Development of antenatal risk prediction model for preeclampsia with severe features in Lagos, Nigeria (PreSev study): protocol of a prospective cohort study

Abstract

Background

Preeclampsia with severe features is a major contributor to maternal and perinatal morbidity and mortality, particularly in low- and middle-income countries (LMICs). Despite improvements in obstetric care, early identification of women at highest risk remains limited, and existing prediction tools lack contextual specificity for African populations.

Objective

This study describes the protocol for the PreSev Study, which aims to develop and validate a machine-learning risk prediction model for preeclampsia with severe features among pregnant women in Lagos, Nigeria.

Methods

This multicentre prospective cohort study will recruit at-risk pregnant women aged ≥ 18 years between 14 and 24 weeks’ gestation from four major hospitals in Lagos State between September 2025 and December 2026. A total of 938 women will be enrolled for model development (training set n = 750) and independent validation (n = 486). Data on epidemiologic and clinical predictors of preeclampsia will be collected using REDCap. Five machine-learning models, random forest, extreme gradient boosting (XGBoost), multilayer perceptron (MLP), LightGBM, and glmnet, will be trained using 10-fold cross-validation and hyperparameter optimisation. Model performance will be assessed using area under the curve (AUC), calibration, decision curves, and clinical performance metrics. SHAP values will be used to evaluate feature importance. The best-performing model will be converted into an accessible online risk-prediction tool.

Conclusion

The PreSev Study will generate a context-specific, clinically applicable machine-learning model to support early risk stratification for preeclampsia with severe features, with potential to enhance maternal outcomes in Nigeria and similar low-resource settings.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12884-026-09537-9.

Introduction

In 2017, an estimated 295,000 women worldwide lost their lives due to complications related to pregnancy and childbirth, with approximately 94% of these deaths occurring in resource-limited settings [1]. Sub-Saharan Africa and Southeast Asia accounted for 254,000 of these deaths, representing about 86% of the global maternal mortality burden [1]. Hypertensive disorders of pregnancy (HDP), including preeclampsia with severe features, affect approximately 10.0% of pregnant women globally [2, 3] and account for 14·0% of maternal mortality globally [4].

Preeclampsia is a progressive multisystemic complication of pregnancy characterised by hypertension and often proteinuria, typically occurring after 20 weeks of gestation [5]. It remains a leading cause of maternal and perinatal morbidity and mortality worldwide, particularly in low- and middle-income countries [6], including Nigeria. Preeclampsia with severe features represents an escalation of preeclampsia severity and is associated with a higher risk of adverse maternal and fetal outcomes [7, 8]. Severe features of preeclampsia include a systolic blood pressure of at least 160 mmHg or a diastolic blood pressure of at least 110 mmHg, thrombocytopenia, impaired liver function, severe persistent right upper quadrant pain, progressive renal insufficiency, pulmonary oedema, and new-onset cerebral or visual disturbances [9].

Despite significant advancements in obstetric care, the incidence of preeclampsia with severe features remains an important public health challenge [10], and effective strategies to accurately identify women at risk of developing the disease in obstetric populations remain limited, particularly in low-resource settings. In Nigerian surveillance for maternal near-miss and maternal death involving over 100 thousand admissions, 6753 (7%) had HDP, consisting of 54.6% preeclampsia cases [11]. Of women with HDP, 298 and 289 women had maternal near-miss and maternal death, respectively [11]. In a study on pregnancy complication-readiness in Lagos, only half of 524 pregnant women were reported to be complication-ready [12]. Hence, early identification of at-risk women could improve maternal and perinatal outcomes through the early introduction of effective prophylactic treatments such as low-dose aspirin [13], the commencement of close monitoring of fetal growth and development, the provision of timely delivery, and the mitigation of long-term complications for mothers, such as chronic hypertension and cardiovascular disease [10]. However, current literature and clinical guidelines often rely on clinical risk factors [14–16], which may lack specificity for local populations, thus necessitating the need to develop context-specific risk prediction models. Therefore, the proposed PreSev study will use known maternal risk factors of preeclampsia, including nulliparity, advanced age, high body mass index (BMI), prior hypertension, diabetes mellitus, prior preeclampsia, and medical conditions [17–21], to develop and validate a machine learning prediction model for preeclampsia with severe features and then design a risk-scoring system among a cohort of at-risk pregnant women in the second trimester of pregnancy (between 14- and 24-weeks’ gestation) of four selected hospitals in Lagos, Nigeria.

