Complete blood count as a biomarker for preeclampsia with severe features diagnosis: a machine learning approach

Abstract

Objective

This study introduces the complete blood count (CBC), a standard prenatal screening test, as a biomarker for diagnosing preeclampsia with severe features (sPE), employing machine learning models.

Methods

We used a boosting machine learning model fed with synthetic data generated through a new methodology called DAS (Data Augmentation and Smoothing). Using data from a Brazilian study including 132 pregnant women, we generated 3,552 synthetic samples for model training. To improve interpretability, we also provided a ridge regression model.

Results

Our boosting model obtained an AUROC of 0.90±0.10, sensitivity of 0.95, and specificity of 0.79 to differentiate sPE and non-PE pregnant women, using CBC parameters of neutrophils count, mean corpuscular hemoglobin (MCH), and the aggregate index of systemic inflammation (AISI). In addition, we provided a ridge regression equation using the same three CBC parameters, which is fully interpretable and achieved an AUROC of 0.79±0.10 to differentiate the both groups. Moreover, we also showed that a monocyte count lower than $490 /mm^{3}$ yielded a sensitivity of 0.71 and specificity of 0.72.

Conclusion

Our study showed that ML-powered CBC could be used as a biomarker for sPE diagnosis support. In addition, we showed that a low monocyte count alone could be an indicator of sPE.

Significance

Although preeclampsia has been extensively studied, no laboratory biomarker with favorable cost-effectiveness has been proposed. Using artificial intelligence, we proposed to use the CBC, a low-cost, fast, and well-spread blood test, as a biomarker for sPE.

Introduction

Preeclampsia (PE), which affects 5-8% of pregnancies, is the leading cause of maternal and perinatal death, intrauterine growth restriction, and preterm birth [1]. Annually, PE is responsible for over 70,000 maternal and 500,000 fetal deaths [2], predominantly in low-income countries. However, many of these outcomes are potentially preventable [3].

The diagnosis of PE hinges on elevated blood pressure ($\ge 140/90mmHg$) with or without proteinuria. In the absence of proteinuria, symptoms such as liver dysfunction or neurological problems including headaches and scotoma become crucial for diagnosis [4]. When blood pressure levels exceeds 160/100mmHg, the condition evolves to PE with severe features (sPE) [5], which may progress to catastrophic outcomes such as eclampsia, HELLP syndrome (hemolysis, elevated liver enzymes, low platelet), and disseminated intravascular coagulation. Eclampsia presents with neurological symptoms, potentially leading to seizures, coma, and death [6]. PE’s severity in pregnant women can intensify unexpectedly, without changes that lead to this suspicion, underscoring the associated maternal and fetal death risks [6]. Therefore, it is paramount to monitor pregnant women with PE, to alert to the evolution to sPE and catastrophic outcomes.

Currently, no cost-effective laboratory markers for PE diagnosis or monitoring have been established. The complete blood count (CBC), a routine and affordable laboratory test, serves as a valuable decision-making tool, since variations in its parameters offer significant insights into the diagnosis of multiple diseases. Highlighting the utility of CBC in PE, meta-analyses conducted by Woldeamanuel et al. revealed that platelet counts in preeclamptic women were significantly lower across 56 studies [7]. In addition, Bulbul et al. identified lymphocytes and MPV (mean platelet volume) as useful predictors of sPE [8]. Furthermore, CBC-derived ratios, such as the systemic inflammation response index (SIRI) [9], the neutrophil-to-lymphocyte ratio (NLR) [8, 10, 11], and the platelet-to-lymphocyte (PLR) ratio [10, 11], are gaining recognition as potential PE markers due the association with immune system changes and systemic inflammation often seen in PE.

Recently, machine learning (ML) techniques have emerged as a valuable tool for the diagnosis and prognosis of multiple diseases [12, 13], given that they can expose non-linear relationships between biomarkers, revealing patterns not immediately apparent. Lu and Hsu used ML to develop a model for predicting the PE severity using the CBC along with maternal-obstetrical characteristics [14]. Jhee et al. developed an ML model for predicting late-onset PE using multiple clinical and laboratory markers, including the CBC [15]. Eberhard et al. developed multiple ML models for PE risk in their longitudinal study, using sociodemographic, clinical diagnoses, family history, laboratory, and vital signs data [16].

