Introduction

The hypertensive disorders of pregnancy (HDP) is observed in 5–10% of antenatal population [1]. Hypertension in pregnancy has been associated with perinatal mortality and morbidity and is a major contributor to stillbirth, particularly in low- and middle-income countries(LMIC) [2]. World Health Organization South-East Asian Region (WHO SEAR) started a newborn-birth defects (SEAR-NBBD) database in 2014. Simultaneously the project involved keeping a database of stillbirth as well. The analysis of the data generated from the study showed that HDP was the most common maternal condition associated with stillbirth [3]. Most of the morbidity could be prevented by recognizing the early warning signs associated with the disease, leading to adequate monitoring and early delivery in HDP cases and thus preventing stillbirth. Previous studies have examined the associations between stillbirths and clinical and non-clinical characteristics such as late or no antenatal care, inadequate management of anemia, infection, diabetes, and fetal growth restriction [3,4,5]. The knowledge generated upon various studies on the subject would have a positive impact on the prevention of stillbirths if these were applied in promoting targeted interventions and closer monitoring by the health care providers. Hence a user-friendly clinical decision-making tool for the identification of such patients might be beneficial in prioritization of care which is particularly relevant in low resource settings.

Previous studies have used the clinical and biochemical parameters which involved biomarkers and Doppler for the prediction of stillbirth among low-risk population [6,7,8]. The biomarkers like PAPP-A, PlGF are costly and Doppler is often not available in areas where the stillbirths mostly happen, therefore, they are not suitable for LMICs, thus, using clinical and socioeconomic parameters would be more pragmatic in our environment [9]. Therefore the study aimed to develop an easy-to-use tool using clinical and biophysical parameters for the prediction of risk of late stillbirth in women with hypertensive disorders of pregnancy.

Materials and Methods

The study was an observational, analytical case–control study performed from November 2015 to December 2020 at a single tertiary care center after ethical clearance, it was as a part of the WHO SEARO, NBBD stillbirth project. All babies delivered after 20 weeks of gestation showing no sign of life after birth were considered stillborn [10]. All women with HDP having a stillborn baby at 28 weeks or more were considered late stillbirths and were included in the study. The age and parity matched women with hypertensive disorders of pregnancy having live birth were taken as controls.

A detailed maternal history with special reference to high risk factors for stillbirths in the present and previous pregnancies was elicited. The period of gestation was calculated according to the last menstrual period (LMP) or first-trimester ultrasound if the LMP was not known. General physical and systemic examination of the mother, time of admission to the hospital was noted. Antenatal records were reviewed to rule out any abnormal clinical findings. All routine investigations along with an ultrasound for confirmation of intrauterine death were done. The expected fetal weight of the babies was estimated by ultrasound using biparietal diameter, abdominal circumference and femur length using Hadlock formula (Aloka, prosound alpha, Hitachi, Japan). The women with HDP having congenital anomaly in fetus were excluded from the study. Hypertensive disorder of pregnancy (HDP) was taken as any rise of blood pressure 140/90 mm Hg on two occasions four hours apart. Preeclampsia was defined according to ACOG 2013 guidelines [1]. Early-onset hypertension was termed as those which occurred before 34 weeks of gestation. Fetal growth restriction was considered when the birth weight was less than the tenth centile for the gestational age according to the Intergrowth 21 chart [11].

The cases and controls were compared using the chi-square test. The univariate odds ratio was calculated along with 95% confidence limits. Further, Multivariate logistic regression was used to identify the impact of all significant risk factors on the outcome. The best model was attained using the Forward LR model selection criteria. The threshold of predicted probabilities for distinguishing stillbirth and live birth was obtained using the ROC curve. A formula was created using the threshold value of 50% obtained under the ROC curve to predict the outcome variable. A p-value < 0.05 was considered statistically significant. IBM SPSS version 20.0 software is used for statistical analysis.

Results

In the study period of five years and one month, there were 69,007 deliveries at the institution, out of which 1691 (24.5/1000) were stillborn. HDP was observed as the most common maternal cause (390/1691, 23.0%), it occurred at or after 28 weeks of gestation in 265/390 (67.4%) cases and hence included as cases in the study. For comparison 288, age and parity matched women with HDP having live birth were selected as controls.