Participants and methods

Study design and settings

The “PreSev Study” is a prospective cohort study of at-risk pregnant women attending antenatal clinics at four hospitals in Lagos, Nigeria, from September 2025 to December 2026. The hospitals – Federal Medical Center (FMC-Eb) in Ebute Meta, Lagos Island Maternity Hospital (LIMH) in Lagos Island, Lagos University Teaching Hospital (LUTH) in Idi-Araba, and Lagos State University Teaching Hospital (LASUTH) in Ikeja – are the leading public health institutions in Lagos State in South-West Nigeria with a population of over 20 million residents. These hospitals have established obstetrics departments with well-equipped and functional maternal units, collectively delivering 12,620 babies annually.

Study population

Potentially eligible participants in the study will include pregnant women aged 18 years or older at 14 to 20 weeks gestational age who attend their routine antenatal care at the participating hospitals and are identified as being at moderate-to-high risk of preeclampsia based on established risk factors such as age 40 years or older, nulliparity, pregnancy interval of more than 10 years, body mass index (BMI) of 35 kg/m² or more at the first visit, family history of pre-eclampsia, multifetal pregnancy, hypertensive disease during a previous pregnancy, and medical conditions such as chronic hypertension, chronic kidney disease, sickle cell anaemia, autoimmune diseases (systemic lupus erythematosus or antiphospholipid syndrome), and type 1 or type 2 diabetes mellitus [17–22]. Inclusion criteria include women free from any mental or physical disabilities that inhibit them from understanding the implications of the study and those not considering relocating from their current residence within the next year. Exclusion criteria are coagulation disorders and hepatic impairment, lack of willingness to continue in the study or withdrawal of consent at any time during the follow-up period, first antenatal presentation after 24 weeks, and termination of pregnancy before 20 weeks.

Sample size calculations

In building the risk prediction model, we estimated the sample size using the maximum modelling principle [23], given that a minimum of 10 events (women with preeclampsia with severe features) would occur for each incorporated prediction variable. Assuming an expected incidence of 20% for preeclampsia with severe features in at-risk pregnant women and potential predictor variables of 15, a minimum sample size of n = 500 would be required as the training dataset to satisfy the model building and then develop a scoring system. Therefore, assuming an attrition rate of 20% between antenatal enrolment at a gestational age of 14–24 weeks and the occurrence of the study outcome (preeclampsia with severe features), we will enrol 938 at-risk pregnant women to achieve the sample size of 750 required for the model building. To validate and assess the model’s predictive accuracy and reliability [24] at a sample area under the curve (AUC) of 0.7 corresponding to a desired type I error rate of 5% at a 95% confidence interval, we will require a minimum sample size of 486 women in the validation dataset.

Sampling and data collection techniques

Prospective participants will be identified by the study coordinators and other clinical team members during their antenatal booking visit at three of the four participating sites (FMC-Eb, LUTH, and LASUTH) for eligibility for inclusion in the training dataset, while eligible participants at the fourth site (LIMH) will be simultaneously enrolled during the same study period for inclusion in the validation dataset. Informed written consent will then be obtained after an explanation of the purpose and procedures of the study, and eligible pregnant women will then be enrolled by the clinical research assistants until the required sample size is achieved. Using a Data Entry form created on REDCap, information on epidemiologic and clinical variables or factors that have been linked to preeclampsia, such as maternal age (in years), gestational age (in weeks), parity, type of pregnancy (singleton or multiple), BMI (in kg/m²), interpregnancy interval (in years), family history of preeclampsia, multifetal pregnancy, previous hypertensive disorders of pregnancy, mean arterial blood pressure, medical conditions such as chronic hypertension, chronic kidney disease, sickle cell anaemia, autoimmune diseases (systemic lupus erythematosus or antiphospholipid syndrome), and type 1 or type 2 diabetes mellitus [25]. The women will be encouraged to adhere to routine antenatal clinic attendance and to present in the labour ward at the onset of labour or planned delivery. The primary outcome, preeclampsia with severe features, is defined as the presence of either systolic blood pressure of at least 160 mmHg or diastolic blood pressure of at least 110 mmHg, thrombocytopenia, impaired liver function, severe persistent right upper quadrant pain, progressive renal insufficiency, pulmonary oedema, and new-onset cerebral or visual disturbances [9].

