Development a nomogram for predicting HELLP syndrome in women complicated with gestational hypertension

Abstract

Objectives

The unpredictability of HELLP syndrome and the severe adverse outcomes for both mother and children make it especially important for us to seek predictive model. This study aimed to develop a clinically accessible prediction model for assessing the risk of HELLP syndrome progression in patients with hypertensive disorders of pregnancy (HDP) and find effective factors that may predict the progression of HELLP within 3 days.

Methods

We used electronic data from Women’s Hospital, Zhejiang University School of Medicine, between January 1,2014 and December 31,2023. A total of 808 patients were included in this study, including 607 patients in the non-HELLP syndrome group and 201 patients in the HELLP syndrome group. We collected clinical and laboratory information, and conducted single- and multiple-factor logistic regression analyses to identify independent factors influencing the occurrence of HELLP syndrome and the onset of HELLP syndrome within 3 days. A nomogram was constructed based on these predictors to provide a visual representation of risk estimation. The model’s performance was evaluated through internal and external validation, with metrics such as the area under the curve(AUC), receiver operating characteristic curve (ROC), precision, recall, and F1 score. Calibration and decision curve analyses were also performed to assess model robustness and clinical utility.

Results

Multiple logistic regression analysis indicated prenatal BMI, neurologic symptoms, other system symptoms, 24-h urine protein, lowest SBP at admission, lowest DBP at admission, prenatal albumin, prenatal platelet and prenatal blood urea nitrogen as independent factors of HELLP syndrome. The prediction model achieved an AUC of 0.975 (95% CI: 0.966–0.985) in the internal validation dataset with a sensitivity of 0.962(95% CI: 0.962–1.000) and specificity of 0.885(95% CI: 0.962–1.000). The AUC of the external validation dataset was 0.838 (95% CI: 0.785–0.892). The optimal cutoff value calculated using Youden’s index was 0.613, with a sensitivity of 0.891(95% CI: 0.473–0.836) and specificity of 0.722(95% CI: 0.667–0.818). In multivariate regression analysis, blood urea nitrogen and the creatinine-to-blood urea nitrogen ratio were significant predictors in predicting HELLP syndrome within 3 days. The sensitivity was found to be 0.68 and 0.65, specificity to be 0.74 and 0.686 respectively.

Conclusions

This study successfully developed and validated a prediction model that can reliably predict the risk of HELLP syndrome in HDP patients. And blood urine nitrogen and the ratio of creatinine over blood urea nitrogen could be efficient predictors of HELLP syndrome occurring within 3 days.

Background

Hypertensive disorders of pregnancy (HDP) are a leading cause of maternal and perinatal morbidity and mortality, affecting approximately 3%–5% of pregnancies worldwide [1]. Among these disorders, preeclampsia can progress to severe complications, including HELLP syndrome, a life-threatening condition characterized by hemolysis, elevated liver enzyme levels, and low platelet counts. HELLP syndrome occurs in 0.5% to 0.9% of all pregnancies and accounts for 10% to 20% of severe preeclampsia cases [2]. It can be life-threatening, with reported maternal mortality rates as high as 24% [3]. The clinical presentation of HELLP syndrome is complex and variable, often manifesting rapidly with right upper quadrant pain, hypertension, proteinuria, general malaise, nausea, and vomiting [4]. However, in some cases, the onset is insidious, with hypertension and proteinuria absent in up to 15% of patients [5]. These atypical presentations contribute to a high misdiagnosis rate, making early identification and intervention challenging.

In clinical practice, for pregnant women with HDP, the management approach depends on the severity and gestational age. If preterm and clinically stable, expectant management with close monitoring is often preferred to optimize fetal maturity [6]. However, once HELLP syndrome develops, immediate delivery is typically required to prevent severe maternal and fetal complications. Given this distinction, accurately predicting which HDP patients will progress to HELLP syndrome is critical for timely intervention and risk stratification. This highlights the need for a predictive model that can assist in early identification and clinical decision-making.

