Medicine in Novel Technology and Devices 15 (2022) 100158

Contents lists available at ScienceDirect

Medicine in Novel Technology and Devices

journal homepage: www.journals.elsevier.com/medicine-in-novel-technology-and-devices/

Full Length Article

Prediction of severe preeclampsia in machine learning

Xinyuan Zhang a, Yu Chen a, c, *, Stephen Salerno b, Yi Li b, Libin Zhou d, Xiaoxi Zeng c,
Huafeng Li e, **
a
Department of Applied Mechanics, College of Architecture and Environment, Sichuan Province Biomechanical Engineering Laboratory, Sichuan University, Chengdu,
610065, Sichuan, China
b
Department of Biostatistics, University of Michigan, Ann Arbor, MI, USA
c
Medical Big Data Center, Sichuan University, Chengdu, 610065, Sichuan, China
d
Department of Computer Science, University of Wisconsin, Madison, WI, USA
e
West China Second University Hospital, Sichuan University, Chengdu, 610041, China

A R T I C L E I N F O A B S T R A C T

Keywords: This study aimed to find out the blood data characteristics of patients and explore the correlation between severe
Preeclampsia preeclampsia and blood index value. Provide assistance for the early attention direction of severe preeclampsia
Screening diagnosis and treatment. 19,653 pregnant women presenting to the West China Second University Hospital,
Prediction
Sichuan University from January 2017 to April 2019. After screening, a total of 248 patients, 124 severe pre-
Blood examination
Data characteristics
eclampsia cases, and 124 controls were selected for this study. Forty-three blood examination variables were
obtained from routine blood work, hepatic, renal and coagulation function examination. Light gradient boosting
machine (light GBM), decision tree and random forest were used for date diving. We randomly divided 35% of the
original data as a testing set to conduct internal validation of the performance of the prediction model. The area
under receiver operating characteristic curve (AUC) was used as the main score to compare the three methods.
Finally, a binary classification light GBM model based on aspartate aminotransferase, direct bilirubin and acti-
vated partial thromboplastin time ratio can predict severe preeclampsia with sensitivity of 88.37%, specificity of
77.27%, AUC of 89.74% and positive predictive value of 65.96%. We believe relevant quantifiable indicators can
establish an effective prediction model, which can provide guidance for early detection and prevention of severe
preeclampsia.

1. Introduction combine markers such as uterine artery pulsatility index (PI) via uterine
artery Doppler, maternal blood pressure, parity, body mass index, and
Hypertensive disorders of pregnancy are one of the leading causes of placental growth factor [3–5]. For example, Goetzinger proposed a risk
maternal and neonatal morbidity, affecting 5–8% of all pregnant women. score with a reported area under the receiver operating characteristic
Roughly 76,000 pregnant women die each year from preeclampsia and (ROC) curve (AUC) of 0.76 [6]. For prediction of preeclampsia in the first
related hypertensive disorders. Preeclampsia is multifactorial, charac- trimester, one study found the neutrophil-to-lymphocyte ratio to be
terized by hypertension with characteristics of systemic endothelial significantly higher in patients with severe preeclampsia, indicating a
activation that begins after 20 weeks gestation. This confers a significant potentially useful laboratory marker [7]. Other research suggests that
impact on maternal and fetal health [1]. As opposed to mild pre- damage to placental circulation may lead to increased mean platelet
eclampsia, severe preeclampsia is characterized by higher blood pres- volume (MPV) [8,9]. In addition, low levels of pregnancy-associated
sures, leading to profound organ dysfunction. Severe preeclampsia plasma protein A in the first trimester are predictive of preeclampsia
constitutes the largest attributable fraction of severe morbidity due to [4,10,11]. Serum beta-trace protein has also been shown to be predictive
hypertensive disorders in pregnancy [2]. of preeclampsia with a diagnostic sensitivity of 91.3% and specificity of
Currently, methods of screening for preeclampsia have high diag- 89.5% [12]. One study found that these risk factors, in addition to family
nostic accuracy, even in the first half of one's pregnancy. These methods history and TIMP-4, a gene which codes for Metallopeptidase Inhibitor 4,

* Corresponding author. Department of Applied Mechanics, Sichuan University, Chengdu, 610065, Sichuan, China.
** Corresponding author.
E-mail addresses: [email protected] (Y. Chen), [email protected] (H. Li).