A key challenge in developing ML models for health disorders like sPE is assembling representative datasets. Data collection and labeling are costly [17, 18], limiting dataset quality and diversity, crucial for reliable ML applications. To mitigate these issues and the high costs of obtaining real-world annotated medical data, synthetic data generation is increasingly utilized [19, 20], creating artificial data with similar statistical properties to original datasets [21].

This study introduces a ML approach for supporting sPE diagnosis using CBC tests from pregnant women in the third trimester. This novel application of ML-powered CBCs has the potential to improve PE monitoring protocols, making them more efficient and broadly accessible. Key contributions include:

• Develop a high-performing ML boosting model for sPE detection using only CBC parameters.

• Utilize a novel synthetic data generation method, DAS, to enhance dataset quality.

• Apply feature selection to identify key CBC parameters and using explainability techniques.

• Introduce a simpler, interpretable ridge regression model.

• Explore clinical links between CBC parameters and sPE with biological plausibility.

Methods

Dataset description

This case-control study was conducted in three Brazilian Public Hospitals, in Belo Horizonte, Minas Gerais, Brazil. Ethical approval was granted by the Federal University of Minas Gerais (UFMG) Research Ethics Committee (ETIC 0618.0.203.000-10), and written informed consent was obtained from all 132 participants. Clinical and CBC data for each study participant were collected from their prenatal card and medical record. All participants were pregnant women at third trimester of gestation, and the inclusion criteria for case (PE with severe features) and control (normotensive pregnant women) groups are detailed below:

• Case Group: Minimum of two episodes of systolic/diastolic pressure $\ge 160/110mm Hg$, measured four hours apart, and proteinuria greater than (++) by the qualitative method in an isolated urine sample, or $\ge 2g$ in 24-hours urine. In cases with absence of proteinuria, liver or neurological disfunction, and low platelet count are imperative for their inclusion [22].

• Control Group: systolic/diastolic pressure $\le 120/80mm Hg$, no previous history of hypertension or PE, and absence of proteinuria.

We collected data from 65 cases and 67 controls, matched for age and gestational age. The exclusion criteria for both groups were advanced labor, previous hypertension, bleeding, and inter-current chronic conditions/diseases such as gestational diabetes mellitus (GDM), obesity, thrombosis, cancer, and cardiovascular, kidney, liver, or autoimmune diseases. Additionally, pregnant women with active infections, clinical symptoms indicative of infection such as fever, or those using medications like anti-inflammatories were also excluded.

To ensure the integrity and transparency of our study, we followed the TRIPOD guidelines proposed by Collins et al. (2016) [23].

Preprocessing

Some CBC parameters were missing for certain patients. Therefore, before generating synthetic data to construct the synthetic training sets, as detailed in “Steps” section, and to prevent data leakage, we split the dataset and substituted the missing values in the seed sets with the average values of their respective parameters. Similarly, we replaced the missing values in the test sets with the averages of their corresponding parameters. Since the dataset was already balanced (49% cases and 51% controls), there was no need to address data imbalance.

The CBC-derived ratios and their respective formulas are presented in Table 1.

Table 1 CBC-derived ratios

DAS - data augmentation and smoothing

In a nutshell, the general idea of DAS is to use a random weighted average of the samples of each class as a new synthetic sample, which could build an infinite number of synthetic samples. We were inspired by the framework proposed by Forestier et al. (2017) [24] for generating synthetic time series under DTW (dynamic time warping).

Formalism

Consider an original dataset ${D_{orig} \subset {\{X\} }_{i=1}^{m}}$ consisting of n i.i.d. samples and m attributes. The task of synthetic data generation tackled in this paper is to generate a synthetic dataset $D_{syn}$, which has the same number and type of attributes and an automatically estimated number of samples. Further, a model trained on $D_{syn} \sim p_{syn}$ could be later applied to a real dataset $D_{real} \sim p_{real}$, that may or not be $D_{orig}$. We have proposed an approach that allows straightforward transfer learning from synthetic to real data, thus building the synthetic data distribution $p_{syn}$ sufficiently wide and varied so that the model trained on $p_{syn}$ will be robust enough to work well on $p_{real}$.

