In-birth CPR in neonates; Predictable or Unpredictable?

Background: Anticipating on in-birth Cardiopulmonary Resuscitation(CPR) in neonates is very important and complex. Timely identication and rapid CPR in neonates in the delivery room signicantly affect reducing the mortality and other neurological disabilities. The aim of this study is to create a prediction system for identifying the need to in-birth CPR in neonates based on Machine Learning(ML) algorithms. Methods: In this study, 3882 neonatal medical records were retrospectively reviewed. Records were extracted from the maternal, fetal, and neonatal registry of Valiasr hospital in Tehran. A total of 60 risk factors were extracted, and ve ML algorithms including J48, Naïve Bayesian, Multilayer Perceptron (MLP) (cid:0) Support Vector Machine (SVM) and Random Forest were compared to predict the need to in-birth CPR in neonates. Also, using 10 feature selection algorithms, the features were ranked based on the importance, and using the ML algorithms, the important risk factors were identied. Results: In order to predict the need to in-birth CPR in neonates, SVM using all risk factors reached the accuracy of 88.43% and F-measure of 88.4%, while MLP using the 15 rst important features reached the accuracy of 90.86% and the F-measure of 90.8%. The most important risk factors included gestational age, delivery type, presentation, steroid administration, macrosomia, prenatal care, infant number and rank, mother addiction, maternal chronic disease history, fetal hydrops, amniotic uid, gestational hypertension, infertility and placental abruption. Conclusions: The proposed system can be useful in predicting the need to CPR in neonates in the delivery room. around 0.1% of term neonates require advanced CPR at the moment of childbirth(4, 5). On the other hand, these statistics are very higher for preterm infants; 6–7% of preterm neonates (GA < 32 weeks) (6) and around 6–10% of Very Low Birth Weight (VLBW) and Extremely Low Birth Weight (ELBW) infants require advanced CPR, i.e. chest compression with or without injecting epinephrine (7). Many studies have been performed on CPR consequences, and it has been found that mortality, neurological morbidity, neurodevelopmental impairment, lower motor scores and ROP are more prevalent among the preterm infants who have received CPR(6, 8, 9). Thus, timely identication and rapid CPR of neonates in the delivery room can reduce the neonatal mortality and morbidity(7).


Background
Annually, around 1 million neonates die because of birth asphyxia worldwide (1). According to the WHO guideline on basic newborn resuscitation, although around one fourth of neonatal mortality is due to birth asphyxia, effective CPR at the childbirth moment can prevent a large number of these deaths (2). Most neonates enter from intrauterine to extrauterine life with no special assistance. However, less than 1% of all neonates (3) and around 0.1% of term neonates require advanced CPR at the moment of childbirth (4,5). On the other hand, these statistics are very higher for preterm infants; 6-7% of preterm neonates (GA < 32 weeks) (6) and around 6-10% of Very Low Birth Weight (VLBW) and Extremely Low Birth Weight (ELBW) infants require advanced CPR, i.e. chest compression with or without injecting epinephrine (7). Many studies have been performed on CPR consequences, and it has been found that mortality, neurological morbidity, neurodevelopmental impairment, lower motor scores and ROP are more prevalent among the preterm infants who have received CPR (6,8,9). Thus, timely identi cation and rapid CPR of neonates in the delivery room can reduce the neonatal mortality and morbidity (7).
Currently, in-birth CPR is suggested for neonates with asystole, profound bradycardia (HR < than 60 per minutes), and pulseless electrical activity despite effective ventilation. Absence of heart rate or other vital signs, which is recorded as zero APGAR, can also be used as a guideline for decisionmaking on beginning CPR (8). Different studies have shown that the severity scoring systems have many limitations, and the systems based on machine learning have a better performance in prediction (10,11). Accordingly, considering the importance of in-birth CPR, usage of ML based systems can be useful to help anticipating on the need to neonatal CPR. Application of ML algorithms in the medicine and especially in neonatal medicine has shown that these techniques have a very suitable performance in prediction and diagnosis.
Nevertheless, sparse studies have dealt with CPR in neonates, most of which have a small set of samples and risk factors because of the challenges in data collection (12)(13)(14)(15)(16). The aim of most of them is only identifying the risk factors affecting the need to CPR (12,13,(17)(18)(19). Further, most studies have dealt with neonatal CPR in NICU, and fewer of them have addressed in-birth CPR, due to examine in-birth CPR, only maternal and fetal factors should be considered. To the best of our knowledge, no study has been performed on predicting the need to in-birth CPR in neonates using machine learning algorithms. Accordingly, our aim is to design a ML-based Clinical Decision Support System (CDSS) which predicts the need to inbirth CPR in neonates based on maternal and fetal factors.

Method
This retrospective study was conducted based on the maternal, prenatal and the fetal data, with the aim of predicting the need to in-birth CPR in neonates. In order to develop the prediction model, ML algorithms were used. Also, the models were evaluated to examine the performance and to determine the best model. The details related to the data, setting, method of development, and evaluation of the prediction models have been presented in this section.