https://doi.org/10.1016/j.medntd.2022.100158
Received 7 November 2021; Received in revised form 5 July 2022; Accepted 17 July 2022
2590-0935/© 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
X. Zhang et al. Medicine in Novel Technology and Devices 15 (2022) 100158

were predictive [13]. Studies have also investigated the relationship Table 1
between serum C-reactive protein, triglyceride, and homocysteine levels Comparison of maternal characteristics and routine blood examination indices
with severity of hypertension and the incidence of complications in between severe preeclampsia cases (n ¼ 124) and controls (n ¼ 124) in the study
preeclampsia [14–16]. Others observed that the differential expressions sample (Indicator values with significant differences were highlighted in bold).
of Fas-related proteins in fetal membranes, decidua, and placentas were Variables SPE Cases (n ¼ Controls (n ¼ P Value*
consistent with increased apoptosis in pregnancy-induced hypertension 124) 124)
[17]. Lastly, differential serum total adiponectin and urine albumin Maternal characteristics indics
excretion has been noted in preeclampsia patients [18]. However, these Maternal age, year 33.04 ± 5.23 28.85 ± 3.28 <0.001
indicators are not frequently part of routine maternal health examina- Parity 0.651
tions. In China, health authorities recommend that pregnant women 0 94 (75.8%) 97 (78.2%)
establish care in their first trimester. Many of these women present to the 1 30 (24.2%) 27 (21.8%)
Gestational age, weeks 0.191
hospital in advance for a comprehensive physical examination or sec-
<15 97 (78.2%) 105 (84.7%)
ondary to nausea and vomiting from pregnancy. This study aims to 15 27 (21.8%) 19 (15.3%)
identify biomarkers of severe preeclampsia collected during these Body mass index (BMI), kg/m2 21.66 ± 2.89 19.79 ± 2.19 <0.001
routine prenatal visits in the first 10–15 weeks of pregnancy. Identifying <18.5 13 37
potential abnormalities in these indicators early on may provide the basis 18.5–24 89 82
24 22 5
for future effective interventions to prevent this disease in advance. Number of fetus <0.001
1 92 (74.2%) 124 (100%)
2. Materials and methods 2 32 (25.8%) 0 (0%)
Maternal complications
Thyroid disease 22 N/A N/A
2.1. Study design and sample
Diabetes mellitus 21 N/A N/A
Intrahepatic cholestasis of 15 N/A N/A
The data were collected from the medical records of pregnant women pregnancy
presenting to the West China Second University Hospital of Sichuan Pregnancy outcome <0.001
University between January 2017 and April 2019. Severe preeclampsia Adverse pregnancy outcomes 78 (62.9%) 1 (0.8%)
Normal pregnancy outcomes 46 (37.1%) 123 (99.2%)
cases were first identified. While diagnosis of severe preeclampsia usu-
ally occurs after 20 weeks of pregnancy, confirmed cases in this study Routine blood examination P
variable Value**
were identified during their first 10–15 weeks of pregnancy. Healthy
Routine blood work examination variables
pregnant women, defined as having no relevant comorbidities, were
selected from the same data as controls. In the context of China's family White blood cell count (WBC), 8.43  1.96 8.23  1.67 0.390
109/L
planning policy and serious conflicts between doctors and patients, the Percent neutrophils (NEUT%), % 74.24  5.52 74.63  4.79 0.550
serious dangers of preeclampsia may make Chinese obstetricians more Percent lymphocytes (LYMPH 19.38  4.93 19.59  4.40 0.719
cautious. According to China's pregnancy examination requirements, %), %
pregnant women will receive about 10 health services from 12 weeks to Percent monocytes (MONO%), 5.23  3.37 4.49  1.07 0.021