The intuition behind our approach is to take a set of h original samples from the same class in $D{_{orig}}$ and randomly generate for each synthetic sample $s_i$ a weights vector $W_i$ of size h, further selecting one original sample (that is, one position of $W_i$) that will receive about 50% of the sum of weight. Thus, each synthetic sample $s_i$ will be composed of a weighted sum of all h samples. Further, each synthetic sample receives a random variation (of up to a percentage selected by the user).

To avoid the addition of new synthetic samples that are very similar to another synthetic sample in the final set, we only select a synthetic sample if, for all samples in the final set, the absolute sum of reasons between its feature values and the sample’s feature values is at least higher than the hundredth part of the log of the size of the original set.

Consider that:

• X is the set of original samples from the same class;

• n is the number of samples in the original set X;

• m is the number of features in X;

• S is a newly generated synthetic sample;

• Z is the final set of synthetic samples with k samples, and may or not receive S.

We are going to add S into Z only if, for each one of i samples in Z, the following is true:

$$\begin{aligned} \sum \limits _{j=1}^{m} \dfrac{S_{j}}{Z_{i,j}} > \dfrac{\log {n}}{100} \end{aligned}$$

(1)

Steps

We have detailed the steps taken for applying DAS and performing ten-fold cross-validation.

• 1. Divide the original dataset $D{_{orig}}$ into k non-overlapping folds to perform further cross-validation (a popular default, here chosen, is k=10 [25]) .

• 2. Partition each i-th fold of $D{_{orig,i}}$ into its j classes (in this particular problem, there are only two classes: preeclampsia vs control patients).

• 3. For each j-th class of the i-th fold $D{_{orig,i,j}}$, fill missing values with a methodology for data imputation (we used feature averages).

• 4. For each j-th class of i-th fold $D{_{orig,i,j}}$, generate synthetic data samples using DAS.

• 5. Combine the generated synthetic data samples from both classes into $D_{syn,i}$.

• 6. To perform cross-validation, combine generated synthetic data samples from all folds, except i, as $D_{train}$. The corresponding validation set will be $D{_{orig,i}}$.

The steps 1-5 above are summarized in Fig. 1.

Fig. 1

Steps for synthetic data generation

DAS algorithm

DAS pseudocode, accounting for steps 3 and 4 of “Steps” section, is presented in Algorithm 1.

Algorithm 1 DAS pseudocode

Boosting models

Our models were trained using synthetic data and tested using real data. As shown in “Steps” section, we conducted ten-fold cross-validation to ensure that the models were never trained using synthetic data generated with a seed of the test set. The models were built using the LightGBM algorithm [26], which uses boosting to compose a complex model using simple decision trees [27]. Classification performance was measured using the area under the curve (AUROC) [28]. For comparison, we also generated the same number of synthetic samples using SMOTE [29], and ADASYN [30] - the most popular methods for tabular synthetic data generation [31]. Finally, we showed the AUROCs obtained when the LightGBM algorithm is trained with each dataset and using only the real dataset.

Feature selection

To reduce the combinatorial model space, we sampled it as a directed acyclic graph in which each node represents a distinct feature subset, and vertex A $\rightarrow$ B is connected if B can be reached by simple feature addition from A. This approach delivers a partial ordering starting from the root node (that is, the first chosen feature), which allows us to search for the following features in an orderly manner [12]. Thus, we selected the feature that gives the highest AUROC model at each iteration, combined with the previously selected features. This feature selection method mitigates the inclusion of redundant features, ensuring that each newly incorporated feature provides a unique contribution to the model’s performance.

Explanations

To assess the importance and individual contribution of each feature and thus extract intuitive insights from the boosting model, we used the SHAP algorithm [32], which assesses the relative significance of each feature by taking it out from the model. SHAP provides an explanation vector for each input, called SHAP values, which has the same dimension as the inputs and indicates the importance of the corresponding feature in a particular prediction. However, although each feature is evaluated separately, non-linear interactions among all features occur in the model. Thus, the SHAP value for a feature should not be considered as its isolated effect but as its compound effect when interacting with the other features.

Moreover, we have also provided one fully interpretable model: a ridge regression. Despite its lower accuracy compared to the boosting model, it can be helpful by providing clinicians with more interpretable indications [33].

