Early Prediction of Preeclampsia Using a Multimodal Machine Learning Model Integrating Clinical, Hemodynamic, and Biochemical Data

Preeclampsia (PE) is a multisystem hypertensive disorder of pregnancy that typically manifests after 20 weeks of gestation and arises from a complex combination of genetic, immunological, inflammatory, and environmental mechanisms. Central to its pathophysiology is inadequate remodeling of the uterine spiral arteries and systemic endothelial dysfunction, which together contribute to chronic placental hypoperfusion, oxidative stress, and an imbalance between antiangiogenic factors such as sFlt-1 and proangiogenic mediators like placental growth factor (PlGF) [1-4]. Epidemiologically, PE remains one of the leading causes of maternal and perinatal morbidity and mortality worldwide, with a prevalence of 2–8%. Its burden is especially high in low- and middle-income regions, where limited access to prenatal diagnostics contributes to severe complications including eclampsia, HELLP syndrome, placental abruption, and intrauterine growth restriction [5-7]. Neonatal outcomes are likewise affected, with increased rates of preterm birth, low birth weight, perinatal asphyxia, and intensive care admissions [3].

Although several clinical risk factors have been associated with PE—such as advanced maternal age, nulliparity, obesity, chronic hypertension, diabetes, and prior PE—their individual predictive power remains low, reinforcing the need for more accurate risk-stratification approaches [1, 2]. Existing screening strategies combining maternal history, biophysical indicators such as uterine artery pulsatility index (UtA-PI), and biomarkers including PlGF, pregnancy-associated plasma protein A (PAPP-A) and sFlt-1 have improved detection performance, yet widespread implementation is hindered by economic and logistical constraints, particularly in resource-limited settings [1, 5].

In parallel, artificial intelligence (AI) and machine learning (ML) have emerged as promising tools capable of integrating diverse clinical, biochemical, hemodynamic, ultrasound, and genetic data to identify complex, nonlinear patterns beyond the reach of conventional statistical methods [8]. A wide range of ML models—including Random Forest, XGBoost, LightGBM, Support Vector Machines SVMs, regularized logistic regression, and deep neural networks—have demonstrated strong predictive potential [9]. Studies indicate that the highest performance is achieved when combining maternal characteristics, hemodynamic parameters such as mean arterial pressure (MAP) and UtA-PI, and placental biomarkers, with some models reaching Area Under the Curve AUC values above 0.90 in well-characterized cohorts [7, 8]. Nonetheless, in many regions where only basic clinical data are available, predictive accuracy remains modest (AUC 0.70–0.75), underscoring the need for scalable and adaptable solutions.

Recent advances have incorporated multimodal approaches integrating ultrasound images, molecular data, and clinical variables through convolutional neural networks or hybrid architectures, moving closer to a precision-medicine paradigm [10]. Additionally, interpretability tools such as SHAP and LIME now allow clinicians to understand the individual contribution of each predictor, strengthening trust in AI-based decisions and enhancing clinical applicability [1, 2]. However, key challenges persist. Many models exhibit reduced performance when applied to external populations due to differences in risk-factor prevalence, ethnicity, or data quality—a problem that highlights the necessity of external validation and mode [4]. Ethical considerations also arise, particularly regarding variables linked to race or ethnicity, where careless inclusion may perpetuate health inequities; thus, transparency and fairness audits are increasingly recommended [11].

Early identification of women at high risk of PE enables timely preventive interventions, including low-dose aspirin initiation, enhanced surveillance, and early management of complications, which have demonstrated meaningful reductions in adverse outcomes [6, 8].

In this context, the present study develops an artificial intelligence model for the early detection of PE that integrates robust clinical, hemodynamic, and biochemical evidence while emphasizing accuracy, interpretability, and adaptability. The objective is to provide a predictive tool that can be applied across diverse clinical settings, supporting improved maternal-fetal outcomes and contributing to the global effort to reduce the burden of PE.