To date, many researchers have attempted to predict the onset of HELLP syndrome. Harun Egemen Tolunay found the first-trimester aspartate aminotransferase (AST) to platelet ratio index (APRI) might serve as a predictor of HELLP syndrome [7]. Göksun İpek pointed that systemic immune-response index (SIRI) and other inflammatory indices could also serve as potential predictive markers [8]. Hromadnikova identified microRNAs that were upregulated in the first trimester of pregnancies that later developmed HELLP syndrome, demonstrationg thier high predictive potential [9]. Additionally, some studies have explored potential genetic biomarkers for HELLP syndrome using massive parallel sequencing [10]. Although these studies provide valuable insights into early risk identification, most focus on determining whether a pregnant individual will develop HELLP syndrome based on potential biomarkers. However, fewer studies have addressed the progression from hypertensive disorders of pregnancy (HDP) to HELLP syndrome, despite HELLP being one of the most severe complications of HDP. Cecilia Villalain developed a machine learning-based model to predict the onset of HELLP syndrome in patients with preeclampsia [11]. Malte also found that C-terminal pro-endothelin-1(CT-pro-ET-1), soluble FMS-like tyrosine kinase- 1(sFlt-1)could predict whether PE would develop into HELLP syndrome within one week and two weeks with AUC 0.94 and 0.83 [12]. However, these methods have limitations in terms of widespread applicability and cost-effectiveness, restricting their clinical utility. Therefore, there is an urgent need for more accessible, efficient, and cost-effective predictive markers to identify high-risk individuals with HDP who are at risk of developing HELLP syndrome.

Therefore, our objective is to develop a predictive model for HELLP syndrome onset in women diagnosed with HDP using more universal predictors and validate it in an external population. Additionally, we aim to identify key characteristics of individuals at risk for its rapid development.

Methods

Study population

The retrospective cohort study involved patients complicated with gestational hypertension including gestational hypertension, preeclampsia, eclampsia, chronic hypertension with superimposed preeclampsia, pregnancy complicated by chronic hypertension and HELLP syndrome at Women’s Hospital, Zhejiang University School of Medicine, between January 1,2014 and December 31,2023. According to whether progressed to HELLP syndrome, all included patients were divided into the non-HELLP syndrome group and HELLP syndrome group. The study was approved by the ethical committee of Zhejiang University (IRB- 20230423-R), and the requirement for informed consent was waived. All the information was obtained from the hospital's electronic database.

External validation dataset: a total of 198 (78.26%) non-HELLP syndrome patients and 55 (21.74%) HELLP syndrome patients who visited Lishui Maternity and Child Health Care Hospital and The First People’s Hospital of YongKang from January 2020 and December 2022 were included for external validation.

Inclusion criteria

(1) Age ≥ 18 years old. (2) all relevant information of patients was available. (3) Patients discharged with a diagnosis of “HDP” at admission and “disease progression form HDP to HELLP syndrome”.

The diagnosis criteria for gestational hypertension including gestational hypertension, preeclampsia, eclampsia, chronic hypertension with superimposed preeclampsia, pregnancy complicated by chronic hypertension was suggested by the American College of Obstetricians and Gynecologists (ACOG) reported in 2020 [5].

HELLP syndrome was defined as follows: platelet count less than 100,000/mm3, aspartate aminotransferase(AST) levels greater than 70 U/L, and l-lactate dehydrogenase(LDH) levels greater than or equal to 600 U/L according to Tennessee classification [13].

Exclusion criteria

(1) Diagnose HELLP syndrome directly at admission. (2) Related information was incomplete. (3) secondary hypertension(such as Cushing's syndrome, primary aldosteronism, chronic kidney disease, and so on). (4) Other obstetric emergencies (such as amniotic fluid embolism, intrahepatic cholestasis of pregnancy, acute fatty liver of pregnancy, and so on). (5)Patients diagnosed as multiple organ injury, severe infection were excluded in our study.