The epidemiological profile of cases and controls are given in Table 1. The maternal age ranged from 19 to 40 years, the mean age was 27.14 ± 4.61 in cases. As the cases and controls were age and parity matched, there was no significant difference between the two groups (P > 0.05). There were significantly more women with previous two or more abortions among cases (31/265, 11.7%) compared to controls (19/288, 6.6%) (p = 0.038). There were considerably more subjects with poor prenatal care among cases (128/265, 48.3%) compared to controls (43/288, 14.9%) (p < 0.001). Although there was no significant difference in the family income between cases and controls (p = 0.849), there were significantly more women among cases who were either illiterate or studied only till the primary level (58.5% Vs 36.5%)). There was no significant difference in maternal weight and height between cases and controls. The mean systolic and diastolic blood pressure among cases was significantly higher than controls (p < 0.001). The delivery occurred significantly early in gestation among cases (mean 34.6 ± 3.3 weeks) compared to the controls (36.9 ± 3.4 weeks) (p < 0.001).

Table 1 The epidemiological characteristics of cases and controls
Full size table

Table 2 shows the details pertaining to history and clinical examination of the cases and controls. The women with a family history of hypertension had significantly increased chances of stillbirth (p = 0.007). The stillbirths were considerably more in chronic or gestational hypertensive mothers compared to those having preeclampsia (p = 0.013), and the antihypertensives were given prior to stillbirth in 123/265 (46.4%) cases compared to 156/288 (54.2%) controls, however the difference did not reach statistical significance (P = 0.074). Other associated co-morbidities such as deranged blood sugar and anemia were more common among cases, but the difference was not statistically significant (p > 0.05). Known case of antiphospholipid syndrome was two each case and controls. Almost half of the cases of preeclampsia had associated complications. Abruption was the most common complication (19.6%), followed by eclampsia (3.8%), these complications were considerably more in cases compared to controls (p = 0.042), there was no significant difference in deranged renal function test among cases and controls.

Table 2 Relevant history and clinical features with associated complications in cases with hypertensive disorders of pregnancy having stillbirth (cases) and live birth (controls)
Full size table

To find out about the degree to which each factor contributed to the stillbirth, the multivariate regression analysis was done and the odds ratio was calculated using binary logistic regression with forward-LR model selection criteria. The best model was obtained in 6 steps wherein explanatory power increased from 27.5 to 45.8% (p < 0.001) (Table 3). Also, the concordance between observed and model's stillbirth outcomes increased from 74.5 to 80%. With the fetal weight less than 2000gms, the odds in favor of the stillbirth were 10.3 times that of live birth (p < 0.001). Due to poor antenatal care, the odds in favor of stillbirth were 5.9 times that of live birth (p < 0.001). If there was a family history of hypertension, the odds in favor of stillbirth were 4.4 times that of live birth (p = 0.018). In the presence of abruption, the odds in favor of stillbirth were 3.4 times that of live birth (p = 0.002). Due to gestational hypertension, the odds in favor of stillbirth were 2.2 times that of live birth (p = 0.001). If the woman was illiterate or had education up to primary level only, the odds in favor of the stillbirth were 1.7 times that of live birth (p = 0.024). The ROC which was constructed showed the composite AUC of all factors to be 0.846 ( 95% CI 0.755–0.835) which was in the region of good prediction(Fig. 1). The fitted equation is given by:

$$ P\left[ {Still \, birth} \right] \, = \frac{1}{{1 + e^{2.550 - 1.212*Abruption - 0.800*Gest.Htn - 1.49*FamilyHist.Htn - 0.539*Education - 1.767*NoANCvisits < 4 - 2.331*FetalWt < 2000} }} $$
Table 3 The Logistic Regression being applied to the variables to find out their odds of resulting in stillbirth in women with hypertensive disorders of pregnancy
Full size table
Fig. 1
figure 1

The Receiver operating curve showing the AUC of all combined predictor probabilities

Full size image

Using the above equation a software using Microsoft excel sheet was created. It was modeled to work as software where the six variables could be clicked as 'yes' or 'no' options. The result would flash as chances of live birth or stillbirth along with the probability in percentage (Fig. 2, excel sheet). The prediction of stillbirth using the software had a sensitivity of 80.3% and specificity of 70.0%.