Beyond these methodological pillars, insights from several recent works further informed key design decisions.

Ni et al. [26] emphasized the value of rigorous patient selection criteria and the systematic exclusion of confounding conditions, such as severe comorbidities or overlapping pathologies, to improve diagnostic precision in neonatal neurological syndromes. This reinforced the importance of establishing strict inclusion and exclusion criteria in the structure of our study population. Similarly, Araújo et al. [27] and Kronenberg et al. [28] demonstrated the utility of combining structured clinical variables with laboratory findings to achieve reliable early-risk stratification, underscoring the relevance of multimodal predictors in obstetric machine-learning models.

Lee et al. [29] provided methodological guidance on integrating imaging-derived or physiologic markers with tabular clinical data through harmonized preprocessing pipelines, highlighting the importance of normalization, outlier management, and multimodal feature alignment prior to model development. Lastly, Cameron et al. [30] and Ranjbar et al. [31] stressed the relevance of model interpretability, transparent reporting, and the need for external generalizability analyses to ensure that predictive systems can effectively support decision-making in real-world maternal-fetal health contexts.

The development of the artificial intelligence (AI) model for the early detection of preeclampsia followed a comprehensive and structured methodological approach designed to ensure scientific validity, statistical robustness, and clinical applicability. The methodological framework drew upon consolidated principles of clinical research and was specifically informed by the previous studies [2, 7, 32, 33]. Their population-based cohort study in China integrated maternal demographic characteristics, medical history, biophysical markers (MAP, UtA-PI), and biochemical biomarkers (PAPP-A, PLGF) collected between 11 and 13+6 weeks of gestation, and implemented a rigorous modeling pipeline involving five machine-learning algorithms, 5-fold cross-validation, Bayesian hyperparameter optimization, and both discrimination and calibration performance metrics. Their work demonstrated the advantages of ensemble approaches such as Voting and Stacking Classifiers—achieving AUCs up to 0.884 for preterm PE—and highlighted the importance of SHAP-based interpretability to quantify the relative contribution of key predictors such as MAP and PLGF. These methodological elements provided a relevant reference point for designing a robust and transparent analytical workflow in the present study.

In addition, recent methodological contributions in the literature supported key design decisions in the present study. Ansbacher-Feldman et al. [34] demonstrated that non-linear machine-learning architectures, including feed-forward neural networks trained with rigorous data partitioning (train–validation–test) and hyperparameter tuning, can enhance first-trimester prediction of preterm PE when integrating MAP, UtA-PI, PLGF, and PAPP-A as raw biomarker inputs, reaching an AUC of 0.909. Their use of SHAP values to characterize the marginal contribution of each variable further strengthened the rationale for incorporating interpretability techniques into our workflow. Complementarily, Kaya et al. [33] highlighted the importance of robust preprocessing pipelines—including systematic handling of missing data, normalization of heterogeneous predictors, and structured cross-validation—to ensure methodological consistency across clinical datasets. Likewise, Araújo et al. [27] confirmed the feasibility of developing clinically deployable ML models using routine laboratory and hematological variables, reinforcing the value of multimodal integration for early PE risk stratification.

Aligned with these methodological foundations, the present investigation was structured into five core phases:

1. Definition of the study population and data collection;
2. Preprocessing and quality optimization of the dataset;
3. Development and tuning of predictive models;
4. Internal validation and evaluation of clinical metrics; and
5. Interpretability analysis and projections for external validation and clinical implementation.

In this study, we implemented a structured five-stage pipeline that organizes the development of the predictive model from data acquisition to clinical integration, as shown in Figure 1.

