that increase the likelihood of newborns being admitted to the NICU, which could assist in predicting the need for additional neonatal care during delivery and in advising women on the chances of NICU admission.

Introduction

During the neonatal phase, which is a crucial period in an individual’s life, a newborn must acclimate to a new environment and undergo various physiological changes necessary for survival. The mortality rate among newborns is a key factor in under-five mortality rates¹. In 2018, it was reported that over 2.4 million children passed away before reaching their second month of life². Neonatal medical information is utilized for a range of reasons, including assessing the well-being of newborns, guiding clinical decisions, identifying illnesses, and aiding in patient care and treatment. Through the examination of critical details within this data using artificial intelligence algorithms, significant goals like early detection of diseases, formulation of effective treatment strategies, and promotion of favorable growth and development may be achieved^3,4. Scientists are aiming to create a system that can predict the likelihood of newborns being admitted to the neonatal intensive care unit (NICU) by combining various statistical methods. This prediction could potentially reduce the negative health outcomes, deaths, and medical costs associated with NICU stays by detecting potential cases early⁵.

NICU admission entails risks for newborns and stress for their families, at a high cost to the healthcare system. The majority of studies that have evaluated admission and risk factors for NICU use have focused primarily on preterm infants⁶. A more comprehensive understanding of the factors that increase the risk of NICU admission can help in developing strategies to enhance the survival rates of newborns. The aim of this study is to identify predictors of NICU admission using a machine learning algorithm.

Methods

This was a secondary analysis of our previous study⁴. The study utilized a retrospective cohort design. The primary outcome of the research focused on admissions to the NICU. Real -time data such as age, place of residence, maternal education, medical insurance, nationality, infertility history, neonatal death history, labor onset, gestational age, parity, multiple pregnancies, attendance at prenatal education classes, presence of a doula during labor, childbirth method, shoulder dystocia, analgesia during cesarean sections, maternal anemia, cardiovascular disease, substance use, prolonged membrane rupture, diabetes, maternal body mass index (BMI), preeclampsia, hypothyroidism, placental abruption, meconium in amniotic fluid, fetal distress, intrauterine growth restriction, chronic hypertension, newborn weight, and newborn gender were extracted from the electronic health records of Khaleej-e-Fars Hospital in Bandar Abbas, Iran, a facility that offers specialized medical services. Electronic health data for all deliveries is gathered and managed by midwives as a regular part of clinical care. The study was approved by the Research Committee Board of Hormozgan University of Medical Sciences. We received electronic health records for all deliveries between January 2020 and December 2022. Women with a cephalic presentation were included, while fetal malformations were used as exclusion criteria.

The first step of the analysis involved comparing healthy babies to those admitted to the NICU. Variables that had a significant p-value (less than 0.05) were selected as features for the machine learning approach. The input data was utilized to train nine different machine learning models: linear regression, logistic regression, decision tree classification, random forest classification, XGBoost classification, permutation classification- KNN, light gradient boosting, deep learning, and support vector machine. With the exception of tree-based models, all machine learning models underwent L2 normalization for feature standardization. The output of each machine learning model ranged from 0 to 1.

Because NICU admission is not common, we expected that healthy newborns would be larger in size than newborns admitted to the NICU. To ensure that the model development and performance evaluation were not affected by imbalanced datasets, random undersampling with a 1:10 ratio of critical to non-critical groups was performed. The full dataset was utilized for creating and testing the model using a five-fold cross-validation method. This approach helped ensure the proper separation of training and testing data, minimizing any distortion in the results due to specific divisions. More specifically, the data was divided into five sections, with four sections used for training and one for testing. The five-fold testing was conducted without any overlap.