Data source
The data were obtained through the neonatal registry system in Valiasr hospital a liated to Tehran University of Medical Sciences (TUMS). This registry includes the information related to all neonates hospitalized in the NICU of Valiasr Hospital that has a grade of B3. The data related to the mother and fetus are recorded by pediatric residents in the data collection form, and are then entered into the registry by its person in charge. In this retrospective study, the data available in this registry were retrieved anonymously from March 2016 to March 2020. Consent form has also been lled by the father or mother of the infants before loading the data into the registry.
The study was approved by the TUMS institutional review board (Approval ID: IR.TUMS.VCR.REC.1398.591).

Inclusion and exclusion criteria
All neonates hospitalized in NICU of Valisasr Hospital from March 2016 to March 2020 were included in this study. The post-delivery data such as Apgar score, height, and weight of the neonate were excluded.

De nition
In this study, delivery room CPR and CPR immediately after birth have been examined. CPR refers to any activity in the delivery room taken to simulate the cardiorespiratory activity of neonates who met the conditions of CPR according to the American Academy of pediatrics (AAP) guidelines (20,21). These activities can be categorized into two groups: primary CPR (use of oxygen mask, Nasal continuous positive airway pressure (CPAP) and positive pressure ventilation(PPV)) and advanced CPR (primary CPR plus epinephrine injection, chest compression and intubation) (22).

Data extraction and preprocessing
After removing the any identi ers, the data was extracted from the registry as a .sav le and includes six groups: The outcome variable is whether or not CPR is performed for a baby in the delivery room. The discrete and continuous missing values in the data set were imputed by the mode and mean of each variable, respectively.

Prediction models construction
In order to develop the prediction model for the need to in-birth CPR, ML algorithms were used including J48, MLP, SVM, Naïve Bayesian (NB) and Random Forest (RF). All of these algorithms were implemented for the original data set. Next, using Feature Selection (FS) algorithms, the importance of each feature in predicting CPR was examined. For this purpose, lter FS algorithms including Relief, Correlation-based feature selection (CFS), Pearson's Correlation, gain ratio, info gain, OneR and symmetrical uncertainty as well as wrapper methods using three classi ers SVM, J48 and NB were used (Table 1). Then, the risk factors were organized based on the total importance resulting from implementing 10 FS algorithms. Based on the ordered list of variables, various data subsets were created, and ML algorithms were implemented on both the original data set as well as these data subsets. For the continuous data, mean and standard deviation, while for discrete data, frequency and percentage have been reported. To investigate the distribution of variables in the two groups (neonates receiving CPR vs those not receiving CPR), independent samples t-test, chi-square, Fisher exact and Mann-Whitney tests have been used. The signi cance level in all tests was considered p < 0.05. All statistical analyses were performed by SPSS 20. After analyzing the role of risk factors in predicting the outcome and developing the prediction models for the need to in-birth CPR, the performance of the developed models was evaluated based on the accuracy, precision, sensitivity, speci city and F-measure criteria and using the 10-fold cross validation method. The role of variables was analyzed using the FS algorithms in WEKA software. Development and assessment of models were also performed using R v3.4.1.
2.7. Clinical decision support system design After selecting the best algorithm in the predicting the need to delivery room CPR in neonates, the system user interface was designed based on the best algorithm in Visual Studio platform 2015.

Results
A total of 3882 infants were included into the study according to the inclusion/exclusion criteria, out of them 2011(51.8%) had received delivery room CPR. The efforts were taken for the CPR, in the order of frequency, were: nasal CPAP (n = 1120, P = 28.8%), PPV (n = 891, P = 22.9%), Oxygen mask (n = 723, P = 18.6%), intubation (n = 494, P = 12.7%), chest compression (n = 86, P = 2.2%) and epinephrine injection (n = 68, P = 1.7%). Data statistics are shown in Tables 2 and 3. In order to develop the prediction model for the need to in-birth CPR, ML algorithms were used. Figure 1 displays the results obtained from implementing these algorithms on the original data set. Figure 1 shows that based on all performance criteria the SVM method has had the best performance in predicting the need to in-birth CPR. Also, J48 method reached comparable results. In the next step of simulation, FS algorithms were employed. For this purpose, seven lter FS algorithms including Relief, Correlation-based feature selection (CFS), Pearson's Correlation, gain ratio, info gain, OneR and symmetrical uncertainty as well as three wrapper methods using three classi ers SVM, J48 and NB were implemented (appendix A). Then, for each risk factor, the total importance resulting from implementing 10 FS algorithms was calculated. Table 4 presents the rank resulting from implementing the FS algorithms as well as the averaged rank for each variable. The averaged rank was calculated using the following relation, where r i represents the rank of variable in the i th feature selection algorithm.  53  49  43  33  42  22  41  36  25  3  34.7   37  (Pre) Eclampsia  54  29  24  32  24  46  27  42  16  58  35.2   38  Thyroid Disorders  19  20  32  36  22  43  24  55  48  53  35.2   39  Eye diseases  33  32  53  40  52  40  52  37  12  7  35.8   40  Abortion  30  43  45  43  44  41  44  48  13  10  Averaged rank: (r 1 + r 2 + … + r 10 ) / 10 According to Table 4, GA is the most important risk factor based on implementation of all FS algorithms. Also, the averaged rank of the "maternal kidney disease" variable was the lowest, suggesting that it is the least important risk factor. The variables have been sorted based on the averaged rank, then, 20 feature subsets were created, including 1, 2, …, 10,15,20,25,30,35,40,45,50,55 and 60 important variables respectively. According to these subsets, 20 data subsets were created, and ML algorithms were implemented on these data subsets. Figures 2 and 3 Table 5 provides the best results of every algorithm. Comparing Fig. 1 and Table 5, it is found that use of FS algorithms has caused increased accuracy and F-measure by 4.7% and 4.9% respectively on average. Graphical user interface of the proposed CDSS was designed based on the best algorithm in Visual Studio 2015 and shown in Fig. 4.

