Predicting the Risk of Adverse Events in Pregnant Women With Congenital Heart Disease.

Background Women with congenital heart disease are considered at high risk for adverse events. Therefore, we aim to establish 2 prediction models for mothers and their offspring, which can predict the risk of adverse events occurred in pregnant women with congenital heart disease. Methods and Results A total of 318 pregnant women with congenital heart disease were included; 213 women were divided into the development cohort, and 105 women were divided into the validation cohort. Least absolute shrinkage and selection operator was used for predictor selection. After validation, multivariate logistic regression analysis was used to develop the model. Machine learning algorithms (support vector machine, random forest, AdaBoost, decision tree, k-nearest neighbor, naïve Bayes, and multilayer perceptron) were used to further verify the predictive ability of the model. Forty-one (12.9%) women experienced adverse maternal events, and 93 (29.2%) neonates experienced adverse neonatal events. Seven high-risk factors were discovered in the maternal model, including New York Heart Association class, Eisenmenger syndrome, pulmonary hypertension, left ventricular ejection fraction, sinus tachycardia, arterial blood oxygen saturation, and pregnancy duration. The machine learning-based algorithms showed that the maternal model had an accuracy of 0.76 to 0.86 (area under the receiver operating characteristic curve=0.74-0.87) in the development cohort, and 0.72 to 0.86 (area under the receiver operating characteristic curve=0.68-0.80) in the validation cohort. Three high-risk factors were discovered in the neonatal model, including Eisenmenger syndrome, preeclampsia, and arterial blood oxygen saturation. The machine learning-based algorithms showed that the neonatal model had an accuracy of 0.75 to 0.80 (area under the receiver operating characteristic curve=0.71-0.77) in the development cohort, and 0.72 to 0.79 (area under the receiver operating characteristic curve=0.69-0.76) in the validation cohort. Conclusions Two prenatal risk assessment models for both adverse maternal and neonatal events were established, which might assist clinicians in tailoring precise management and therapy in pregnant women with congenital heart disease.

area under the receiver operating characteristic curve

congenital heart disease

Eisenmenger syndrome

least absolute shrinkage and selection operator

machine learning

New York Heart Association

pulmonary arterial hypertension

arterial blood oxygen saturation

Clinical Perspective

What Is New?

This cohort study of 318 pregnant women with congenital heart disease who gave birth after 28 gestational weeks found that 12.9% women experienced adverse maternal events, and 29.2% neonates experienced adverse neonatal events.

The models constructed by our study for both adverse maternal and neonatal events had obtained a good prediction accuracy in the development and validation cohorts.

What Are the Clinical Implications?

We established the first prenatal risk assessment model for pregnant women with congenital heart disease during the last trimester of pregnancy.

More importantly, the models in our study are particularly suitable for clinical use in developing countries where prepregnancy counseling and pregnancy monitoring systems are deficient.

These models may also be used in the design of optimal treatment strategies for pregnant women with congenital heart disease.

With the progress of medical care for patients with congenital heart disease (CHD), their life expectancy has increased significantly over the past decades, and more women with CHD survive into their childbearing years. Between 1998 and 2007, the number of deliveries of pregnant women with CHD increased by 34.9% compared with 21.3% in the general population in the United States. For patients with a significant residua or those who are surgically uncorrected, pregnancy is often considered to be contraindicated. In these cases, pregnancy is accompanied by an increase in maternal morbidity and mortality. According to the different kinds of underlying defect and previous treatment strategies, pregnancy is a physiological stress, and complications during pregnancy such as gestational diabetes mellitus or pregnancy‐induced hypertension place those women at a higher risk for adverse events in the third trimester. , Despite earlier interventional or operative therapy, pulmonary arterial hypertension (PAH) is common and predisposes women with CHD to more symptoms and further clinical deterioration. Women with CHD have an increased risk of poor pregnancy outcomes, so clinical counseling and multidisciplinary specialist care should be provided before conception. Meanwhile, preexisting heart disease should be highlighted before conception counseling, and necessarily, information on the potential risk of adverse obstetric and fetal outcomes should be provided to pregnant women with CHD.