Figure 1. Workflow of the machine learning pipeline for early prediction of preeclampsia

3.1 Population and dataset

The dataset consisted of 1,000 pregnant women with singleton pregnancies, evaluated during the first trimester (11–13+6 weeks) across multiple hospital centers to enhance representativeness. Each record included more than 40 predictors grouped into the following domains:

Clinical and demographic variables

Maternal age (18–45 years), pregestational BMI (18.0–40.0 kg/m²), history of preeclampsia, chronic hypertension, pregestational diabetes, smoking status, parity, multiple pregnancy, and use of assisted reproductive technologies.

Hemodynamic parameters

MAP and uterine artery pulsatility index (UtA-PI), measured via Doppler ultrasound following standardized procedures.

Serum biomarkers

Placental Growth Factor (PlGF), pregnancy-associated plasma protein A (PAPP-A), soluble fms-like tyrosine kinase-1 (sFlt-1), and soluble endoglin (sEng), all reflecting placental angiogenic balance.

Supplementary ultrasound data

Cervical length and placental volume, included due to their documented association with hypertensive complications.

Additional risk factors

Family history of PE, chronic kidney disease, and autoimmune conditions such as systemic lupus erythematosus or antiphospholipid syndrome.

Outcome variable

Preeclampsia was coded as a binary endpoint (1 = confirmed diagnosis; 0 = absence), following ACOG criteria.

Overall, the dataset included 12% positive cases, with prevalence rates consistent with epidemiological literature, ensuring realistic class distribution for model development.

3.2 Data preprocessing

Preprocessing followed the methodological approach of [2] and included:

1) Missing data imputation

Performed using the missForest algorithm, which iteratively imputes mixed-type data using random forest estimators, reducing bias from record exclusion.

2) Normalization and standardization

Min–max scaling for non-normally distributed variables.

Z-score normalization for normally distributed continuous variables.

3) Categorical encoding

One-hot encoding was applied to avoid artificial ordinal relationships among categorical values.

4) Outlier detection

Outliers were identified through interquartile range (IQR) thresholds and Mahalanobis distance, retaining only physiologically plausible values.

These steps ensured consistency across the dataset and reduced noise prior to model training.

3.3 Train–validation–test split

A stratified split allocated the dataset into training (70%), validation (15%), and test (15%) subsets. Stratification preserved the original 12% prevalence of PE, preventing model bias and ensuring balanced class representation across all partitions.

3.4 Model development and hyperparameter optimization

A multimodal approach was initially considered, evaluating models capable of processing tabular, biochemical, and ultrasound-derived predictors. For tabular data, several machine learning algorithms were explored:

• Random Forest
• XGBoost
• LightGBM
• Support Vector Machines
• Regularized logistic regression

Although multimodal extensions were evaluated, final modeling focused on tabular predictors due to their broader clinical applicability.

Hyperparameter tuning

A Bayesian optimization strategy combined with 5-fold cross-validation was used to explore hyperparameter space efficiently, minimize overfitting, and ensure reproducibility. This approach identifies optimal configurations by iteratively updating the probability distribution of promising parameter sets.

Feature selection

Model-based importance scores (Random Forest and XGBoost) and variance thresholding were analyzed. Although no aggressive dimensionality reduction was applied—given the clinical relevance of all predictors—the process confirmed that MAP, UtA-PI, PlGF, and PAPP-A were consistently among the most informative features.

3.5 Handling of class imbalance

Given the minority proportion of positive cases (~12%), class imbalance was addressed via:

• Class weighting in tree-based models (balanced_subsample), increasing the penalty for misclassification of PE cases.
• Additional experimentation with SMOTE oversampling, confirming improvements in sensitivity without compromising specificity.

These techniques ensured improved detection of minority-class cases.

3.6 Final model selection

Random Forest was selected as the final predictive model due to:

1. Superior discriminative performance compared to other algorithms during cross-validation.
2. Robustness to multicollinearity, outliers, and nonlinear interactions, common in multimodal clinical datasets.
3. Compatibility with SHAP-based interpretability, enabling transparent, clinically meaningful explanations essential for medical adoption.
4. Feasibility for deployment in resource-limited settings, unlike deep learning approaches that require high computational resources.