General information collection

In our study, the general information included the different types of gestational hypertension(gestational hypertension, preeclampsia, eclampsia, chronic hypertension with superimposed preeclampsia, pregnancy complicated by chronic hypertension), age, height, prepregnancy and premartum body mass index(BMI), hightest and lowest systolic blood pressure(SBP) at admission, hightest and lowest diastolic blood pressure(DBP) at admission, gravida, parity, assisted reproductive technology(ART), twin pregnancy, recurrent spontaneous abortion(RSA), inevitable fetal dismiss, preterm birth or stillbirth, placental dysfunction, immune disorders before pregnancy, neurological symptom or other systemic dysfunctions, hypothyroidism, hyperthyroidism, 24 h urine protein quantification, medication use during pregnancy: aspirin, low molecular weight heparin, hydroxychloroquine or prednisone. Hemogram, biochemistry and full urinalysis tests before and after delivery are requested and recorded, and we obtained these data from our hospital electronic records. Pregnancy outcomes included gestational diabetes, placenta previa, abnormal umbilical cord blood flow, oligohydramnios, fetal growth restriction(FGR), fetal distress, fetal dismiss, placental abruption, postpartum HELLP and postpartum hemorrhage. Neonatal outcomes included gestational age, gender, delivery mode, birth weight, admission to the neonatal intensive care unit (NICU), APGAR score at 1 and 5 min.

Statistical analysis

All statistical analyses were performed using R software (version 4.4.2). Continuous variables were presented as mean ± standard deviation (SD) or median with interquartile range (IQR), depending on the data distribution. Categorical variables were expressed as frequencies and percentages. Differences between groups were assessed using the Student’s t-test or Mann–Whitney U test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables. The deletion method was applied to handle missing values.

To construct the predictive model, multivariate logistic regression analysis was performed. Model performance was evaluated using the area under the curve(AUC), receiver operating characteristic curve (ROC) along with sensitivity, specificity. Additionally, precision, recall, and the F1 score were calculated to assess model performance comprehensively. Calibration was assessed using the Hosmer–Lemeshow goodness-of-fit test and a calibration curve. Precision-recall curve was to test involving sample. The nomogram was constructed based on the final logistic regression model.

To ensure model robustness, K-fold cross-validation (K = 10) was employed, and key metrics such as the root mean square error (RMSE), mean absolute error (MAE), and R-squared values were calculated for both the training and validation datasets. Decision curve analysis (DCA) was performed to evaluate the model’s clinical utility by assessing net benefit across different threshold probabilities. Statistical significance was defined as p < 0.05.

Results

Comparison of data between HELLP group and non-HELLP group

A total of 808 patients were included in this study, including 607 patients in the non-HELLP syndrome group and 201 patients in the HELLP syndrome group. The AUC of precision-recall curve is 0.938(showed in supplementary material). We believe this stability mitigates concerns related to sample imbalance.Demographic details, pregnancy indexes and neonatal outcomes are shown in Tables 1, 2 and 3. The comparison of general data between the two groups indicated that, there was a statistically significant difference in terms of hight, gravida, highest SBP, lowest SBP, SBP variability, highest DBP, lowest DBP, DBP variability, neurological symptoms, other systemic symptoms, 24 h urine protein, absent or reversed umbilical artery flow, FGR, fetal distress, fetal dismiss, placental abruption, postpartum HELLP, postpartum hemorrhage, gestational age at birth, mode of delivery, birth weight, APGAR scores, and NICU admissions (P < 0.001) (Tables 1 and 2). In the aspect of laboratory tests, there was a significantly statistical difference in terms of albumin(ALB), platelet(PLT), alanine aminotransferase(ALT), aspartate aminotransferase(AST), blood urea nitrogen(BUN), creatinine, fibrinogen, total bilirubin(TBil), direct bilirubin(DBil) before and after delivery between HELLP group and Non HELLP group (P < 0.001) (Table 3).