In our assessment, we used accuracy as a measure, which indicates the percentage of correct predictions out of all predictions made. We also considered the area under the receiver operating characteristic curve (AUC), precision (the ratio of correct predictions for a class to the total number of predictions for that class), recall (the ratio of correct predictions for a class to the total number of actual instances of that class), and F1-Score (the weighted mean of precision and recall, with an F1 score ranging from 0 to 1, where 1 is the ideal value) to evaluate effectiveness. SPSS (version 25.0, IBM Corp, Armonk, NY, United States) and Python software (version 3.7.0) were used for conducting all statistical analyses.

Results

During the study period, the rate of NICU admission in our center was 477 out of 3,062 deliveries (15.5%). In Table 1, you can find the maternal demographic characteristics associated with NICU admission. Maternal age was found to be related to NICU admission. The newborns of adolescent and advanced-aged mothers were at a higher risk of NICU admission.

Table 1.

Demographic factors associated with the neonatal intensive care admission.

Demographic characteristics	Well-baby (n = 2585)	NICU-admission (n = 477)	P-value
Age (years)
13–19	96 (3.7)	27 (5.7)	0.007
20–34	1882 (72.8)	314 (65.8)
35–40	507 (19.6)	119 (24.9)
Above 40	100 (3.9)	17 (3.6)
Residency place
Urban	1950 (75.4)	370 (77.6)	0.352
Rural	635 (24.6)	107 (22.4)
Education
Primary	599 (23.1)	100 (29.9)	0.338
High school/diploma	1326 (51.3)	255 (53.5)
Advanced	660 (25.5)	122 (25.6)
Medical insurance
Yes	2298 (88.9)	429 (89.9)	0.576
No	287 (11.1)	48 (10.1)
Prenatal education course
Yes	99 (3.8)	13 (2.7)	0.717
No	2486 (96.2)	464 (97.3)
Nationality
Iranian	2565 (99.2)	473 (99.2)	0.781
Non-Iranian	20 (0.8)	4 (0.8)

Data are presented as n (%).

NICU neonatal intensive care unit admission.

Table 2 shows a connection between obstetric factors and NICU-admission. Factors such history of neonatal death, onset of labor, gestational age, parity, multiple pregnancy, the presence of a doula, the method of childbirth, and analgesia during cesarean section were all linked to NICU-admission.

Table 2.

Obstetric factors associated with the neonatal intensive care unit admission.

Variables	Well-baby (n = 2585)	NICU-admission (n = 477)	P-value
History of infertility
Yes	18 (0.7)	12 (2.5)	0.060
No	2567 (99.3)	465 (97.5)
History of neonatal death
Yes	9 (0.3)	6 (1.3)	0.020
No	2576 (99.7)	471 (98.7)
Onset of labor
Induced	1305 (50.5)	257 (53.9)	0.015
Spontaneous	693 (26.8)	98 (20.5)
Cesarean before the onset of labor	587 (22.7)	122 (25.6)
Gestational age (week)
Late-term (more than 41)	306 (11.5)	43 (9.0)	< 0.001
Term (37+ 1-41)	2034 (78.7)	239 (50.1)
Preterm (24–37)	245 (9.5)	195 (40.9)
Parity
Primiparous	659 (25.5)	147 (30.8)	0.017
Multiparous	1926 (74.5)	330 (69.2)
Multiple pregnancy
Yes	42 (1.6)	50 (10.5)	< 0.001
No	2543 (98.4)	427 (89.5)
Attending of doula
Yes	602 (23.3)	56 (11.7)	< 0.001
No	1983 (76.7)	421 (88.3)
Method of childbirth
Vaginal delivery	1599 (61.9)	166 (34.8)	< 0.001
Operative vaginal delivery	9 (0.3)	1 (0.2)
Cesarean section	977 (37.8)	310 (65.0)
Analgesia during cesarean section
General	34 (3.5)	18 (5.8)	0.028
Spinal	943 (96.5)	292 (94.2)

Data are presented as n (%).

NICU neonatal intensive care unit admission.

Table 3 displays the connection between maternal and newborn clinical factors and admission to the NICU. Factors such as maternal diabetes, preeclampsia, hypothyroidism, meconium in the amniotic fluid, fetal distress, shoulder dystocia, newborn weight, and sex were associated with a higher probability of NICU admission.