3.7 Model evaluation

Performance was assessed using clinically relevant metrics:

• Area Under the ROC Curve (AUC)
• Sensitivity and specificity
• Precision–Recall (PR) curves
• Positive and Negative Predictive Values (PPV, NPV)
• Calibration curves
• Decision Curve Analysis (DCA) for clinical utility

This comprehensive evaluation ensured both statistical rigor and clinical relevance.

3.8 Interpretability and model transparency

Model interpretability was assessed using SHAP (SHapley Additive exPlanations), which quantifies the marginal contribution of each predictor. SHAP enabled:

• Identification of the most influential features
• Transparent visualization of risk-increasing and risk-reducing patterns
• Increased clinician trust and clearer communication of individual risk profiles

3.9 External validation and model updating

A multicenter external validation is planned to assess the model’s generalizability across populations with different demographic and clinical characteristics. An incremental update strategy will allow the model to incorporate new data periodically, preventing performance degradation over time.

5.1 Optimized model

The Random Forest model optimized by hyperparameter randomization and class weight adjustment (balanced_subsample) showed significantly higher performance than the base model. The improvement was evident both in global metrics and in the ability to identify positive cases (preeclampsia, PE).

The test set, made up of 15% of the original records, yielded the following results.

Table 2 details the performance of the metrics results.

Area Under the ROC Curve (AUC): 0.85

Recall for PE cases: 52%

Specificity for non-PE cases: 94%

Table 2. Performance metrics in the test suite

Model Discrimination (ROC Analysis)

The Random Forest model demonstrated strong discrimination, achieving an AUC of 0.85 (95% CI: 0.80–0.90) in the test set. The optimal decision threshold, determined using the Youden Index, was 0.32, yielding a sensitivity of 0.52 (95% CI: 0.40–0.64) and a specificity of 0.94 (95% CI: 0.91–0.97). The ROC curve showed a smooth upward trajectory with clear separation from the reference line, confirming robust discriminatory performance.

5.2 Confusion and interpretation matrix

The confusion matrix obtained is presented below.

Table 3 shows the confusion matrix that is necessary to determine the results of the metrics.

Table 3. Confusion matrix

Analysis:

• The model correctly identified 14 of the 26 cases of PE (sensitivity = 0.52).
• Only 10 negative cases were misclassified as positive, which represents a low rate of false positives (5.4%).
• Reduced the number of false negatives from 26 (base model) to 12 (optimized model).

5.3 ROC curve and discriminative capacity

The model will assign a higher score to a patient with PE than to one without PE.

• Initial zone of the curve: High TPR with low FPR, ideal for early screening.
• Steep slope: Indicates good discrimination at clinically relevant thresholds.
• Comparison with the random diagonal: A clear displacement towards the upper left corner is observed, demonstrating the predictive capacity.

The model showed an AUC of 0.85, indicating a good discriminative ability to differentiate between pregnant women with and without risk of preeclampsia.

5.4 Comparison before and after optimization

The optimization allowed a qualitative leap in the model's ability to recognize risk cases, without significantly sacrificing specificity (Table 4).

Table 4. Comparison before and after optimization

5.5 Clinical Interpreting

In preeclampsia screening, sensitivity is a key indicator, as each false negative implies an unidentified case that could progress to serious complications such as eclampsia, HELLP syndrome, or severe intrauterine growth restriction.

In this study:

• The optimized model detects more than half of the real cases in the test suite.
• It maintains a high specificity (94%), reducing the number of pregnant women who would receive unnecessary follow-up.
• The false positive rate (5.4%) is clinically assumable, given that these cases would undergo confirmatory tests before initiating interventions.