Discussion And Conclusions
This paper dealt with a prediction system for the need to neonatal CPR immediately after birth in delivery room. In order to achieve a system with proper performance, various ML algorithms were compared with different sets of risk factors to identify the best system as well as the most effective factors in CPR predicting. According to the obtained results, in order to predict the need to CPR in neonates, SVM using all risk factors reached accuracy of 88.43% and F-measure of 88.4%, while MLP using the rst 15 most important variables reached accuracy of 90.86% and F-measure of 90.8%.
Feature ranking was performed using 10 FS algorithms and the most effective risk factors were gestational age, delivery type, presentation, steroid administration, macrosomia, prenatal care, number of infant, mother addiction, fetal hydrops, rank of infant, amniotic uid status, gestational hypertension, infertility, placental abruption and maternal chronic disease history, respectively. According to the sixth's edition of the textbook of neonatal resuscitation(23) and the International Liaison Committee On Resuscitation (ILCOR) guideline (24), the risk factors of GA, delivery type, presentation, macrosomia, prenatal care, multiple gestation, fetal hydrops, amniotic uid status, hypertension, placental abruption and maternal chronic disease history all may contribute to increasing the need to in-birth CPR in neonates. In the study by Afjeh et al. the risk factors affecting the CPR in neonates were examined, whereby placental abruption, multiple gestation, delivery type and infertility were identi ed as the risk factors which may contribute to increasing the need to delivery room CPR (17). Also, in the study by Jiang et al., it was found that diabetes, hypertension and delivery type affect the need to CPR in neonates (18).
The prevalence of mortality, neurodevelopmental impairment, respiratory support at 28 days, days to full oral feeds and length of stay are very high among the neonates who have received in-birth CPR (8,25). Even the national institute of child health and developmental neonatal research reported that CPR in delivery room is a prognostic factor for morbidity and later complications up to 18 months of age (26). Thus, the healthcare system should be able to predict which neonates require CPR before delivery, so that the neonatal resuscitation team would present in time (27). Previous studies have shown that antenatal transfer of high-risk mothers reduces pre-discharge neonatal mortality (28,29). Thus, predicting the need to inbirth CPR can be very effective, as it increases the preparation of the neonatal resuscitation team and provides the possibility of consultation with the family before delivery (27). Therefore, according to the results obtained from this study, use of the proposed system for predicting the need to inbirth CPR in neonates will have a great signi cance in reducing the adverse outcomes in childbirth and better preparation of the CPR team.
In addition, the coordination between the CPR team and obstetricians can lead to reduced adverse events in the delivery room and improve the care (27). In the Draper's study, intrapartum deaths in the UK were examined and it was found around 25% of mortalities have been due to lack of suitable communication between the multidisciplinary team during delivery (30). Thus, the proposed system can be used as an infant pre-resuscitation guide in order to ensure coordination between the CPR team and obstetricians.
In spite of the great signi cance of CPR prediction, very few studies have dealt with neonatal CPR, out of which mostly have addressed CPR in NICU (12)(13)(14)(15), which have samples with a small sample size and few risk factors because of the challenges in data collection (12)(13)(14)(15)(16). However, in this study, in addition to considering a sample with suitable size, attempts were made to capture all fetal and maternal risk factors mentioned in credible guidelines which also had demonstrated their importance in previous studies.
The main limitation of this study, like most previous studies, was that the data related to only one center were examined. Thus, it is suggested to conduct studies with a more diverse sample extracted from multiple centers with different grades of NICU. Comparison of the results can be useful in identifying signi cant risk factors affecting the need to CPR and thus in its prediction.