Table 3.

Clinical factors associated with the neonatal intensive care unit admission.

Variables	Well-baby (n = 2585)	NICU-admission (n = 477)	P-value
Maternal anemia
No	2537 (98.1)	466 (97.7)	0.471
Yes	48 (1.9)	11 (2.3)
Maternal cardiovascular disease
No	2569 (99.4)	470 (98.5)	0.055
Yes	16 (0.6)	7 (1.5)
Prolonged rupture of membrane
No	2553 (98.8)	465 (97.5)	0.055
Yes	32 (1.2)	12 (2.5)
Maternal diabetes
No	2059 (79.7)	363 (76.1)	0.047
Yes	526 (20.3)	114 (23.9)
Maternal body mass index (kg/m2)
Less than 18.5	113 (4.4)	19 (4.0)	0.207
18.5–24.9	1640 (63.4)	281 (58.9)
25-29.9	677 (26.2)	142 (29.8)
30 and above	155 (6.0)	35 (7.3)
Preeclampsia
No	2504 (96.9)	444 (93.1)	< 0.001
Yes	81 (3.1)	33 (6.9)
Maternal hypothyroidism
No	2272 (87.9)	400 (83.9)	0.017
Yes	313 (12.1)	77 (16.1)
Placenta abruption
No	2563 (99.1)	466 (97.7)	0.012
Yes	22 (0.9)	11 (2.3)
Meconium amniotic fluid
No	2367 (91.6)	415 (87.0)	0.020
Yes	218 (8.4)	62 (13.0)
Fetal distress
No	2488 (96.2)	425 (89.1)	< 0.001
Yes	97 (3.8)	52 (10.9)
Intrauterine growth retardation
No	2506 (96.9)	460 (96.4)	0.567
Yes	79 (3.1)	17 (3.6)
Chronic hypertension
No	2560 (99.0)	469 (98.3)	0.222
Yes	25 (1.0)	8 (1.7)
Substances use
No	2583 (99.9)	472 (99.0)	0.010
Yes	2 (0.1)	5 (1.0)
Newborn sex
Male	1303 (50.4)	275 (57.7)	0.004
Female	1282 (49.6)	202 (42.3)
Newborn weight (gr)
Less than 1500	18 (0.7)	37 (7.8)	< 0.001
1501–2500	247 (9.6)	112 (23.5)
2501–4000	2278 (88.1)	323 (67.7)
Above 4000	42 (1.6)	5 (1.0)
Shoulder dystocia
No	2581 (99.8)	472 (99.0)	0.007
Yes	4 (0.2)	5 (1.0)

Data are presented as n (%).

NICU neonatal intensive care unit admission.

The area under the curve (AUC) for each model was as follows : linear regression (0.79), logistic regression (0.79), decision tree classification (0.83), random forest classification (0.87), XGBoost classification (0.86), permutation classification - KNN (0.81), light gradient boosting (0.80), deep learning (0.90), and support vector machine (0.79). In comparison to other models, the random forest classification exhibited the highest accuracy (0.87) and AUC (0.87) for predicting NICU admission. The machine learning models showed varying performance, as indicated in Table 4.

Table 4.

The performance of machine learning models.

Machine learning model	Accuracy	Precision	Sensitivity	Specificity	AUC	FPR	FNR
Linear regression	0.79	0.72	0.80	0.76	0.79	0.25	0.17
Logistic regression	0.79	0.75	0.83	0.75	0.79	0.24	0.16
Decision tree classification	0.83	0.76	0.90	0.76	0.83	0.24	0.09
Random forest classification	0.87	0.80	0.94	0.80	0.87	0.20	0.05
XGBoost classification	0.86	0.81	0.92	0.81	0.86	0.19	0.08
Permutation classification, knn	0.81	0.67	0.95	0.68	0.81	0.32	0.04
Light gradient-boosting	0.79	0.73	0.86	0.73	0.80	0.27	0.13
Deep learning	0.82	0.80	0.84	0.80	0.90	0.19	0.15
Super vector machine	0.79	0.74	0.85	0.74	0.79	0.25	0.15

AUC area under curve, FPR false positive rate, FNR false negative rate.