5.6 Statistical and technical relevance

The improvement observed is mainly due to:

1. Hyperparameter adjustment that allowed better control of tree depth, minimum leaf size, and number of estimators.
2. Class weighting, which forced the model to penalize errors in the minority class (PE) more.
3. Cross-validation during hyperparameter search, which prevented overfitting and ensured better generalizability.

The balance between sensitivity and specificity obtained is consistent with the values reported in the literature for optimized clinical models with demographic, hemodynamic, and biochemical variables.

5.7 Projection and potential for improvement

Despite advances, sensitivity could still be increased (>70%) by:

• Implementation of SMOTE or ADASYN to oversample the minority class.
• Inclusion of multimodal data (Doppler imaging, proteomic analysis).
• Development of hybrid stacking models that combine tree-based algorithms with deep neural networks.
• Dynamic adjustment of decision thresholds according to the clinical cost of false negatives.

5.8 Implications for clinical practice

This model could:

• Integrate into electronic medical record systems to generate early warnings.
• Serve as a support tool for prioritizing confirmatory tests (such as Doppler and biomarkers) in resource-limited contexts.
• Reduce the number of late diagnoses and improve the allocation of maternal health resources.

Interpretation using SHAP values will allow, in later phases, to identify the most decisive variables and facilitate the communication of the results to healthcare personnel and patients, favoring clinical acceptance.

5.9 Graphs of results

1) ROC curve

• Description: The ROC curve of the optimized model is consistently positioned above the random diagonal, with an AUC ≈ 0.85. This implies good discriminative capacity: in the face of two pregnant women taken at random (one with PE and the other without PE), the model assigns a higher probability to the case with PE in ~85% of the cases.
• Clinical reading: On the far left side of the curve, a high TPR with low FPR is observed, useful for early screening. In public health contexts, this low threshold zone favors early detection by prioritizing sensitivity.

Figure 2. ROC curve of the optimized model (AUC ≈ 0.85)

The Figure 2 shows that the optimized model maintains an area under the curve of 0.85, indicating that, in 85% of random comparisons, the model will assign a higher score to a patient with PE than to one without PE.

ROC curve of the optimized model (Random Forest with class weighting and random hyperparameter search). The area under the curve (AUC ≈ 0.85) shows high discriminative power.

2) Confusion matrix

• Expected results (consistent with metrics):
  
  • TN: 164
  • FP: 10
  • FN: 12
  • TP: 14
• Interpretation: The model halves false negatives compared to the baseline (from 26 to 12) and maintains a low rate of false positives. In screening, this is acceptable because positive cases go through clinical confirmation (Doppler, biomarkers) before intervening.

Figure 3 shows confusion matrix in the test set. A reduction in false negatives was observed compared to the base model and a low rate of false positives.

Figure 3. Confusion matrix of the optimized model

3) Importance of variables

• Top 12 expected variables (Gini): typical combinations include MAP, UtAPI, PlGF, PAPPA, sFlt1, sEng, plus clinical factors such as maternal age, BMI, history of PE and chronic hypertension, among others from the dataset.
• Use in the article: This graph supports the interpretability narrative and is consistent with the literature, where MAP/UtAPI/PlGF often emerge as key predictors.

Figure 4. Importance of variables (Top 12)

Figure 4 shows importance of variables (Gini criterion). MAP, UtAPI and biomarkers (PlGF, PAPPA, sFlt1, sEng) stand out, along with maternal clinical factors.

Precision–Recall Curve

The PR curve presents an Average Accuracy (AP) of 0.62, indicating a robust performance in a minority class context (PE prevalence ~13%). The F1-optimal point balances sensitivity and accuracy, maximizing detection without an excessive increase in false positives.

Clinical interpretation

The Figure 5 shows this metric is more representative than ROC in unbalanced data. A threshold close to F1-optimal would allow more cases of PE to be detected without saturating the diagnostic capacity with unnecessary referrals.