%
delivery. Which means, the data selected in this article are those gener-
Percent eosinophils (EOS%), % 1.00(0.90) 0.80(0.80) 0.069
ated by China's national conditions. In the diagnosis of preeclampsia, the Percent basophils (BASO%), % 0.20(0.18) 0.20(0.10) 0.337
diagnosis is made according to the international guidelines for related Neutrophil count (NEUT), 109/L 6.30  1.73 6.18  1.45 0.532
diseases. Monocyte count (MONO), 109/L 0.41  0.12 0.37  0.11 0.005
Table 1 report maternal characteristics, including, demographics, Lymphocyte count (LYMPH), 1.59  0.43 1.58  0.38 0.877
109/L
parity, gestational age, pregnancy outcome, maternal complications, and Eosinophil count (EOS), 109/L 0.08(0.08) 0.06(0.07) 0.077
number of fetus. Routine blood work, hepatic, renal, and coagulation Basophil count (BASO), 109/L 0.01(0.01) 0.02(0.01) 0.428
studies were recorded with each patient's medical history. These char- Red blood cell count (RBC), 4.19  0.37 4.10  0.35 0.079
acteristics were based on maternal medical records and broadly consisted 1012/L
Hemoglobin (HGB), g/l 127.47  10.14 125.71  9.68 0.164
of the following measured and calculated laboratory values. The specific
Hematocrit (HCT), % 37.22  2.91 36.40  2.66 0.022
indicators of each examination are shown in Table 1. Mean corpuscular volume 89.10  3.99 88.84  3.12 0.560
(MCV), fl
2.2. Diagnosis and classification Mean corpuscular hemoglobin 30.52  1.50 30.68  1.17 0.376
(MCH), pg
Mean corpuscular hemoglobin 342.52  7.28 345.29  7.70 0.004
Clinically, severe preeclampsia is defined as new onset hypertension concentration (MCHC), g/L
after 20 weeks gestation in the presence of proteinuria. No cutoff value Red cell distribution width-CV 13.59  1.06 13.53  0.70 0.597
for blood pressure is formally defined since treatment for end-organ- (RDW-CV), %
damage (proteinuria) should be pursued regardless. For the purposes of Red cell distribution width-SD 44.00  3.46 43.74  2.96 0.530
(RDW-SD), fL
our study, severe preeclampsia was defined as a systolic blood pressure of
Platelets (PLT), 109/L 215.18  55.01 199.30  47.37 0.016
at least 160 mmHg or a diastolic blood pressure of at least 110 mmHg, in Platelet distribution width 13.94  3.08 14.22  2.94 0.457
addition to proteinuria exceeding 2.0 g/24 h, serum creatinine exceeding (PDW), fL
106 μmol/L, urine output of less than 17 mL/h, a platelet count of less Mean platelet volume (MPV), Fl 11.13  1.29 11.25  1.17 0.414
Platelet large cell ratio (P-LCR), 33.89  10.17 34.89  9.28 0.418
than 100,000/mm3, and elevated levels of alanine aminotransferase
%
(ALT), aspartate aminotransferase (AST), or lactate dehydrogenase. This Plateletocrit (PCT), % 0.24  0.05 0.22  0.04 0.013
definition of severe preeclampsia is based on standard guidelines in
Hepatic and renal function examination variables
China. Our outcome variable was an indicator of severe preeclampsia,
whereby each record was labeled as (1) cases and (0) for controls. Alanine aminotransferase 23.00(24.00) 15.00(10.00) <0.001
(ALT), U/L
Severe preeclampsia group inclusion criteria: 1) severe preeclampsia
Aspartate aminotransferase 23.00(10.75) 20.00(6.00) <0.001
diagnosed during pregnancy; 2) complete maternal health records of the (AST), U/L
West China Second University Hospital from January 2017 to May 2019. Total bilirubin (TB), umol/L 9.05(3.60) 10.50(3.88) <0.001
Exclusion criterion: 1) previous history of hypertension, diabetes, or 2.30(1.30) 3.50(1.57) <0.001
heart disease before pregnancy; 2) history of other metabolic disease (continued on next page)