In developing countries, the long‐term health follow‐up monitoring system for patients with CHD is not perfect. For pregnant women with CHD who lack prepregnancy assessment and health monitoring during pregnancy, medical management in the perinatal period is a huge challenge for healthcare professionals. For attending cardiologists and obstetricians, adequate risk assessment is critical in optimizing pregnancy management, especially for pregnant women who are about to give birth shortly after diagnosis.

Therefore, the aim of the present study is to develop prognostic models to optimize the prenatal management of pregnant women with CHD and to obtain better prognostic outcomes for mothers and infants.

Methods

Researchers may contact the corresponding authors for the data within the article for future analysis.

Study Population

Patients with CHD who gave birth after 28 gestational weeks in Qilu Hospital of Shandong University from January 2004 to June 2019 were recruited. The model development cohort included 213 pregnant women with CHD, and data were collected from January 2004 to May 2016; an independent validation cohort included 105 patients with CHD, and data were collected from June 2016 to June 2019. A summary of the research procedure is shown in Figure 1. All patients had echocardiography results and were diagnosed with CHD by cardiologists. This study was guided by the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis statement (Table S1).

Figure 1

Flowchart of the study. (A) datasets; (B) model development and validation; (C) model training and validation based on machine learning. AUC indicates area under the receiver operating characteristic curve; CHD, congenital heart disease; LASSO, least absolute shrinkage and selection operator; NPV, negative predictive value; and PPV, positive predictive value.

Data Collection

Therapeutic records were used to collect patients’ information. From the original list of risk factors, the following candidate predictors were selected: maternal age, parity, New York Heart Association (NYHA) functional status, type of maternal congenital heart lesion, Eisenmenger syndrome (ES), history of cardiac surgery before pregnancy, preeclampsia, gestational diabetes mellitus, mode of delivery, pulmonary infection, hemoptysis, left ventricular ejection fraction, ascending aorta diameter, mitral regurgitation, PAH, sinus tachycardia, ectopic cardiac rhythm, arterial blood oxygen saturation (SaO2), hemoglobin, platelet, total serum protein, and pregnancy duration.

Definitions and Outcomes

Adverse maternal and neonatal events for each case were defined as the outcomes. Maternal and neonatal outcomes were composites of major adverse events. Maternal outcomes included cardiac death, heart failure, arrhythmia requiring treatment, and peripartum cardiomyopathy. Neonatal outcomes included preterm labor (<37 gestational weeks), small‐for‐gestational‐age birth weight (<10th percentile), low birth weight (<2500 g), intrauterine fetal death, and neonatal death. The cause of death of each patient was also collected in detail. Patients and children were following up for 6 weeks after delivery.

Ethics Statement

This retrospective study was approved by the Ethical Committee of Qilu Hospital of Shandong University (protocol number 2018 064) and obtained a waiver for informed consent.The names of the patients and their hospital admission numbers were anonymized before the analysis.

Statistical Analysis

Univariable logistic regression analysis was used for preliminary screening of clinical features, variables with a P<0.10 were selected. In addition, to find the optimal predictor selection algorithm, we compared the performances of several predictor selection methods, including least absolute shrinkage and selection operator (LASSO), maximum relevance minimum redundancy, and random forest. A comparison of these methods was also performed, and LASSO performs best in predictor selection. Afterward, multivariate logistic regression analysis was used to develop the risk prediction models for both mothers and offspring. The results are described as odds ratio, 95% CI, and P values. Then we developed 2 nomogram lists to predict the individual incidence of adverse events for each patient. The receiver operating characteristic curves and the area under the receiver operating characteristic curve (AUC) values were used to evaluate the classification of the model in both the development and validation cohorts. If the AUC was closer to 1, the model was seen as having a good efficacy ability of classification. The calibration slope with pointwise 95% confidence limits was used to estimate the calibration ability of the model.