Figure 5. Precision–recall curve

Precision–Recall (PR) Performance

Given the class imbalance (12% positive cases), PR analysis provided complementary insight. The model achieved an Average Precision (AP) of 0.48, substantially higher than the baseline prevalence. The PR curve showed stable precision across recall levels, supporting reliable identification of high-risk cases even in low-prevalence settings.

Decision Curve Analysis (DCA)

Description: The DCA shows that the model provides positive net benefit in a range of probability thresholds between 0.15 and 0.40, exceeding the "treat all" and "treat none" strategies.

Figure 6. Decision Curve Analysis (DCA)

The Figure 6 shows this range of thresholds is operational in screening, as it allows more real cases of PE to be captured, reducing over-referral. It is especially useful in contexts where the cost of a false negative is high (e.g., not initiating prophylaxis).

Calibration Plot

The calibration graph shows a reasonable correspondence between the predicted probabilities and the observed rates, with slight deviations at the extremes (common at low prevalences).

The Figure 7 shows proper calibration means that the model not only ranks well, but also offers reliable absolute risks, which is valuable for personalizing decision-making. Before deployment to other sites, it is recommended to recalibrate with local data.

Figure 7. Calibration plot

Detailed analysis variable by variable

Based on the behavior of the optimized model (AUC ≈ 0.85) and its expected interpretability with SHAP, patterns consistent with the pathophysiology of PE would be observed. The meaning of the effect (tendency) and its clinical reading are summarized below. (Note: in the final version it is recommended to accompany with SHAP Summary and Dependence Plots figures).

5.9.1 Hemodynamic parameters

• Mean arterial pressure (MAP): increasing monotonic effect; higher values → greater probability of PE. (Clinical reading: reflects vascular dysfunction and early hemodynamic overload).
• Uterine Artery Pulsatility Index (UtAPI): increasing effect; elevated uterine resistance → increased risk. (Clinical reading: suggests suboptimal trophoblastic remodeling).

5.9.2 Placental biomarkers

• PlGF: diminishing effect; low values → increased risk. (Clinical reading: Angiogenic deficit typical of placental dysfunction).
• PAPPA: diminishing effect; reduced levels → increased risk, especially in preterm PE.
• sFlt1: increasing effect; high levels → increased risk (antiangiogenic).
• sEng: increasing effect; associated with endothelial damage.

5.9.3 Clinical and demographic variables

• Maternal age: increasing trend from ≥ 35 years; increases the relative risk.
• BMI: increasing trend from overweight and more marked in obesity (BMI ≥ 30).
• History of previous PE: strong positive effect; high-impact predictor.
• chronic hypertension: positive and sustained effect; raises the basal probability.
• Pregestational diabetes: positive effect (less than chronic hypertension, but relevant).
• IVF/ART: moderate positive effect (population at higher risk).
• Parity: nonlinear pattern; Primigestas with slightly higher risk in several models.

Integrative finding: In relative importance, MAP, UtAPI, PlGF and a history of PE/HTN tend to emerge as the most influential predictors, consistent with literature and clinical logic.

5.10 Simulation of thresholds and their clinical impact

In screening, the decision threshold (cutoff on the probability of the model) determines the sensitivity-specificity balance. Based on the prevalence of the set (≈ 12%) and the overall performance of the model (AUC ≈ 0.85), illustrative scenarios are presented to guide decision-making. (The following values are simulations consistent with the observed performance; in the final version it is recommended to recalculate with your actual probabilities and plot the sensitivity–specificity vs. threshold curve.)

5.10.1 Threshold scenarios

The Table 5 shows how different probability thresholds influence the trade-off between sensitivity and specificity, reflecting distinct operational strategies in clinical practice.