As shown in Fig. 1, our findings indicate that the most significant predictors of NICU admission among various maternal and clinical factors were gestational age, maternal age, parity, history of neonatal death, onset of labor, multiple pregnancy, fetal distress, meconium in amniotic fluid, method of childbirth, neonatal weight, and sex.

Fig. 1.

Feature importance of the random forest classification in the prediction of Nicu-admission.

Discussion

We conducted research in which we created and evaluated multiple machine learning models to predict admissions to the NICU using real-time data. To the best of our knowledge, this is the second study that used machine learning model to predict NICU-admission. A previous study indicated that factors such as preterm deliveries, hypertension, amniotic fluid index, birth weight, mode of delivery, and maternal complications could potentially increase the risk of NICU-admission. In comparison to other models, the decision tree had the highest accuracy (0.921) and AUC (0.966) for predicting NICU-admission⁷.

According to our findings, the random forest classification had the highest accuracy (0.87) and AUC (0.87) for predicting NICU admission. The most weighted predictors of NICU admission were gestational age, maternal age, parity, a history of neonatal death, onset of labor, multiple pregnancy, fetal distress, meconium in amniotic fluid, method of childbirth, neonatal weight, and sex.

The main factor determining the need for admission to the NICU is generally agreed upon to be the gestational age of the infant. Compared to term infants, preterm newborns have a significantly increased risk of NICU admission. Preterm newborns are more likely to experience respiratory problems, necessitating specialized care⁸. Because multiple pregnancies are often terminated earlier than single pregnancies, it was anticipated that there would be a higher rate of NICU admissions for multiple pregnancies. In line with our findings, a recent study demonstrated an increase in the rate of NICU admissions for multiple pregnancies due to lower birth weights and shorter gestational ages⁹.

Neonatal weight is an important factor in determining if critical care is required. Most babies with low birth weight are born prematurely. Furthermore, low birth weight neonates are more likely to experience respiratory distress syndrome and birth asphyxia¹⁰. These factors are among the main reasons for admission to the NICU.

The mothers included in this research were between 13 and 47 years old, falling within the age range for women in their fertility years. However, just over 14% of the babies needing NICU care were born to mothers aged 20–34. Many studies in-line with our findings, have shown a strong link between teenage pregnancy and NICU-admissions¹¹, while previous study has not found teenage pregnancy to be a major factor^12,13. Likewise, some research has indicated a higher risk of NICU-admissions for babies born to older mothers^13,14, although other studies have not found a significant association between advanced maternal age and NICU-admissions^15,16.

There is no consistent evidence to support the relationship between parity and negative neonatal outcomes, but many studies have shown that multiparous women are more likely to have their newborns admitted to the NICU compared to nulliparous women^17,18. Some believe that the interaction between maternal age and parity is related to the higher risk of adverse maternal conditions such as gestational hypertension, eclampsia/pre-eclampsia, placenta previa, and preterm birth, ultimately resulting in an increased risk of NICU admission¹⁹. However, the risk of NICU admission was found to be elevated in previous studies among nulliparous women²⁰.

As we observed, the rate of NICU admissions was higher in mothers with a history of neonatal death. Several factors need to be considered: initially, it is possible that previous infant fatalities were caused by abnormalities or recurring health conditions present in the current pregnancy. Secondly, the mother may have underlying health conditions that have impacted both past and current pregnancies, leading to an increased reliance on NICU admissions. Lastly, the oversensitivity of healthcare providers in caring for the newborn of a mother with a previous history of neonatal death may lead to an increase in cases of misdiagnosis and hospitalization rates.