Hyperparameters optimization

For the hyperparameters optimization, we proposed using the Weighted Balanced Accuracy (WBA) metric [13], defined below. Considering the difficulties in screening for PE, false negatives are more undesirable than false positives, and therefore a greater weight to sensitivity than specificity should be given. By performing a grid search on the model hyperparameters while optimizing WBA, we optimized the sensitivity and specificity using the weights of 0.77 and 0.33, respectively. Table 2 shows the searched and selected parameters.

$$\begin{aligned} WBA = \frac{2*Sensitivity+Specificity}{3} \end{aligned}$$

(2)

Table 2 Hyperparameter search space of the best model

Results

We examined data from 132 pregnant women, of whom 65 had severe PE and 67 were normotensive pregnant women. Table 3 shows the CBC statistical parameters for the two groups. Although almost all CBC parameters had p-values $<0.05$, except for LMR, SII, SIRI, and RBC count, only monocytes had an AUROC greater than 0.70 in the linear separation of both groups. Considering the monocyte cutoff point that maximized the WBA, set at $<490 /mm^{3}$, for the classification of sPE, the accuracy was 0.71, sensitivity 0.71, and specificity 0.72.

Table 3 Statistics for both groups for CBC ratios and raw values. Student’s t-test was used to compare variables

Boosting model performance

Using data from 65 women diagnosed with sPE and 67 normotensive pregnant women as seeds, we employed DAS to generate 2,309 cases and 1,243 controls of synthetic data. The generation process, detailed in the Methodology section, involved dividing the original dataset into ten parts. Nine parts were used as seeds to create synthetic data, while one part served as a test set for feature selection. This cycle was repeated nine times to ensure every segment of the dataset served as a test set, with the seed dataset for generating synthetic data excluding the test data in each iteration.

The selected features were two raw CBC parameters, neutrophils and mean corpuscular hemoglobin (MCH), plus a CBC-derived ratio, aggregate index of systemic inflammation (AISI). This model had an AUROC of 0.90 ± 0.10, sensitivity of 0.95, specificity of 0.79, accuracy of 0.87, and precision of 0.82.

Figure 2 shows the ROC curves obtained when training the model with synthetic data generated by DAS, SMOTE, and ADASYN, and also when training the model using only the real dataset (using cross-validation of 10 folds). The AUROCs obtained were respectively 0.90±0.10, 0.88±0.08, 0.86±0.09, and 0.86±0.10. Although DAS a priori outperformed the other databases, the paired samples t-tests showed that the differences were not significant. Regarding ADASYN and SMOTE hyperparameters, we used the nearest neighbors number of three. Because both algorithms require a pre-defined number of samples to be generated, we set them to the same numbers automatically generated by DAS (2,309 cases and 1,243 controls). For DAS, we selected a random variation of up to 5%. For all methods, missing values were imputed using the parameters’ averages.

Fig. 2

ROC curves using synthetic data generation methods of DAS, SMOTE, and ADASYN for augmenting the training set and the real dataset

Explanations

The contribution of each parameter - neutrophils, MCH and AISI - was graphically represented in the SHAP summary plot (Fig. 3), where parameters are depicted in the order of importance. Red dots are associated with patients for whom the corresponding parameter has a relatively higher value. On the contrary, blue dots are associated with patients for whom the corresponding parameter has a relatively lower value. A vertical line separates dots (that is, patients) - the ones on the left side are those for whom the model provided a lower risk of PE, and on the right side, a higher one. All the available data was used to plot SHAP, both synthetic and real.

Fig. 3

SHAP Summary Plot. Abbreviations: AISI= Aggregate índex of systemic inflammation; MCH= Mean corpuscular hemoglobin

Table 4 Pearson correlations and p-values for each parameter and its respective SHAP values. Abbreviations: AISI= Aggregate índex of systemic inflammation; MCH= Mean corpuscular hemoglobin

The summary plot in Fig. 3 enables us to grasp general patterns about PE risk, as for all three parameters, one side has significantly more blue or pink dots than the other side. These visual patterns can be verified in Table 4, which shows the Pearson correlations and respective p-values for each parameter and its SHAP values. For AISI and MCH, we can see more blue dots on the right and red on the left, meaning that relatively lower values of these parameters are associated with PE. The Pearson correlations of these parameters with their SHAP values are -0.90 and -0.88, being considered very strong [34], thus confirming the observed visual patterns. For neutrophils, the opposite pattern is observed, i.e., relatively higher values are found in preeclamptic women, with a Pearson correlation of 0.90, which is considered very strong.