2
X. Zhang et al. Medicine in Novel Technology and Devices 15 (2022) 100158

Table 1 (continued ) for different model hyperparameters were averaged over five validations.
Variables SPE Cases (n ¼ Controls (n ¼ P Value* This allows the selection of the optimal hyperparameter combination
124) 124) that best fits a machine learning model.
Direct bilirubin (DBIL), umol/
The true disease status of the patients in the testing set was then used
L to obtain the false positive rates (FPR) and true positive rates (TPR) of the
Indirect bilirubin (IDIL), umol/L 6.50(2.70) 7.00(2.40) 0.159 models. Predictive accuracies for the tests were assessed via AUC for
Total protein (TP), g/L 71.28  4.41 72.02  3.88 0.162 these methods and variable importance, which measures the association
Albumin (ALB), g/L 42.05 ± 2.53 43.47 ± 2.86 <0.001
between a given variable and the accuracy of the prediction based on
Globulin (GLB), g/L 29.31  3.18 28.63  2.70 0.070
Glutamyl transpeptidase (r-GT), 15.00(19.75) 12.00(9.00) 0.014 percent increase in mean squared error [23]. Machine learning algo-
U/L rithms were performed in Python. All analysis was carried out in SPSS,
Lactate dehydrogenase (LDH), 167.49  28.25 163.92  20.67 0.257 version 16.0 (SPSS Inc., Chicago, IL, USA) [24].
U/L
Urea (UN), mmol/L 3.08  0.70 3.23  0.80 0.108
Creatinine (Cr), umol/L 42.36  6.41 42.50  5.63 0.858
2.4. Ethical considerations
Fasting plasma glucose (FPG), 4.49  0.69 4.34  0.31 0.024
mmol/L This study received approval from the Ethics Committee of Sichuan
Coagulation function examination variables University (K2019043). The need for informed consent was waived by
the Medical Ethics Committee, as this was an observational, retrospective
Fibrinogen (Fg), mg/dL 395.24  64.36 373.06  52.20 0.003
Thrombin time (TT), s 16.51  0.62 16.42  0.48 0.175
study using a database from which all personally identifiable information
Prothrombin time (PT), s 11.65 ± 0.61 11.37 ± 0.55 <0.001 had been removed.
International normalized ratio 0.95  0.05 0.97  0.05 0.015
(INR) 3. Results
Activated partial 27.92 ± 4.10 32.62 ± 3.06 <0.001
thromboplastin time
(APTT), s 3.1. Exploratory analysis
APTT ratio (APTT-R) 0.93 ± 0.11 1.03 ± 0.10 <0.001
A total of 19,653 medical records were obtained on pregnant women
Continuous statistics are expressed as mean  SD for normally distributed vari-
ables and as median (IQR) when the normality assumption is violated. Classifi- presenting to the West China Second University Hospital from January
cation statistics are expressed as n (%). 2017 to April 2019. According to the hypertensive disorders of preg-
*P values were obtained using chi-square test or the t-test as appropriate. **P nancy disease diagnosis criteria, 377 patients with severe preeclampsia
values were obtained using Mann-Whitney U test or the t-test as appropriate. were screened, of which 124 were severe preeclampsia. In these records,
SPE: Severe preeclampsia, SD: Standard deviation, IQR: Interquartile range. we excluded records of heart disease, diabetes, complications during
pregnancy, postpartum dysfunction, etc., and finally got 8638 healthy
such as autoimmune disease, chronic kidney disease, or antiphospholipid pregnant women records. In order to match the specific gestational week
syndrome during pregnancy. of the severe preeclampsia group, we further screened and selected 455
Control group inclusion criteria: 1) complete maternal health records healthy pregnant women for data statistics, of which 230 non-
of the West China Second University Hospital from January 2017 to May primiparous women and 45 cases of data were excluded. Finally, 124
2019; 2) during pregnancy and childbirth, the mother has no disease and people in the control group were obtained. These records were screened
in a state of complete health. Exclusion criterion: 1) maternal health as shown in Fig. 1. After removing patients who did not meet the in-
records of the West China Second University Hospital are incomplete. clusion criteria or did not have complete medical records, 124 cases and
124 healthy controls were carried forward as the study sample.
Table 1 report descriptive statistics on maternal characteristics and
2.3. Statistical analysis laboratory values, stratified by severe preeclampsia cases versus controls.
As shown, the mean age value of patients with severe preeclampsia was
Exploratory data analysis was first conducted to examine the unad- greater than that of healthy controls (33.04 versus 28.85 years; p <
justed associations between severe preeclampsia and the predictors of 0.001). And they are more likely to be overweight or obese than normal
interest. Bivariate tests consisted of Student's t-test for normal distribu- pregnant women. Twenty-two in patients with severe preeclampsia had a
tion and the Mann-Whitney U test for non-normal distribution. Fisher's BMI over 24, compared to five of healthy controls. They more often had
exact test was performed when appropriate. A Bonferroni correction was twin pregnancies (25.8% versus 0%, p < 0.001). 12–17% of severe
used to adjust the significance level for multiple comparisons. Prognostic preeclampsia patients also have other maternal complications, like thy-
modeling was then compared between three approaches: decision trees roid disease and diabetes mellitus. Total bilirubin (TB), direct bilirubin
(DT), light gradient boosting machines (LightGBM), and random forests (DBIL), and albumin (ALB) levels were significantly lower among cases
(RF). These approaches all represent various machine learning tech- (p  0.001), while ALT and AST were significantly higher (p  0.001).
niques which are known to have high predictive accuracy and compu- Significant differences in coagulation function were also found with
tational feasibility [19–21]. respect to prothrombin time (PT), activated partial thromboplastin time
The data were randomly stratified into training (65%) and testing (APTT), and the APTT ratio (APTT-R) between cases and controls. We
(35%) sets with a 1:1 ratio of cases to controls, and the model perfor- also tested whether the distributions of the blood examination variables
mance results of different models and different division ratios are shown differed significantly. As compared to the control group, only the dis-
in Supplementary Table S1. All three methods were first fit to the training tribution of AST range counts in the case group differed significantly (p
datasets to establish criteria for detecting severe preeclampsia. The < 0.001), while the other blood examination variables showed no sig-
resulting models were then applied to the testing datasets, and k-fold nificant difference. Significant positive correlations were found between
cross validation was used to tune optimal predictive models [22]. We are severe preeclampsia and AST, while negative correlations were found
using a five-fold cross-validation method. The training set is randomly between severe preeclampsia, DBIL (p < 0.001), and APTT-R (p < 0.001).
divided into five non-intersecting sets, four of which are used as a new
small training set, and the remaining one is used as a new small test set to 3.2. Predictive modeling
verify the performance of different model parameters on the new small
test set. Each cross-validation will be performed five times (five-fold), Based on the significant associations presented in Table 1 and stan-
and each piece of data will be used as a new small test set in turn. Results dard clinical diagnostic criteria [25], 20 clinical metrics were carried

3
X. Zhang et al. Medicine in Novel Technology and Devices 15 (2022) 100158

Fig. 1. Derivation of the study population. This figure displays the initial population of 19,653 records obtained on pregnant women presenting at the West China
Second University Hospital, Sichuan University, from January 2017 to April 2019, the relevant exclusions, and the final study population. HDP indicates hypertensive
disorders of pregnancy.