According to the present study, more neonates born through induced labor required NICU admission compared to those born through spontaneous labor. Many studies confirmed this association^21,22. However, some previous studies showed no association²⁰.

Meconium-stained amniotic fluid was another predictor of NICU admission based on our findings. It has been reported that around 10–15% of newborn babies are delivered through amniotic fluid that is stained with meconium, and of those, 3–9% end up suffering from meconium aspiration syndrome²³. Between 8% and 20% of infants born through meconium-stained amniotic fluid show signs of depression and lack of activity, with symptoms such as a slow heart rate, insufficient breathing, and weak muscle tone²⁴, which are all indications for NICU admission.

The presence of low oxygen levels during labor or before it can lead to issues such as cognitive impairment and cerebral palsy in newborns²⁵. Obstetricians and gynecologists are focused on anticipating the outcomes for babies during childbirth. Many maternity health centers utilize fetal monitoring to detect fetal distress during labor. Despite being a long-standing practice, discussions regarding the accuracy of this test in predicting outcomes continue. Nevertheless, our findings suggest that detecting fetal distress through continuous or intermittent monitoring of the fetal heart rate was associated with newborns requiring NICU care.

In recent years, there has been an examination of the impact of the method of delivery (such as vaginal birth or cesarean section) on newborns, primarily driven by the rising rates of cesarean sections. The rate of neonatal morbidity differs depending on the method of delivery. Babies delivered via cesarean section have a higher risk of respiratory morbidity. On the other hand, babies born vaginally are more prone to intracranial hemorrhage, brachial plexus injury, and culture-positive neonatal sepsis²⁶. According to our research results, the method of delivery can help predict whether a newborn will require admission to the NICU. Infants born via cesarean section are more likely to be admitted to the NICU.

According to our findings, more male newborns were observed among the neonates admitted to the NICU compared to female newborns. The sex of the fetus is linked to pregnancy complications and affects the outcomes of newborns. Studies show that females tend to have better outcomes during childbirth compared to males²⁷.

The innovative research on employing machine learning models in maternal and neonatal health demonstrates encouraging outcomes that inspire researchers to conduct further studies with this method to draw more conclusive insights about their effectiveness^28–31. The use of nine different machine learning models in our research is a significant advantage. Nevertheless, there are a few limitations to our study. One issue is the retrospective nature of the design. To minimize selection bias, we attempted to encompass all successive mothers who delivered during the research period. Essential variables that could enable the evaluation of the fetus, such as the biophysical profile, amniotic fluid index, and corticosteroid usage before delivery in cases of preterm births, were not included in the machine learning dataset. These limitations need to be considered in future research.

Conclusion

We have identified several important factors that increase the likelihood of newborns being admitted to the NICU, which could assist in predicting the need for additional neonatal care during delivery and in advising women on the probabilities of NICU admission. Utilizing a clinical database and various machine learning algorithms to assess the risk of NICU admission has potential benefits. Further studies incorporating intrapartum clinical characteristics are necessary to enhance the accuracy of predictions.

Abbreviations

NICU

Neonatal intensive care unit

AUC

Area under the curve

BMI

Body mass index

Author contributions

N.M. and F.D. wrote the proposal. F.D. and V.M. contributed significantly to data collection. The findings were analyzed and interpreted by M.V. F.D., the primary contributor to the manuscript’s commenting and editing. F.A. assessed the manuscript’s scientific content critically. The final manuscript for submission was read and approved by all authors.

Funding

All of the authors acknowledged Hormozgan University of Medical Sciences.

Data availability

The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

This research adhered to the Declaration of Helsinki and was conducted following the approval of the Ethics Committee. The Research Committee Board of Hormozgan University of Medical Sciences approved the study (IR.HUMS.REC.1402.332). The files of every patient who provided informed consent for the use of their data in research were examined. For individuals under eighteen years old, consent was obtained from their guardians. Statistical analysis was conducted while maintaining patient anonymity in accordance with the Ethics Committee guidelines.