Fully interpretable model

In addition to presenting the SHAP explanations for the boosting model, we also provide a ridge regression model. Although lower in performance, it is considered a fully interpretable model and may help in clinical interpretation.

This ridge regression model obtained AUROC of 0.79±0.10, and may be expressed with the simple equation that follows:

$Y = 1.71 - 2.84 \times ||MCH|| - 7.93 \times ||AISI|| + 4.78 \times ||Neutrophils||$

where:

Y > 0:

means high risk of PE

Y$\le$0:

means low risk of PE

Markers were normalized using the following equations:

$$\begin{aligned} ||MCH|| = \frac{MCH-19.8}{17.1} \end{aligned}$$

(3)

$$\begin{aligned} ||AISI|| = \frac{AISI-569.8}{2,905,089.6} \end{aligned}$$

(4)

$$\begin{aligned} ||Neutrophils|| = \frac{Neutrophils-3391.4}{16,096.6} \end{aligned}$$

(5)

Confirming the tendencies observed in the SHAP summary plot, we can see that MCH and AISI have negative signals in the equation, and thus higher relative values of these parameters contribute to decreasing the PE risk, while neutrophils have a positive signal.

Discussion

Main findings

This study aimed to develop an ML model based on CBC for sPE detection. Although PE has been extensively studied, no laboratory biomarker that presents favorable cost-effectiveness has been proposed so far. Some biophysical and biochemical markers, such as the angiogenic factors ratio (PlGF/sFlit-1) and the resistance/pulsatility index, have been proposed as screening tests in the first trimester of pregnancy. However, none of them showed satisfactory predictive potential in multicenter studies. Besides, PE diagnosis continues to be established based on blood pressure measurement and determination of proteinuria, plasma creatinine and liver enzymes, platelet count, and clinical data [22]. It is noteworthy that many diseases, including COVID, mimic these laboratory and clinical findings, which complicates, even more, the PE diagnosis [1, 35].

Our study showed that the CBC could be useful for supporting the sPE diagnosis. Using ML, we developed a boosting model that uses neutrophils count, AISI, and MCH. This model achieved AUROC, sensitivity, specificity, accuracy, precision, and F1 of 0.90, 0.95, 0.79, 0.87, 0.82, and 0.88, respectively, being the first ML model for sPE detection using CBC. This model could be a potentially practical tool for obstetricians. Furthermore, we have provided a simple equation that achieved 0.79 AUROC using the three same CBC parameters; and also showed that a monocyte cutoff point of $<490 /mm^{3}$ achieved an accuracy of 0.71, emerging for the first time a low monocyte count as a significant indicator of sPE.

To address the challenge of a limited database size, we employed the DAS methodology to generate 3,552 synthetic samples from an original set of 132. While DAS outperformed the original dataset and those augmented by SMOTE and ADASYN in AUROC scores, the improvements were not statistically significant, indicating that data augmentation may not be necessary in this particular context.

Clinical explanations

As shown in the analysis, neutrophils - the most frequent subtype of circulating white blood cells - are increased in preeclamptic women with severe features. Knowing that embryo implantation in the uterus and placental development are inflammatory events, the activation of the maternal immune system is essential for a healthy pregnancy. However, the physiological balance between proinflammatory/regulatory responses, with a shift toward a regulatory state, is crucial for pregnancy success [36]. In PE, it has been observed that this alteration does not occur, or it is reverted in the very early stages of the disease, and in consequence, it leads to a proinflammatory state [37]. In addition, several pieces of evidence support a balance of proinflammatory and anti-inflammatory/pro-resolving pathways in healthy pregnancy due to the well-functioning resolution of inflammation mechanisms. It leads to a state of controlled inflammatory response in healthy pregnant [36]. On the other hand, our group reported that pro-resolution pathways might be compromised in PE women [38]. As a consequence, higher plasma levels of proinflammatory cytokines [37, 39, 40] and lower levels of pro-resolution factors [41] in PE women compared to those with normal pregnancy have been found. It is well-established that inflammatory processes increase the number of circulating neutrophils. Wang et al. showed that neutrophils produced significantly more IL-6 when cells were primed by preeclamptic placental conditioned medium compared to normal placental conditioned medium. This finding provided evidence that the placenta of PE women plays a role in inducing neutrophil activation [42]. In light of this knowledge, our finding of neutrophilia in preeclamptic women is wholly justified.