forward for predictive modeling. These included: percent monocytes, endorsed as an acceptable prediction level [26]. The AUC of all three
percent eosinophils, eosinophil count, red blood cell count, hematocrit, models exceeded 70%, suggesting that all three approaches may
plateletocrit, globulin, glutamyl transpeptidase, fasting plasma glucose, reasonably predict early severe preeclampsia. As shown, LightGBM using
fibrinogen, international normalized ratio, platelets, monocyte count, AST, DBIL, and APTT-R, had the highest predictive accuracy, with a
mean corpuscular hemoglobin concentration, ALT, AST, TB, DBIL, ALB sensitivity of 88.37%, a specificity of 77.27%, an AUC of 89.74%, an
and APTT-R. Important predictors of severe preeclampsia, as determined accuracy of 74.67%, and a positive predictive value of 65.96%. The
by DT, LightGBM, and RF, are presented in Table 2. Among the top 10 sensitivity, specificity, AUC, accuracy, and positive predictive values
predictors for each model, ordered by variable importance, AST, DBIL, were 76.74%, 70.45%, 71.41%, 70.67%, and 64.29% for DT and 86.05%,
and APTT-R appeared consistently across the three methods. This means 75.00%, 89.79%, 81.33%, 73.81% for RF, respectively.
they are the common variables in the top ten variables of the three
models and thus serve as the final predictors of the models. In a subse- 4. Discussion
quent analysis, the three methods were then refit using just these highly
predictive variables. 4.1. Main findings
The sensitivity and specificity of the refit DT, LightGBM, and RF
models are given in Fig. 2. A benchmark AUC of 70% has previously been In China, given the restrictions imposed by the family planning policy
and the potential high risks associated with pregnancy, an emphasis is
placed on early detection and treatment of maternal and infant disease.
Table 2 In line with this, a more extensive prenatal health exam is carried out
Important predictors resulting from predictive modeling (Importance values of
the 10 most influential predictors of each model are given in descending order).
Decision tree model LightGBM model Random forest model

(AUC ¼ 71.41%) (AUC ¼ 89.74%) (AUC ¼ 89.49%)

(ACC ¼ 70.67%) (ACC ¼ 74.67%) (ACC ¼ 81.33%)
(PPV ¼ 64.29%) (PPV ¼ 65.96%) (PPV ¼ 73.81%)

Predictor Importancea Predictor Importancea Predictor Importancea

APTT-R 1628.64 DBIL 122 DBIL 0.1211

INR 1064.77 APTT-R 118 APTT-R 0.1009
FPG 406.53 ALT 76 r-GT 0.0676
TB 75.27 RBC 61 ALB 0.0626
GLB 59.70 AST 30 ALT 0.0613
DBIL 58.00 TB 29 Fg 0.0587
RBC 30.89 EOS 23 AST 0.0492
PLT 16.16 ALB 20 TB 0.0490
HCT 14.96 INR 17 FPG 0.0455
AST 14.18 Fg 14 HCT 0.0447

LightGBM: Light gradient boosting machine AUC: Area under the ROC curve;
ACC: Accuracy; PPV: Positive predictive value; APTT-R: Activated partial
thromboplastin time ratio; INR: International normalized ratio; FPG: Fasting
plasma glucose; TB: Total bilirubin; GLB: Globulin; DBIL: Direct bilirubin; RBC:
Red blood cell count; PLT: Platelets; HCT: Hematocrit; AST: Aspartate amino-
transferase; ALT: Alanine aminotransferase; EOS: Eosinophil count; ALB: Albu- Fig. 2. Receiver operating characteristic curves (ROC) and area under the curve
min; Fg: Fibrinogen; r-GT: Glutamyl transpeptidase. (AUC) for the decision tree (DT), light gradient boosting machine (LightGBM),
a
Feature importance scores. and random forest (RF) models.