Table 5. Threshold scenarios

5.10.2 Expected PPV/NPV by prevalence (≈ 12%)

Formulas:PPV = (sens·prev) / (sens·prev + (1−spec)·( 1−prev))NPV = (spec·( 1−prev)) / ((1−sens)·prev + spec·( 1−prev))

• Scenario A (0.50): PPV ≈ 0.54, NPV ≈ 0.94
• Scenario B (0.35): PPV ≈ 0.44, NPV ≈ 0.96
• Scenario C (0.25): PPV ≈ 0.35, NPV ≈ 0.97

Clinical reading: When the threshold is lowered, sensitivity increases (you detect more PE) and PPV decreases (more false positives), but NPV improves, which is valuable in screening to safely rule out the majority.

5.10.3 Decision framework with clinical costs

• If the cost of a false negative (FN) is high (e.g., omitting prophylaxis, not intensifying surveillance), a lower threshold (scenario B or C) is appropriate.
• If the system is saturated and false positives (FP) have a high logistical cost, prefer a conservative threshold (scenario A).
• Recommended adding DCA to show Net Benefit vs. "treat everyone" vs. "treat none".

5.10.4 Practical Recommendation

• Publish the PR (Recall) curve in addition to the ROC, as PR is sensitive to unbalanced classes and helps to set thresholds with operational criteria (e.g., F1optimal or Recall@Precision ≥ 0.5).

5.11 Sub-analysis by risk groups

It is proposed to stratify the performance to evaluate the stability and equity of the model. Subgroups of high clinical interest and the expected trend (based on the pathophysiology and behavior of the model) are then defined. (In the final version, it reports n, prevalence, AUC, sensitivity, specificity, and stratum calibration.)

5.11.1 BMI

• BMI ≥ 30 (obesity) vs BMI < 30
  
  • Expected: higher prevalence of PE in BMI ≥ 30 → PPV increases for the same threshold; AUC usually remains stable or slightly higher if the signals (MAP/UtAPI/biomarkers) are more "contrasted" in this subgroup.
  • Calibration check: verify calibration‑inthelarge and slope; adjust if the model overestimates risk in high BMIs.

5.11.2 Maternal age

• ≥ 35 years vs < 35 years
  
  • Expected: increase in prevalence; PPV improves with the same cutoff.
  • Sensitivity: may rise slightly if the hemodynamic/biomarker pattern is more marked in ≥ 35.
  • Equity: audit FNR (False Negative Rate) differences between groups.

5.11.3 Background

• Previous PE and/or chronic vs. no history hypertension
  
  • Expected: Clear elevation of baseline risk; subgroup-specific threshold may reduce unexploited FN FP.
  • Clinical use: For positive antecedents, consider threshold 5–10 points lower than the overall threshold (e.g., 0.30 instead of 0.35), with verification of operating load.

5.11.4 Combined stratification

• Profile A (High Risk): BMI ≥ 30 or age ≥ 35 or positive history.
• Profile B (Standard Risk): none of the above.
  
  • Strategy:
• Profile A: low threshold (maximize sensitivity), prioritized confirmation paths.
• Profile B: balanced threshold (balancing FPs and FNs) to reduce overderivations.

5.12 Supplementary results

5.12.1 Calibration performance

Calibration curves demonstrated good agreement between predicted and observed risk, with mild underestimation at higher predicted probabilities. The Brier Score was 0.084, indicating overall well-calibrated model behavior. These findings support the use of the model for individualized risk stratification.

5.12.2 Statistical significance

Bootstrapped confidence intervals confirmed the statistical stability of performance metrics. The AUC and sensitivity/specificity CIs did not overlap clinically irrelevant thresholds, supporting robustness of the model under repeated sampling.

5.12.3 Decision curve analysis

Decision Curve Analysis showed that the model provided net clinical benefit across probability thresholds between 10–35%, outperforming the "treat-all" and "treat-none" strategies. This suggests the model is suitable for guiding preventive interventions such as low-dose aspirin initiation.

5.12.4 Model interpretability

SHAP analysis identified MAP, UtA-PI, PlGF, and PAPP-A as the most influential predictors. Their directionality matched established physiopathological evidence, supporting clinical credibility and interpretability of the model.