Machine learning (ML) algorithms can perform better than traditional regression methods when the research aims to generate a model that can predict an outcome more accurately. ML offers an alternative approach to standard prognostic modeling, and its potential has been demonstrated in some recent studies. , , Therefore, we further used ML algorithms to model the selected high‐risk factors. We selected 7 widely used ML algorithms (support vector machine, RF, AdaBoost, decision tree, k‐nearest neighbor, naïve Bayes, and multilayer perceptron) to comprehensively evaluate our hypothesis. Tenfold cross validation was applied as a criterion for each classifier in the development cohort. The patients in the development cohort were randomly partitioned into 10 equal‐sized subsamples, where 9 subsamples were used as the training data, and 1 single subsample was retained as the validation data for testing. The average and standard deviation of the AUCs over the 10 tests performed in the multiple rounds of cross validation, as well as the corresponding sensitivity, specificity, positive predictive value, negative predictive value, and overall accuracy, were used to assess the performance of the 2 models. Then we retrained the models in the development cohort using the same hyperparameters as those in cross‐validation and evaluated their predictive ability in the independent validation cohort.

Statistical analysis was conducted with IBM SPSS statistics (version 24.0), R software (version 3.6.1), and Python (version 3.6.4) ML library scikit‐learn (version 0.19.1).

Results

The clinical characteristics of all patients in the development and validation cohorts are summarized in Table 1. The median age at the time of pregnancy was 27 years (range, 16–45 years). Vaginal delivery was observed in 32 (10.1%) patients, and cesarean section was observed in 286 (89.9%) patients. Two hundred sixty‐five (83.3%) patients delivered after 36 gestational weeks. Forty‐one (12.9%) women experienced adverse maternal events, and 93 (29.2%) neonates experienced adverse neonatal events. The details of the adverse maternal cardiac and neonatal events encountered during the perinatal period are presented in Table S2. Maternal mortality occurred in 13 patients, 2 of whom terminated the pregnancy attributable to intrauterine death and death caused by irreversible heart failure after cesarean section. One patient abandoned treatment because of family financial burden and cardiac arrest on the way home (Table S3).

Table 1

Patient’s Characteristics (Before Delivery)

Characteristic	Total (n=318)	Development Cohort (n=213)	Validation Cohort (n=105)
Age at delivery, y	27 (16–45)	26 (16–45)	29 (19–41)
Parity
0	220 (69.2)	160 (75.1)	60 (57.1)
≥1	98 (30.8)	53 (24.9)	45 (42.9)
Cardiac functional status
NYHA class I–II	248 (78.0)	159 (74.6)	89 (84.8)
NYHA class III–IV	70 (22.0)	54 (25.4)	16 (15.2)
Maternal congenital lesion
Atrial septal defect	123 (38.7)	76 (35.7)	47 (44.8)
Ventricular septal defect	101 (31.8)	75 (35.2)	26 (24.8)
Persistent ductus arteriosus	28 (8.8)	20 (9.4)	8 (7.6)
Tetralogy of Fallot	23 (7.2)	13 (6.1)	10 (9.5)
Ventricular outflow tract obstruction*	19 (6.0)	9 (4.2)	10 (9.5)
Other †	24 (7.5)	20 (9.4)	4 (3.8)
PAH, mm Hg
PAH <30	153 (48.1)	97 (45.5)	56 (53.3)
30 ≤PAH <60	87 (27.4)	58 (27.2)	29 (27.6)
60 ≤PAH <90	39 (12.3)	31 (14.6)	8 (7.6)
90 ≤PAH	39 (12.3)	27 (12.7)	12 (11.4)
Eisenmenger syndrome	26 (8.2)	18 (8.5)	8 (7.6)
Cardiac surgery before pregnancy
Corrected	111 (34.9)	72 (33.8)	39 (37.1)
Uncorrected	207 (65.1)	141 (66.2)	66 (62.9)
Preeclampsia	31 (9.7)	22 (10.3)	9 (8.6)
Gestational diabetes mellitus	13 (4.1)	8 (3.8)	5 (4.8)
Mode of delivery
Vaginal	32 (10.1)	21 (9.9)	11 (10.5)
Cesarean	286 (89.9)	192 (90.1)	94 (89.5)
Pregnancy duration, wk
28 ≤GW <32	47 (14.8)	30 (14.1)	17 (16.2)
32 ≤GW <36	6 (1.9)	5 (2.3)	1 (1.0)
36 ≤GW	265 (83.3)	178 (83.6)	87 (82.9)
Adverse maternal cardiac event	41 (12.9)	29 (13.6)	12 (11.4)
Adverse neonatal event	93 (29.2)	63 (29.6)	30 (28.6)

Values are median (range) or n (%). GW indicates gestational weeks; NYHA, New York Heart Association; and PAH, pulmonary hypertension.