Table 3 presents CBC ratios and raw values. A comparative analysis of data from normal pregnant women versus those with sPE revealed that AISI, neutrophil-to-lymphocyte ratio (NLR), derived NLR (dNLR), neutrophil-to-lymphocyte ratio (NLR), and platelet-to-lymphocyte ratio (PLR), in addition to hematocrit, hemoglobin, leukocytes, lymphocytes, mean corpuscular volume (MCV), MCH, mean corpuscular hemoglobin concentration (MCHC), monocytes, neutrophils, and platelets counts were significantly different between the two groups. PE has been reported to be a consequence of systemic inflammation, oxidative stress, and endothelial dysfunction, with evidence suggesting an exaggerated inflammatory response during pregnancy due to an imbalance of innate and adaptive immune responses [38, 43]. Based on our data, in preeclamptic women, CBC ratios suggested a greater inflammatory process attested by neutrophilia and lymphocytosis and a reduction in the number of platelets. The thrombocytopenia, defined as a platelet count lower than 150 $\times$ $10^3 /{mm}^3$, probably is caused by platelet consumption due to platelet activation caused by PE, in addition to immune mechanisms or severe vasospasm with resultant endothelial damage that may contribute to the reduced platelet number [44]. It should be noted that thrombocytopenia in PE can be a very important finding, as this hematological change may indicate a coexisting systemic or gestational disorder, with chances of potential maternal intervention reducing the risks for the fetus [45].

In line with our results, Yakistiran et al. (2022) reported similar findings comparing NLR, dNLR, and PLR in preeclamptic and healthy pregnant women [11]. In addition, these authors observed moderate negative correlations between maternal PLR (r=-0.231, p= 0.002), uric acid (r= 0.332, p=0.000), and adverse neonatal outcomes. Also, Taşkömür & Erten (2021) observed that the neutrophil-to-lymphocyte ratio (NLR) was significantly higher in the severe PE than in healthy pregnancies; however, no difference was observed for mild PE [46]. Unlike us, these authors did not find a difference between the groups regarding PLR. The lack of significant difference between the two groups for PLR is unclear, since thrombocytopenia is a common finding in PE [45, 47, 48]. In contrast, and according to our data, PLR was significantly lower in severe PE than in mild PE (p $< 0.001$) [49, 50]. Our findings for NLR were also in agreement with those observed in a meta-analysis including 3,982 patients [51], and in other studies [50], whose values were higher in preeclamptic women, especially those with severe PE.

Seyhanli showed that the systemic inflammatory index (SIRI) was statiscaly significant in differentiating between the preeclampsia and control groups [9] . Regarding the LMR and the other systemic inflammatory indices (AISI, and SII), no studies were found in the literature linking them to PE or sPE, which makes our findings unique.

In our study, in addition to elevated NLR in preeclamptic women, increased dNLR were also observed, as expected. We found decreased AISI in studied PE women. AISI is calculated as $neutrophils * monocytes * platelets/lymphocytes$. Our statistical analysis showed that monocytes and platelets are significantly decreased in PE women, while neutrophils and lymphocytes are increased. Although the number of neutrophils increased in preeclamptic women compared to normal pregnant women, the lower quotient (lower AISI) presented by the first group compared to the second one, was the result of the lower number of platelets and monocytes, besides the increased number of lymphocytes.

On the other hand, CBC parameters have indicated mild anemia in sPE women compared to healthy pregnant women. Probably this is due to hemolysis, in line with both reduced hematocrit and hemoglobin results, as well as decreased MCV, MCH, and MCHC. It should be noted that although these parameters are reduced in PE, they are still within the reference ranges.

It is known that pregnancy is a high-risk period for the development of different types of thrombotic microangiopathy [52]. This disorder and PE share clinical and laboratory characteristics, consequent to endothelial damage and formation of fibrin thrombi in the microcirculation resulting in fragmentation of erythrocytes (schistocytes). This fact may partially explain the hemolysis and consequent drop in parameters associated with erythrocytes. In parallel with damage to erythrocytes, platelet aggregation occurs, leading to ischemic injury to target organs, i.e., kidneys and brain [52].