Table 1 Clinical characteristics for HELLP-group and non-HELLP-group

Table 2 Neonatal-Obstetric outcomes for HELLP-group and non-HELLP-group

Table 3 Laboratory parameters for HELLP-group and non-HELLP-group

Multivariate logistic regression analysis

By multivariate logistic regression analysis, only prepartum BMI ≥ 24(kg/m^2) (OR:0.259(0.069–0.977), P = 0.046), lowest SBP at admission (OR: 1.103(1.053–1.156), P < 0.001), lowest DBP at admission (OR: 1.067(1.007–1.129), P = 0.027), patients with neurological symptoms (OR: 3.693(1.246–10.941), P = 0.018), patients with other systemic symptoms (OR: 5.475 (1.650–18.165), P = 0.005), urine protein (OR: 1.178 (1.039–1.335), P = 0.01), prenatal ALB (OR: 0.851(0.759–0.954), P = 0.006), prenatal PLT (OR: 0.964 (0.954–0.975), P < 0.001), prenatal BUN (OR: 1.328(1.019–1.732), P = 0.036) were risk factors for HELLP syndrome, seen in Table 4.

Table 4 Multifactorial analysis of HELLP occurrence

Internal validation of the prediction model

The independent influencing factors, including prepartum BMI ≥ 24, lowest SBP at admission, lowest SBP at admission, lowest DBP at admission, patients with neurological symptoms, patients with other systemic symptoms, urine protein, prenatal ALB, prenatal PLT, prenatal BUN, obtained from the multivariate logistic regression analysis were introduced into R software for model evaluation. According to the ROC curve analysis based on the predicted probability of the model, the AUC was 0.975(95% CI: 0.966–0.985) (Fig. 1), indicating good discrimination ability and suggesting good predictive performance of the model. The optimal cutoff value was calculated using Youden's index, which was 0.172 with a sensitivity of 0.962(95% CI: 0.9624–1.000) and a specificity of 0.885(95% CI: 0.944–0.985) and a Youden's index of 0.847. Our prediction model demonstrated strong performance across multiple metrics. The model achieved a precision of 0.887, a recall of 0.812 and an F1 score of 0.848, indicating a robust balance between positive predictive value and sensitivity.The Hosmer–Lemeshow goodness-of-fit test was used to assess the fitting degree of the model, and the test statistic was 5.363 with a P value of 0.718, suggesting good agreement between the predicted and actual outcomes, thereby confirming the model’s appropriate fit. The calibration curve (solid line in Fig. 2) closely aligned with the ideal curve (dashed line in Fig. 2), indicating satisfactory predictive performance and strong reproducibility.

Fig. 1

The ROC Curve of Training Set

Fig. 2

The Calibration Curve of Training Set

External validation of the prediction model

The prediction factors prepartum BMI ≥ 24, lowest SBP at admission, lowest SBP at admission, lowest DBP at admission, patients with neurological symptoms, patients with other systemic symptoms, urine protein, prenatal ALB, prenatal PLT and prenatal BUN in the external validation dataset were entered into the prediction model to calculate the predicted probability, and the ROC curve was plotted based on the predicted probability (Fig. 3). The AUC was 0.838 (95% CI: 0.785–0.892). The optimal cutoff value was calculated using Youden's index, which was 0.526 with a sensitivity of 0.891(95% CI: 0.473–0.836) and a specificity of 0.722(95% CI: 0.667–0.818) and a Youden's index of 0.613. The model exhibited a precision of 0.462, a recall of 0.891 and an F1 score of 0.609, reflecting a notable strength in identifying true positive cases with minimal false negatives. The results indicate that the model has robust predictive ability, particularly in identifying high-risk patients, and demonstrates promising generalizability when applied to external validation data.