Ventricular outflow tract obstruction including aortic valve stenosis and pulmonary valve stenosis.

Other including Marfan syndrome, mitral regurgitation, single ventricle, atrioventricular septal defect, congenitally corrected transposition of the great arteries, transposition of the great arteries, and so on.

Model Development and Validation

Risk factors related to adverse events are shown in Table S4. The results of predictor selection were summarized in Table S5. In the development cohort of both mother and offspring, the P values of LASSO, maximum relevance minimum redundancy, and random forest were all <0.0001. Using the AUC value as an evaluation, the LASSO method showed the best performances. In the validation cohort of both mother and offspring, the P value for LASSO method was the lowest, while its AUC was the highest. Therefore, the LASSO method was chosen as the optimal method in our study.

After predictor selection by LASSO analysis (Figure S1), the following 7 predictors were included in the maternal events model, including NYHA class, ES, PAH, left ventricular ejection fraction, sinus tachycardia, SaO2, and pregnancy duration. In addition, 3 high‐risk factors showed a correlation with adverse neonatal events, including ES, preeclampsia, and SaO2. Two nomogram lists were built according to the regression coefficients of the models (Figure 2A and 2B). The maternal events model yielded an AUC of 0.92 (95% CI, 0.86–0.97) in the development cohort and 0.80 (95% CI, 0.64–0.97) in the validation cohort. The AUC of the neonatal events model was 0.77 (95% CI, 0.70–0.84) in the development cohort and 0.76 (95% CI, 0.66–0.87) in the validation cohort (Figure 2C and 2D). The risk score for each patient and risk calculation equations of the prediction models are shown in Figure S2, and the calibration curves are illustrated in Figure S3.

Figure 2

Nomogram lists of the maternal model (A) and neonatal model (B); ROC curves of the maternal model (C) and neonatal model (D). Example of the maternal model in (A): A 37‐week pregnant woman with CHD (0 points) who had NYHA class III (12.5 points) without ES (0 points), and had a PAH of 35 mm Hg (2.5 points) and a left ventricular ejection fraction of 40% (70 points) with symptoms of sinus tachycardia (25 points) and an SaO2 of 98% (1 point) has a total score of 111 points; the corresponding probability of experiencing adverse events in this pregnancy was more than 50%. Example of the neonatal model (B): the pregnant woman described above who did not have ES (0 points) but had preeclampsia (15.5 points) with an SaO2 of 98% (2.5 points) has a total score of 18 points, and the corresponding probability of experiencing adverse neonatal events was more than 60%. AUC indicates area under the receiver operating characteristic curve; CHD, congenital heart disease; ES, Eisenmenger syndrome; GW, gestational week; LVEF, left ventricular ejection fraction; NYHA, New York Heart Association; PAH, pulmonary hypertension; ROC, receiver operating characteristic; and SaO2, arterial blood oxygen saturation.

Model Training and Validation Based on ML

The results of 10‐fold cross validation in the development and validation cohort are shown in Table 2. Seven ML algorithms showed that the adverse maternal model had an accuracy of 0.76 to 0.86 (AUC=0.74–0.87) in the development cohort, and 0.72 to 0.86 (AUC=0.68–0.80) in the validation cohort. In addition, 7 ML‐based algorithms showed that the adverse neonatal model had an accuracy of 0.75 to 0.80 (AUC=0.71–0.77) in the development cohort and 0.72 to 0.79 (AUC=0.69–0.76) in the validation cohort. The diagnostic performance of the logistic regression analysis and the ML algorithms (AUCs) are shown in Figure 3.