MCH, which estimates the average amount of hemoglobin within the red blood cells, is also decreased in studied sPE women. Knowing that iron is the main component of the hemoglobin molecule (since oxygen binds to iron to be delivered to all cells in the body), disturbances in its metabolism can compromise the synthesis of hemoglobin. Uncontrolled, increased proinflammatory responses with cytokine production contribute to the pathogenesis of PE and may explain the reduced supply of iron to the bone marrow, compromising the synthesis of hemoglobin. Therefore, the average amount of hemoglobin within the red blood cells is reduced. As reviewed by Madu & Ughasoro (2017), cytokines and acute phase proteins play essential roles in the pathogenesis of anemia of chronic diseases. In the same context, changes in iron metabolism involving hepcidin and ferritin molecules contribute largely to the anemia that sets in [53]. According to Spence et al. (2021), the proinflammatory cytokines TNF-$\alpha$, IFN-$\gamma$, IL-2, IL-8, and IL-6 are significantly elevated in PE, while lower concentrations of IL-10 in the second trimester may be an early predictor for the development of this disease [54].

In summary, the present study proposed a digital tool for sPE using CBC parameters. We proposed three models of different complexities: (1) a boosting model that achieved 0.90 AUROC; (2) a fully-interpretable equation obtained using ridge regression that achieved 0.79 AUROC; and (3) a simple cutoff point that obtained 0.70 AUROC. Although cutoff points are more common in clinical decision-making, ML models can identify non-linear relationships between CBC parameters, thereby recognizing existing patterns that are less likely to be identified through traditional approaches. Thus, an ML-based tool could achieve better performance and be easily deployed in multiple clinical contexts, in addition to its versatility. In other words, this tool may represent a mobile device with potential application - i.e., a doctor enters the CBC parameters or an API application connected directly to a clinical laboratory running routine tests to label the respective exam with an sPE risk-score, just to name a few use cases.

Limitations

Despite the promising results presented here in the field of PE, some limitations should be mentioned, namely, the limited sample size, the heterogeneity of individual characteristics among pregnant women, and the few studies involving some inflammatory indices.

Additionally, the exclusion criteria of the study, which removed pregnant women with active infections, pro-inflammatory states, and other conditions, could introduce potential biases and limit the model’s applicability in real-world scenarios where such conditions are common.

Another critical limitation is the absence of external validation, which is essential to corroborate our findings across different populations and settings. The lack of external validation means that the model’s robustness and generalizability to diverse clinical environments remain uncertain.

Furthermore, we acknowledge that some confounding factors, such as periodontitis, may still be present despite our efforts to control for them. Periodontitis, although less common in women of childbearing age, may be more prevalent during pregnancy and could influence the model’s accuracy.

Lastly, it should be noted that the input data, specifically the CBC markers, may vary in unit measures across different laboratories, and the naming conventions for these features may differ, posing challenges in applying the model to real-world data.

Conclusions

Currently, one of the most critical goals in obstetrics is identifying pregnant women at increased risk of sPE. The merit of the present study is its unprecedented character, the use of artificial intelligence to give support to sPE diagnosis using the CBC, a low-cost and well-spread blood exam that is already included in the prenatal panel. Our data showed that the number of circulating neutrophils, MCH, and AISI are significantly altered in sPE women in the third trimester of pregnancy. When these parameters are combined in a boosting ML model, it becomes possible to identify PE with 0.95 sensitivity and 0.79 specificity.

These findings encourage a new, carefully designed project to confirm the unprecedented data reported here. Considering the wide distribution and cheapness of CBC, a longitudinal study is welcome, including CBC before the 20th week of pregnancy (when the symptoms appear), from the 21st to the 26th week (the end of the second trimester), from the 26th to 34th weeks and after the 34th week (period that distinguishes early and late PE). Assessment of circulating neutrophils, MCH, and AISI in these four gestational periods will validate the potential of these parameters to track PE.

In conclusion, careful monitoring of the circulating neutrophils count, MCH, and AISI emerges as a useful tool for monitoring pregnant women. The timely detection of these parameters allows for rigorous surveillance and intervention that will undoubtedly benefit pregnant women and their fetuses. Finally, we highlight that DAS could enable data analysis in other medical conditions where it is very difficult or expensive to recruit participants or label outcomes, but further investigations are needed.