4
X. Zhang et al. Medicine in Novel Technology and Devices 15 (2022) 100158

relative to most international standards. Diagnoses are then made are differential changes in the function of the coagulation system among
following routine blood, hepatic, renal, and coagulation function severe preeclampsia cases in the early stages of pregnancy. This may be
examinations. due to blood vessel damage during the vasodilatory process in severe
Between severe preeclampsia cases and healthy controls, we observed preeclampsia, leading to increased activity of the endogenous coagula-
differences between several clinical metrics in the early stages of disease tion system.
development, despite these values falling within the normal range. TB, During pregnancy, the mother's basal metabolic rate increases. This
DBIL, ALB, PT and APTT levels were significantly elevated as compared leads to the enhanced consumption of nutrients, which increases burden
to controls. These observations may provide the basis for a narrower on the liver. ALT and AST are found in hepatocytes, with ALT mainly
normal range in the future. existing in the cytoplasm and AST in the mitochondria. When cells are
Today, machine learning algorithms are being utilized more damaged, such as in hepatitis and myocarditis, ALT enters the blood first.
frequently to explore the risk factors for disease occurrence and devel- Serum levels of AST will rise when a more severe form of cellular damage
opment, to build disease classification and prediction models, and to is achieved. Though average AST and ALT among severe preeclampsia
provide decision support for clinical diagnosis and treatment rationale cases were within reference ranges, we observed significantly higher
[22,27]. This study has established a new approach for early screening of levels as compared to our control group. Total and unconjugated bili-
severe preeclampsia by combining maternal characteristics and routine rubin concentrations may also be slightly decreased throughout gestation
blood examination metrics in a predictive model. Based on the prediction [34]. Our study found that bilirubin levels among severe preeclampsia
model, AST, DBIL, and APTT-R were predictive of severe preeclampsia cases were significantly lower than that of our controls. A previous study
with a combined AUC of 89.74%, accuracy of 74.67%, and positive found that increased levels of ALT, AST, and serum bilirubin should be
predictive value of 65.96%. considered pathologic and further evaluation should be undertaken [35].
Other studies have found that low levels of bilirubin leave the mother and
4.2. Strengths and limitations fetus susceptible to preeclampsia [36]. Synthesizing the above discussion
with our results, we first selected APTT-R, an indicator that reflects the
The main strengths of our study are as follows. (1) Our predictive endogenous coagulation system, followed by DBIL and AST, which reflect
model for severe preeclampsia has a high accuracy rate, and its indicators the hepatobiliary system. Our model using AST, DBIL, and APTT-R was
can be supported clinically, thus demonstrating its reliability. (2) We predictive of severe preeclampsia with an accuracy of 74.67%, and
selected predictors from routine blood, hepatic, renal, and coagulation positive predictive value of 65.96%.
function examinations. These examinations are generally required for all As has been noted, in the early stages of disease development, we
pregnant women in China. This means that the predictive model can be observed differences between several clinical metrics, despite these
applied without additional testing or additional medical fees. values falling within the normal range. TB, DBIL, ALB, PT, and APTT
Some potential limitations must also be considered. First, all samples levels were all found to be significantly higher in severe preeclampsia
come from a single center (the West China Second University Hospital). cases as compared to controls. We believe these measures could help
Future multi-center studies, including more patients, biomarkers, and provide direction for earlier detection and prevention of severe
risk factors are needed to validate the current findings. Second, it is preeclampsia.
unclear as to whether the current model is applicable to patients with
mild preeclampsia. Finally, this study uses indicators derived from lab- Funding sources
oratory tests which cannot be drawn at home, these there is a lag time to
obtaining the test results. National Key R&D Program of China (2018YFC2001800).

4.3. Interpretation Data availability statement

This study has established a new approach for the early screening of The data that support the findings of this study are available on
severe preeclampsia by a combination of maternal characteristics and request from the corresponding author. The data are not publicly avail-
routine blood examination. Gestational hypertension is new onset hy- able due to privacy or ethical restrictions.
pertension developed after 20 weeks of pregnancy with blood pressure
rising up to 140/90 mmHg without proteinuria. And, with the further
CRediT author statement
development of proteinuria, it can progress to preeclampsia [28].
The pathophysiological mechanisms leading to hypertension include
Xinyuan Zhang: paper's conception, data analyses, and manuscript
an increase in vascular resistance, which is largely dependent on the
drafting. Yu Chen and Huafeng Li: the development of the project. Ste-
decrease in blood vessel diameter caused by vasoconstriction and arterial
phen Salerno, Yi Li, Libin Zhou and Xiaoxi Zeng: manuscript editing and
remodeling. During pregnancy, the mother undergoes major anatomical
final approval of the manuscript.
and physiological changes [29]. An adaptive mechanism is the marked
fall in systemic vascular resistance that occurs in week six of gestation,
affecting the renal vascular system [30]. Specifically, relaxin stimulates
Declaration of competing interest
the formation of endothelin, which in turn mediates the vasodilation of
renal arteries through the synthesis of nitric oxide [31]. Arterial
The authors declare that there are no conflicts of interest.
under-filling in pregnancy leads to the stimulation of arterial barore-
ceptors, which activate the renin–angiotensin–aldosterone system
(RAAS) and the sympathetic nervous system. RAAS activation will Acknowledgements
mediate relative resistance to angiotensin II, counteracting the vaso-
constrictive effect and allowing for profound vasodilatation [32]. The We thank the statistician Mr. Chen Ruidong for statistical support.
expansion of vessels affects blood flow, which is one of the principal
factors affecting blood coagulation. When the velocity of blood flow Appendix A. Supplementary data
decreases, coagulation increase [33]. Average PT among severe pre-
eclampsia cases was significantly prolonged as compared to controls, Supplementary data to this article can be found online at https
while average APTT was significantly shorter. This indicates that there ://doi.org/10.1016/j.medntd.2022.100158.