Table 2

Prediction of the 2 Models by LR and ML Analysis

	Threshold	AUC	Sensitivity	Specificity	PPV	NPV	Overall Accuracy
Maternal model in development cohort (n=213)
LR	0.13±0.09	0.85±0.14	0.80±0.11	0.73±0.29	0.95±0.05	0.42±0.21	0.79±0.08
SVM	0.17±0.05	0.85±0.12	0.84±0.08	0.65±0.27	0.94±0.05	0.40±0.16	0.81±0.08
RF	0.41±0.05	0.81±0.17	0.80±0.11	0.69±0.24	0.95±0.04	0.39±0.15	0.78±0.09
DT	0.58±0.05	0.74±0.10	0.83±0.08	0.58±0.20	0.93±0.03	0.37±0.12	0.79±0.07
KNN	0.33	0.77±0.16	0.86±0.09	0.63±0.31	0.94±0.05	0.41±0.19	0.83±0.08
NB	0.10±0.30	0.87±0.18	0.82±0.09	0.78±0.32	0.96±0.05	0.41±0.18	0.82±0.08
Ada	0.49±0.005	0.79±0.21	0.78±0.09	0.68±0.28	0.94±0.05	0.34±0.12	0.76±0.07
MLP	0.47±0.35	0.76±0.17	0.91±0.09	0.53±0.27	0.92±0.04	0.59±0.34	0.86±0.09
Neonatal model in development cohort (n=213)
LR	0.35±0.05	0.77±0.12	0.92±0.06	0.51±0.21	0.82±0.07	0.74±0.22	0.79±0.08
SVM	0.31±0.04	0.77±0.12	0.92±0.06	0.51±0.19	0.82±0.07	0.75±0.19	0.80±0.07
RF	0.50±0.05	0.76±0.11	0.92±0.06	0.51±0.19	0.82±0.07	0.75±0.19	0.80±0.07
DT	0.75±0.03	0.71±0.10	0.92±0.06	0.48±0.19	0.81±0.07	0.73±0.22	0.79±0.07
KNN	0.34±0.08	0.75±0.12	0.85±0.11	0.56±0.23	0.83±0.09	0.66±0.23	0.76±0.10
NB	0.08±0.22	0.74±0.08	0.92±0.06	0.51±0.19	0.82±0.07	0.75±0.19	0.80±0.07
Ada	0.50±0.004	0.77±0.12	0.84±0.10	0.59±0.19	0.83±0.08	0.65±0.20	0.77±0.10
MLP	0.33±0.04	0.72±0.17	0.85±0.11	0.51±0.20	0.81±0.08	0.63±0.25	0.75±0.11
Maternal model in validation cohort (n=105)
LR		0.80 (0.64–0.97)	0.78	0.67	0.95	0.29	0.77
SVM		0.80 (0.64–0.97)	0.87	0.58	0.94	0.37	0.84
RF		0.79 (0.64–0.94)	0.86	0.58	0.94	0.35	0.83
DT		0.68 (0.53–0.84)	0.85	0.50	0.93	0.30	0.81
KNN		0.79 (0.64–0.94)	0.88	0.67	0.95	0.42	0.86
NB		0.78 (0.61–0.95)	0.82	0.67	0.95	0.32	0.80
Ada		0.78 (0.61–0.95)	0.73	0.67	0.94	0.24	0.72
MLP		0.74 (0.59–0.89)	0.92	0.50	0.93	0.46	0.88
Neonatal model in validation cohort (n=105)
LR		0.76 (0.66–0.87)	0.91	0.50	0.82	0.68	0.79
SVM		0.76 (0.66–0.87)	0.91	0.50	0.82	0.68	0.79
RF		0.76 (0.66–0.87)	0.81	0.67	0.86	0.59	0.77
DT		0.69 (0.59–0.79)	0.91	0.47	0.81	0.67	0.78
KNN		0.75 (0.65–0.86)	0.75	0.67	0.85	0.51	0.72
NB		0.76 (0.65–0.86)	0.91	0.50	0.82	0.68	0.79
Ada		0.76 (0.65–0.86)	0.73	0.70	0.86	0.51	0.72
MLP		0.74 (0.63–0.85)	0.75	0.67	0.85	0.51	0.72

Ada indicates AdaBoost; AUC, area under the receiver operating characteristic curve; DT, decision tree; KNN, k‐nearest neighbor; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron; NB, naïve Bayes; NPV, negative predictive value; PPV, positive predictive value; RF, random forest; and SVM, support vector machine.