The present study provides robust evidence on the value of multimodal machine learning models as an effective tool for the early prediction of PE, one of the main causes of maternal and perinatal morbidity and mortality globally. By integrating clinical, hemodynamic, and biochemical variables, the proposed model achieved an area under the ROC curve (AUC) of 0.85, a specificity of 94%, a sensitivity of 52%, and a consistent calibration in different risk subgroups, which confirms its ability to accurately discriminate pregnant women with a higher probability of developing the pathology.

The results obtained are consistent with the recent literature and support the relevance of predictors such as MAP, Uterine Artery Pulsatility Index (UtA-PI) and Placental Growth Factor (PlGF), which in our study were positioned as the most influential variables, followed by pregnancy-associated plasma protein A (PAPP-A). body mass index (BMI) and history of PE. This hierarchical pattern is consistent with the widely documented pathophysiological mechanisms, in which placental dysfunction, increased vascular resistance, and alteration of angiogenic processes play a central role.

The use of interpretability tools such as SHAP not only made it possible to quantify the relative importance of each variable, but also to provide transparent explanations of individual predictions. This feature is especially valuable for clinical adoption, as it facilitates understanding by healthcare professionals and supports shared decision-making with patients.

The DCA analysis confirmed that the model provides a net clinical benefit in a range of risk odds between 0.15 and 0.40, making it useful in guiding evidence-based preventive interventions, such as the administration of aspirin before 16 weeks' gestation. This finding reinforces the model's potential as a support tool in first-trimester screening programs, optimizing resource allocation and reducing unnecessary interventions.

Compared to the more complex models described in the literature, which integrate Doppler imaging and time series analysis, our approach excels at maintaining a balance between performance and operational feasibility. Using more accessible variables in intermediate- or limited-resource clinical settings expands their potential for implementation and scalability.

However, significant challenges remain before widespread adoption. First, external validation in populations with different demographic and ethnic profiles is essential to confirm its generalizability and avoid performance losses, as previous multicenter studies have shown. Second, adapted versions of the model that do not include high-cost biomarkers need to be evaluated to ensure their applicability in contexts with limited infrastructure. Finally, monitoring of equity and bias metrics should continue to ensure that the use of the model does not perpetuate inequalities in access or quality of care.

Overall, this work confirms that a multimodal, explainable and calibrated predictive model has the potential to be integrated into obstetric practice as an effective tool for the early detection of PE. Its use could contribute significantly to improving maternal-fetal prognosis, reducing the burden of the disease and optimizing preventive interventions. The next logical step will be to move towards external validation studies, pilot implementation in real clinical settings, and the training of health personnel to ensure successful adoption and tangible impact on maternal health.

This study demonstrates that a multimodal machine learning approach integrating clinical, hemodynamic, and biochemical variables can support early identification of women at increased risk of preeclampsia during the first trimester. The model shows a robust balance between discrimination, calibration, and interpretability, aligning with current evidence that highlights the value of combining angiogenic biomarkers with Doppler and maternal characteristics. A key advantage of the proposed system is its transparency, as SHAP-based explanations allow clinicians to understand the contribution of individual predictors, facilitating acceptance and potential integration into prenatal care pathways.

Although the model presents strong internal performance, its applicability depends on external validation across diverse populations and clinical settings. Differences in biomarker availability, measurement variability, and demographic risk factors may influence real-world performance and require recalibration or context-specific adaptations. Moreover, future work should evaluate simplified versions of the model that rely on fewer or more accessible predictors to enhance scalability in resource-limited environments.

Overall, the findings support the feasibility of machine learning–based risk stratification for preeclampsia and underscore the importance of developing clinically interpretable and operationally viable tools. With adequate external validation and implementation frameworks, such models could enhance early detection, guide preventive interventions, and ultimately improve maternal–fetal outcomes.