Fig. 2
figure 2

Figure shows the predictive value of the probability of stillbirth based on the presence or absence of variables in the prediction model

Full size image

Discussion

To the best of our knowledge, for the first time, a model for the prediction of the risk of stillbirth in women with HDP has been created. The model is simple, low cost, user-friendly, and can be used in clinical settings, using the parameters involving history, clinical features, and ultrasound biometry for fetal weight estimation. The model allows for early detection of pregnancies at high risk of stillbirth due to hypertension in pregnant women, by a physician or even a health worker and allocation of targeted interventions such as closer monitoring throughout the pregnancy or early delivery. This prioritization of care allocation based on the prediction by the model is particularly relevant in low resource settings.

The data from the last five years of research on the analysis of factors associated with stillbirth under the WHO SEARO project have shown that hypertension in pregnancy was associated with nearly one-quarter of the stillbirths [12], making it the most common maternal cause of stillbirths according to ICD -10 PM coding. In our previous study on the factors associated with stillbirth in hypertensive women, it was noted that 80.6% of them occurred after 28 weeks of gestation, thus, the babies were salvageable and could be saved by timely intervention [13]. Therefore, it was pertinent that the factors associated with salvageable babies be pondered upon to find out ways to prevent them.

In the present prediction model six variables which contributed most to stillbirth on regression model, namely, less than four antenatal visits, gestational/ chronic hypertension, abruption, fetal weight less than 2000gms, education only to primary/ no education, and family history of hypertension were used in a modified excel sheet with options as "yes" and "no" against each. Although, abruption occurs as a late event, it was included with all other significant prenatal factors as its presence would alert the health care personnel regarding a very high risk of impending stillbirth if "yes" was clicked. If abruption would not be present "no" could be clicked.

To date, there have been very few studies regarding the development of a decision-making tool for prediction of the risk of stillbirth. Moreover, these models involved estimation of biomarkers and doppler studies which are difficult to apply in the low-resource settings [6, 7]. In the model developed by Akolekar et al., the prediction stillbirth was studied in low risk women by the maternal factors along with the estimation of biomarker PlGF levels and doppler in the first and second trimester of pregnancy, with the use of the maternal factors alone, 16% of all stillbirths could be predicted at a 5% false-positive rate (FPR), and with the addition of biomarkers, the prediction rate increased to one-third of all stillbirths at the end of the first trimester and nearly half at the end of second trimester [6,7,8]. The model developed in the present study is for women with hypertensive disorder of pregnancy in late second or third trimester.

In our study we found that the area under the receiver operating curve (AUROC) was in good prediction range (0.846). The sensitivity was as high as 80.3% at the specificity of 70.0%. In a study by Koivu et al., a model was developed using clinical parameters with the help of artificial intelligence, in the study the best-performing machine learning models achieved 0.76 AUC for early stillbirth, 0.63 for late stillbirth [14], in another study by Kayode et al., done in a low resource setting, using clinical parameters the discriminative performance of the final basic model was very good with a C-statistic of 0.80 (95% CI 0.78–0.83) [15]. Therefore, in the present study, the sensitivity specificity and the AUC were better than previous studies.

The major strengths of the study were its large sample size and a novel attempt to create a practically feasible model for predicting stillbirth in hypertensive pregnant women for a population that is the most vulnerable to this risk. The major limitation of the study could be the possibility of an overestimation of the performance of screening as it was derived and tested using the same dataset, therefore, consequently, the model needs validation from prospective studies.

The study demonstrates that nearly four fifths of deaths occurring in the study population could have been predicted using a simple, user friendly tool. This model would be particularly effective in low resource settings and would be useful in guiding the clinicians for active intervention and thus aid in its prevention.