Figure 3

ROC curves of the LR and ML analysis. (A), Training cohort of the maternal model; (B) validation cohort of the maternal model; (C) training cohort of the neonatal model; (D) validation cohort of the neonatal model. Ada indicates AdaBoost; AUC, area under the receiver operating characteristic curve; DT, decision tree; KNN, k‐nearest neighbor; LR, logistic regression; MLP, multilayer perceptron; NB, naïve Bayes; RF, random forest; ROC, receiver operating characteristic; and SVM, support vector machine.

Clinical Use

The decision curve analysis for the maternal and neonatal events models are presented in Figure S4. The maternal model showed a positive net benefit when predicted probability thresholds were between 0.05 and 0.92, and the neonatal model showed a positive net benefit when the predicted probability thresholds were between 0.16 and 0.90.

Discussion

This study established 2 prenatal assessment models to predict adverse events in both mothers and offspring, which are particularly suitable for clinical use in developing countries where prepregnancy counseling and pregnancy monitoring systems are deficient. The developed nomogram lists allow the model to be used conveniently in clinical practice. The ML‐based algorithms achieved high prediction accuracy, which suggests the effectiveness of these ML algorithms as well as the strong association between high‐risk predictors and adverse maternal and neonatal events.

Ideally, women with CHD should seek counseling about pregnancy during the pubertal years. The advice given usually includes the importance of pregnancy planning, effective contraception options, and the impact of pregnancy on maternal heart diseases. However, in developing countries, because of the lack of complete health monitoring systems, many patients do not obtain a prepregnancy evaluation. For example, in our study, 216 (68%) patients visited our hospital for the first time and gave birth there, and 114 (36%) patients were unaware of their clinical history of CHD until the time of delivery. For patients with poor cardiac function and a first visit to a hospital, emergency assessment before delivery is critical for both doctors and patients.

Three popular risk assessment criteria are commonly used, including the Zwangerschap bij Aangeboren HARtAfwijkingen I, Cardiac Disease in Pregnancy, and World Health Organization classification systems. In a prospective study by Balci et al, for assessing the cardiovascular events of 203 women with CHD, the AUC was 0.57 (95% CI, 0.43–0.70) for the Cardiac Disease in Pregnancy risk score and 0.71 (95% CI, 0.59–0.83) for the Zwangerschap bij Aangeboren HARtAfwijkingen I risk score; the World Health Organization classification was the best risk assessment model for maternal cardiovascular events (AUC=0.77; 95% CI, 0.67–0.87); for the prediction of adverse events in offspring, there was no functional differentiation among the models. The adequate and accurate prenatal prediction of maternal and descendant risk is crucial for the counseling and management of pregnant women with CHD. The proposed models in our study have high prediction accuracies. Therefore, the 2 models can help cardiologists and obstetricians identify high‐risk pregnant patients with CHD.

The risk assessment studies for pregnancy with CHD are shown in Table S6. In our study, adverse maternal events were observed in 12.9% of pregnancies. International research on the event rate ranges from 4.0% to 23.5%. , , , , , , , , , , Heart failure and arrhythmias requiring treatment were the 2 most common adverse cardiac complications, which is consistent with results from previous research. In this study, women with CHD had a markedly higher risk of death during childbirth (n=13; 4.1%), and this proportion was higher than that reported from other authors, , , , , , , , , , mainly because of the lack of prenatal assessment and close monitoring during the pregnancy period, especially for patients with ES. Engelfriet et al found that 20% of patients with CHD died over a 5‐year follow‐up period, and patients with ES had a higher frequency of major bleeding events and associated right ventricular dysfunction. The European Society of Cardiology and the European Respiratory Society acknowledge that pregnancy is associated with higher mortality in patients with PAH, that the patient should be kept informed of the high risk associated with pregnancy, and that the termination of pregnancy should be discussed.