5
X. Zhang et al. Medicine in Novel Technology and Devices 15 (2022) 100158

References [18] Wu F, Huang J, Zhou J. Inverse correlation between serum adiponectin level and
albuminuria in pregnancy-induced hypertension patients. Clin Exp Hypertens 2019;
41(3):263–7. https://doi.org/10.1080/10641963.2018.1469642.
[1] Ramos JGL, Sass N, Costa SHM. Preeclampsia. Rev Bras Ginecol Obstet 2017;39(9):
[19] Zhang J, Mucs D, Norinder U, et al. LightGBM: an effective and scalable algorithm
496–512. https://doi.org/10.1055/s-0037-1604471.
for prediction of chemical toxicity-application to the Tox21 and mutagenicity data
[2] Siddiqui A, Deneux-Tharaux C, Luton D, et al. Maternal obesity and severe pre-
sets. J Chem Inf Model 2019;59(10):4150–8. https://doi.org/10.1021/
eclampsia among immigrant women: a mediation analysis. Sci Rep 2020;10(1):
acs.jcim.9b00633.
5215. https://doi.org/10.1038/s41598-020-62032-9.
[20] Sufriyana H, Wu YW, Su EC. Artificial intelligence-assisted prediction of
[3] Akolekar R, Syngelaki A, Poon L, et al. Competing risks model in early screening for
preeclampsia: development and external validation of a nationwide health
preeclampsia by biophysical and biochemical markers. Fetal Diagn Ther 2013;
insurance dataset of the BPJS Kesehatan in Indonesia. EBioMedicine 2020;54:
33(1):8–15. https://doi.org/10.1159/000341264.
102710. https://doi.org/10.1016/j.ebiom.2020.102710.
[4] Moon M, Odibo A. First-trimester screening for preeclampsia: impact of maternal
[21] Jhee JH, Lee S, Park Y, et al. Prediction model development of late-onset
parity on modeling and screening effectiveness. J Matern Fetal Neonatal Med 2014;
preeclampsia using machine learning-based methods. PLoS One 2019;14(8):
28(17):2028–33. https://doi.org/10.3109/14767058.2014.978758.
e0221202. https://doi.org/10.1371/journal.pone.0221202.
[5] Skråstad R, Hov G, Blaas H-G, et al. Risk assessment for preeclampsia in nulliparous
[22] Ainscough KM, Lindsay KL, O'Sullivan EJ, et al. Behaviour change in overweight
women at 11–13 weeks gestational age: prospective evaluation of two algorithms.
and obese pregnancy: a decision tree to support the development of antenatal
BJOG 2014;13:1781–8. https://doi.org/10.1111/1471-0528.13194.
lifestyle interventions. Publ Health Nutr 2017;20(14):2642–8. https://doi.org/
[6] Goetzinger KR, Tuuli MG, Cahill AG, et al. Development and validation of a risk
10.1017/S136898001700129X.
factor scoring system for first-trimester prediction of preeclampsia. Am J Perinatol
[23] Fang X, Liu W, Ai J, et al. Forecasting incidence of infectious diarrhea using random
2014;31(12):1049–56. https://doi.org/10.1055/s-0034-1371705.
forest in Jiangsu Province, China. BMC Infect Dis 2020;20(1):222–30. https://
[7] Kang Q, Li W, Yu N, et al. Predictive role of neutrophil-to-lymphocyte ratio in
doi.org/10.1186/s12879-020-4930-2.
preeclampsia: a meta-analysis including 3982 patients. Pregnancy Hypertens 2020;
[24] Quintana DS, Williams DR. Bayesian alternatives for common null-hypothesis
20:111–8. https://doi.org/10.1016/j.preghy.2020.03.009.
significance tests in psychiatry: a non-technical guide using JASP. BMC Psychiatr
[8] Kanat-Pektas M, Yesildager U, Tuncer N, et al. Could mean platelet volume in late
2018;18(1):178. https://doi.org/10.1186/s12888-018-1761-4.
first trimester of pregnancy predict intrauterine growth restriction and pre-
[25] ACOG Practice Bulletin No. 202. Gestational hypertension and preeclampsia. Obstet
eclampsia? J Obstet Gynaecol Res 2014;40(7):1840–5. https://doi.org/10.1111/
Gynecol 2019;133(1):e1–25. https://doi.org/10.1097/AOG.0000000000003018.
jog.12433.
[26] Jayanthi N, Babu B, Rao N. Survey on clinical prediction models for diabetes
[9] Diguisto C, Gouge AL, Piver E, et al. Second-trimester uterine artery Doppler, PlGF,
prediction. J Big Data 2017;4(1):26. https://doi.org/10.1186/s40537-017-0082-7.
sFlt-1, sEndoglin, and lipid-related markers for predicting preeclampsia in a high-