Fig. 3

The ROC Curve of Validation Set

K-Fold Cross-Validation Results

To assess the stability and generalizability of our prediction model, we performed K-fold cross-validation. The mean RMSE for the training set was 0.341 (SD = 0.056), MAE was 0.243 (SD = 0.045) and the mean R-squared was 0.424 (SD = 0.213). For the validation set, the mean RMSE was 0.238 with an SD of 0.045, MAE was 0.110 (SD = 0.027) and the mean R-squared was 0.706 (SD = 0.085). These results indicate that the model maintains robust predictive performance across different data partitions(showed in supplementary material).

Nomogram prediction of HELLP syndrome

We imported prepartum BMI ≥ 24, lowest SBP at admission, lowest SBP at admission, lowest DBP at admission, patients with neurological symptoms, patients with other systemic symptoms, urine protein, prenatal ALB, prenatal PLT and prenatal BUN into the R software to develop the nomogram for the prediction model (Fig. 4). A total score of 300 was calculated by summing the scores for each factor. The final model was Y = − 6.375 + 0.082*lowest SBP + 0.059*lowest DBP + 0.117 * 24 h urine protein + 1.752*other system symptoms (yes = 1,no = 0) + 0.952*neurological symptoms(yes = 1,no = 0)− 0.177*prenatal ALB- 0.029*prenatal PLT + 0.323*prenatal BUN − 1.513*prenatal BMI (less than 24,yes = 1,no = 0). The DCA was also plotted (Fig. 5).

Fig. 4

The Nomogram of the HELLP Syndrome prediction model

Fig. 5

The DCA Curve of Training Set

Risk factors for within- 3-day occurrence of HELLP Syndrome

A key challenge in clinical practice is identifying which patients are likely to progress to HELLP syndrome in a short time. In our study, we divided our HELLP syndrome group into two different group according to the time of HELLP onset. Multivariate regression analysis indicated that BUN and the ratio of creatinine over BUN were significant predictors of HELLP syndrome onset within three days (Table 5). The sensitivity was found to be 68% and 65%, specificity to be 74% and 68.6% respectively (Fig. 6). The cutoff for BUN and the ratio of creatinine over BUN was 5.635 mmol/L and 12.544. The AUC of BUN was 0.673 (95% CI: 0.585–0.761). The optimal cutoff value was calculated using Youden's index, which was 5.635 with a sensitivity of 0.68 and a specificity of 0.74 and a Youden's index of 0.42. The AUC of the ratio of creatinine over BUN was 0.731 (95% CI: 0.650–0.0.813). The optimal cutoff value was calculated using Youden's index, which was 12.544 with a sensitivity of 0.65 and a specificity of 0.686 and a Youden's index of 0.336.

Table 5 Multifactorial analysis of HELLP occurrence within 3 days

Fig. 6

The ROC Curve of BUN and the ratio of creatinine over BUN

Discussion

HELLP syndrome is a life-threatening complication of HDP that necessitates immediate delivery, making early prediction crucial for maternal and fetal outcomes. While identifying those at high risk of developing HELLP syndrome remains a challenge [6]. Therefore, our study aimed to address this gap by developing a predictive model based on clinical and laboratory parameters, providing an accessible tool for early risk stratification.

In this study, we successfully developed and validated a predictive model and a corresponding nomogram to estimate the risk of HELLP syndrome in patients with HDP. Our model integrates common clinical and laboratory parameters that are readily available in routine clinical practice, making it highly practical for real-world use. The model demonstrated robust performance across multiple metrics, with an AUC of 0.975 in the internal validation dataset and 0.882 in the external validation dataset. These results indicate strong discrimination ability, while the Hosmer–Lemeshow test confirmed good calibration (P = 0.718), ensuring reliable predictive accuracy.