The adverse neonatal event rate was 29.2% in our study, which is consistent with previous studies that reported an incidence of 7.4% to 37.3%. , , , , , , , , , Because of the limitations of medical progress in China, the survival rate of fetuses who do not reach 28 weeks is much worse than that of fetuses in developed countries. Moreover, because of the financial burden, many families could not afford the cost of treatment in the neonatal intensive care unit, further increasing neonatal mortality.

NYHA functional class III/IV served as an independent predictor of adverse maternal events in patients with CHD in our study. In a retrospective study of 1302 completed pregnancies (>20 weeks of gestation) in patients with CHD, Drenthen et al proposed NYHA functional class as an independent high‐risk predictor of maternal cardiac complications. In addition, Cardiac Disease in Pregnancy and Zwangerschap bij Aangeboren HARtAfwijkingen I maternal cardiovascular and offspring risk scores also selected NYHA class III/IV as a predictor in pregnant women with CHD. The presence of PAH increases morbidity in patients with CHD, and the end of the spectrum of PAH in the setting of CHD is ES. In our study, a total of 13 patients died, 12 of whom were complicated with ES. At the same time, ES is also a high‐risk factor for adverse neonatal events. Left ventricular ejection fraction and SaO2 were independent predictors in patients with CHD, which was consistent with the finding of previous studies. , Therefore, cardiac function and oxygen saturation in patients with CHD are essential predictors of adverse events in pregnant women and their offspring. Preeclampsia was significantly correlated with neonatal adverse events in our study, and patients with CHD need to pay attention to the detection and treatment of pregnancy complications to reduce the occurrence of neonatal adverse events during pregnancy.

A logistic regression model is commonly used in the field of medical research, so we chose this method to construct predictive models and establish nomogram lists that are convenient for clinician use. ML has emerged as efficient computer algorithms for identifying patterns in large data sets with many variables and facilitating data‐driven prediction or categorical modeling. , In this study, ML algorithms were used to further train the models and verify the high‐risk factors in the 2 models. The 7 ML algorithms obtained similar predictive results in both the development cohort and validation cohort. Analysis of clinical data by ML methods offers considerable advantages for the evaluation of complex healthcare data.

Strengths and Limitations

The strengths of the study are as follows: First, we established the first prenatal risk assessment model for pregnant women with CHD during late pregnancy; second, the 2 prenatal assessment models have good accuracy and can be effectively applied to the assessment of adverse events in the last trimester of pregnancy in patients with CHD, especially in developing countries that lack well‐developed preconception counseling and pregnancy monitoring systems; third, in clinical practice, the application of prediction models can significantly improve the poor maternal and infant prognosis of pregnant women with CHD. This study has some limitations, one of which might be a potential bias caused by the small sample size in the development cohort. The rule of thumb is that the number of events per variable should to reach at least 10 for logistic regression modeling to ensure a small expected relative bias. However, Vittinghoff and McCulloch found that the requirement of events per variable could be relaxed to 5 to 9 in the context of confounder adjustment. In our study, the number of events per variable is 5.9 in maternal model development cohort and 31 in neonatal model development cohort. The sample size in our study meets the requirements. In addition, the nature of retrospective study is inevitably leading to the absence of previous medical history data such as cardiac surgery history details and medications during pregnancy. Increasing the sample size of patients in the development cohort, as well as to conduct prospective validation, will compensate for the above research limitations.

Conclusions

Two prenatal assessment models for mothers and offspring with reliable predictive accuracy were successfully established, which will benefit pregnant women with CHD worldwide, especially in developing countries where preconception counseling is inadequate. The proposed prediction models have benefits in helping clinicians to determine the patients at high risk of adverse events and provide a reference for clinicians’ management decisions. Assessing the accuracy of our prediction model in a prospective validation study is necessary before clinical use.

Sources of Funding

This work was supported by the National Key Technology R&D Program of China (grant number 2019YFC1005200 and 2019YFC1005204), the National Natural Science Foundation of China (grant number 81602286 and U1806202), the Taishan Scholar Youth Project of Shandong Province (grant number tsqn201812130), and the Key Research and Development Program of Shandong Province (grant number 2018GFS118202).

Disclosures

None.

Tables S1–S6

Figures S1–S4

Click here for additional data file.