[27] Han X, Zheng X, Wang Y, et al. Random forest can accurately predict the
risk population. Prenat Diagn 2013;33(11):1070–4. https://doi.org/10.1002/
development of end-stage renal disease in immunoglobulin a nephropathy patients.
pd.4198.
Ann Transl Med 2019;7(11):234. https://doi.org/10.21037/atm.2018.12.11.
[10] Gedikbasi A, Yücel B, Dundar O, et al. The utility of first trimester uterine artery
[28] Williams PJ, Broughton Pipkin F. The genetics of pre-eclampsia and other
Doppler, placental volume and PAPP-A levels alone and in combination to predict
hypertensive disorders of pregnancy. Best Pract Res Clin Obstet Gynaecol 2011;
preeclampsia. Pregnancy Hypertens 2016;6(4):269–73. https://doi.org/10.1016/
25(4):405–17. https://doi.org/10.1016/j.bpobgyn.2011.02.007.
j.preghy.2016.04.007.
[29] Soma-Pillay P, Nelson-Piercy C, Tolppanen H, et al. Physiological changes in
[11] Pihl K, Sorensen S, Stener Jorgensen F. Prediction of preeclampsia in nulliparous
pregnancy. Cardiovasc J Afr 2016;27(2):89–94. https://doi.org/10.5830/CVJA-
women according to first trimester maternal factors and serum markers. Fetal Diagn
2016-021.
Ther 2020;47(4):277–83. https://doi.org/10.1159/000503229.
[30] Tkachenko O, Shchekochikhin D, Schrier RW. Hormones and hemodynamics in
[12] Duan B, Zhang L, Ding X, et al. Serum beta-trace protein as a Novel predictor of
pregnancy. Int J Endocrinol Metabol 2014;12(2):e14098. https://doi.org/10.5812/
pregnancy-induced hypertension. J Clin Hypertens (Greenwich). 2016;18(10):
ijem.14098.
1022–6. https://doi.org/10.1111/jch.12801.
[31] Conrad KP. Emerging role of relaxin in the maternal adaptations to normal
[13] Zhang Y, Ma Q, Yang H, et al. Maternal plasma TIMP-4 levels combined with
pregnancy: implications for preeclampsia. Semin Nephrol 2011;31(1):15–32.
clinical risk factors for the early prediction of pregnancy-induced hypertension.
https://doi.org/10.1016/j.semnephrol.2010.10.003.
Arch Gynecol Obstet 2015;292(5):1043–50. https://doi.org/10.1007/s00404-015-
[32] Gant NF, Worley RJ, Everett RB, et al. Control of vascular responsiveness during
3753-7.
human pregnancy. Kidney Int 1980;18:253–8. https://doi.org/10.1038/
[14] Chen H, Zhang J, Qin F, et al. Evaluation of the predictive value of high sensitivity
ki.1980.133.
C-reactive protein in pregnancy-induced hypertension syndrome. Exp Ther Med
[33] Bal V. Coagulation behavior of spherical particles embedded in laminar shear flow
2018;16(2):619–22. https://doi.org/10.3892/etm.2018.6246.
in presence of DLVO-and non-DLVO forces. J Colloid Interface Sci 2019;564:
[15] Chandi A, Sirohiwal D, Malik R. Association of early maternal hypertriglyceridemia
170–81. https://doi.org/10.1016/j.jcis.2019.12.119.
with pregnancy-induced hypertension. Arch Gynecol Obstet 2015;292(5):1135–43.
[34] Y B, Zarka O, Brechot JF, et al. Liver function tests in normal pregnancy: a
https://doi.org/10.1007/s00404-015-3706-1.
prospective study of 103 pregnant women and 103 matched controls. Hepatology
[16] Maru L, Verma M, Jinsiwale N. Homocysteine as predictive marker for pregnancy-
1996;23(5):1030–4. https://doi.org/10.1002/hep.510230514.
induced hypertension-A comparative study of homocysteine levels in normal versus
[35] Mufti AR, Reau N. Liver disease in pregnancy. Clin Liver Dis 2012;16(2):247–69.
patients of PIH and its complications. J Obstet Gynaecol India 2016;66(Suppl 1):
https://doi.org/10.1016/j.cld.2012.03.011.
167–71. https://doi.org/10.1007/s13224-015-0832-4.
[36] Breslin E, Kaufmann A, Quenby S. Bilirubin influences the clinical presentation of
[17] Koenig JM, Chegini N. Enhanced expression of Fas-associated proteins in decidual
pre-eclampsia. Eur J Obstet Gynecol Reprod Biol 2013;170(1):111–3. https://
and trophoblastic tissues in pregnancy-induced hypertension. Am J Reprod
doi.org/10.1016/j.ejogrb.2013.05.024.
Immunol 2000;44(6):347–9. https://doi.org/10.1111/j.8755-8920.2000.440605.x.