Once a pregnant woman is diagnosed with HDP, dynamic monitoring——including blood pressure, complete blood count with platelet estimate, serum creatinine, LDH, AST, ALT, and proteinuria is necessary to track disease progression. Our study specifically focused on the progression from HDP to HELLP Syndrome, making our predictive model particularly suitable for the management of preterm HDP patients. By predicting the probability of HELLP Syndrome, the safety of pregnant woman is ensured to a greater extent and the adverse effects of premature birth on the fetus are reduced. Given that HELLP syndrome management requires access to neonatal and obstetric intensive care units (ICUs) and specialized personnel, our model can help identify high-risk patients early, allowing timely referral of those in remote areas to appropriate medical centers [14]. Several of our predictors have been previously explored. Muhammed’s study identified some similar risk factors but did not establish a predictive model [15]. Zhaoqi Li established a predictive model for PE complicated with HELLP syndrome, including aspartate-aminotransferase to platelet ratio index(APRI), mean artery pressure(MAP), fibrinogen degradation products(FDP), cholinesterase(CHE), and serum calcium(Ca). However, their study only included patients with gestational hypertension or preeclampsia [16]. Compared to Malte’s study that CT-pro-ET- 1, sFlt- 1 and systolic blood pressure could predict whether PE would develop into HELLP syndrome within one week and two weeks with AUC 0.94 and 0.83 [12]. Our model is more practical, cost-effective, and accessible, especially for use in remote or resource-limited settings.

What’s more, additional multivariate regression analysis indicated that BUN and the ratio of creatinine to BUN were significant predictors of the occurrence of HELLP syndrome within 3 days with cutoff values of 5.635 mmol/L for BUN and 12.544 for the creatinine-to-BUN ratio. The sensitivity was 68% and 65%, while specificity was 74% and 68.6%. Notably, our study is the first to identify these markers as potential predictors of HELLP syndrome progression. This finding aligns with previous research suggesting that renal dysfunction plays a critical role in HELLP syndrome pathophysiology [17]. HELLP syndrome patients’ renal histopathology showed thrombotic microangiopathy (TMA) that coexisted with acute tubular necrosis (ATN) and acute renal cortical necrosis (ARCN) [18]. HELLP syndrome is characterized by microangiopathic hemolysis and endothelial dysfunction, which can lead to renal ischemia and impaired glomerular filtration, resulting in elevated BUN levels [19]. Additionally, intravascular volume depletion due to vasospasm and endothelial leakage may further alter the BUN/Cr ratio, making it a potential early biomarker for HELLP progression.

Based on our predictive model, we developed a software tool to assist obstetricians in calculating the probability of HELLP syndrome(uploaded as Supplementary Software). For example, a 33-year-old patient(height:160 cm, prenatal BMI:22), experiencing her first pregnancy with twin gestation, presented with lowest blood pressure 144/71 mmHg, blurred version and 24-h urine protein of 1.56 g. she had no neurological symptoms, but laboratory tests showed ALB(25.2 g/L), PLT(66 * 10^9/L), BUN(4.67 mmol/L). The model calculated a 0.9944 probability of developing HELLP syndrome. Furthermore, her creatinine-to-BUN ratio was 15.41, exceeding our cutoff value, strongly suggesting imminent onset of HELLP syndrome. Our findings support the predictive utility of this nomogram, which provides an intuitive and clinically applicable method for estimating HELLP syndrome risk.

Although the sample size of our study was the largest among the current studies, our study also has some limitations. Firstly, First, the retrospective nature of the data introduces potential biases, despite efforts to minimize these effects through comprehensive data collection and rigorous analysis. Second, since the data were derived from a single tertiary hospital, its applicability to diverse patient populations remains to be validated. Third, some clinical features in our model rely on patients’ subjective symptoms, which may impact the accuracy of model. In future research, we could integrate advanced machine-learning techniques to further enhance predictive efficiency and validate our model through multi-center, prospective cohort studies.

Conclusions

In conclusion, our study presents a predictive model with strong discrimination, calibration, and clinical applicability for identifying patients with HDP at risk of developing HELLP syndrome. Additionally, we identified BUN and the creatinine-to-BUN ratio as potential early indicators of HELLP syndrome within 3 days. This model, implemented in a software tool, may aid in early risk stratification and clinical decision-making. However, validation of multi-center, prospective cohort studies are needed to enhance its applicability.