Statistical analysis

Data analyses will be performed using Stata version 18.0 (StataCorp LLC, Texas, USA). Missing data will be handled by running a missing values analysis if no pattern is detected and the multiple imputation method if a pattern is detected [26]. Features with a missing value ratio of more than 20% will be excluded from potential predictors. Continuous variables will be tested for normality using the Kolmogorov–Smirnov test with Lilliefors’ significance correction. Subsequently, descriptive statistics will be calculated to summarise the participants’ obstetric and clinical characteristics. We will then develop machine learning (ML) algorithms, including random forest (RF), extreme gradient boosting (XGBoost), multilayer perceptron (MLP), LightGBM, and glmnet models. RF is an ensemble method that constructs multiple decorrelated, unpruned decision trees to enhance prediction accuracy [27]. XGBoost is a gradient-boosting model that builds simple decision trees sequentially, with each tree improving the performance of the previous tree [28]. MLP is an artificial neural network comprising an input layer, one or more hidden layers, and an output layer to create complex, layered architectures [29]. LightGBM is a gradient-boosting framework that employs Gradient-based One-Side Sampling and Exclusive Feature Bundling to reduce data volume and the number of features, enabling faster training [30]. Lastly, glmnet applies penalised maximum likelihood estimation to fit various regression models, including linear, logistic, multinomial, Poisson, and Cox regression [31]. We will train all classifiers on the training dataset using 10-fold cross-validation to evaluate the bias and variance of each model [32]. This process will be repeated 10 times to account for the variability inherent in cross-validation and to minimise the risk of overfitting or underfitting [33]. Hyperparameter tuning will be performed using a grid search method to maximise the area under the receiver operating characteristic (AUC) curve. The ML algorithm with the highest AUC will be selected for further model development with reduced features. Feature importance will be evaluated using SHapley Additive exPlanations (SHAP) values, a post hoc explanation tool that visualises each predictor’s contribution to the ML models [34]. The final model’s performance will be validated on an independent (or validation) dataset, assessing metrics such as AUC, calibration curves, and decision curves. The AUC of the final model will be compared to the original model using DeLong’s non-parametric test [35]. We have predefined an AUC threshold of 0.7 as a benchmark for acceptable model performance. A calibration plot will be generated by stratifying data into two risk groups (low- versus high-risk) based on predicted probabilities of preeclampsia with severe features. The plot will compare predicted probabilities (with 95% confidence intervals) against observed probabilities, with an ideal calibration curve appearing as a 45-degree straight line. To provide additional insights into model performance, sensitivity, specificity, accuracy, negative predictive value, and positive predictive value with 95% CI will be calculated using the bootstrapping method. The optimal cutoff point for the final model will be determined using the Youden index [36]. The final prediction model for preeclampsia with severe features will then be deployed as an accessible online application.

Ethical considerations

This study will be conducted in accordance with the ethical principles outlined in the Declaration of Helsinki, the Council for International Organizations of Medical Sciences (CIOMS) International Ethical Guidelines for Health-related Research Involving Humans, and Nigeria’s National Code of Health Research Ethics. The protocol has received ethical and administrative approvals from the Health Research Ethics Committee (HREC) of the participating health institutions (Protocol Approval Numbers: ADM/DSCST/HREC/APP/7474 and LREC/06/10/2928).

Discussion

This protocol outlines the design and methods for developing and validating a machine-learning–based antenatal risk prediction model for preeclampsia with severe features among pregnant women receiving care in Lagos, Nigeria. Early and accurate identification of women at high risk is essential for reducing the substantial maternal and perinatal morbidity and mortality associated with hypertensive disorders of pregnancy, which remain a leading cause of adverse outcomes in low- and middle-income countries. Existing risk assessment tools are largely derived from high-income settings, may not incorporate context-specific epidemiologic patterns, and often demonstrate suboptimal predictive performance in African populations. The PreSev Study addresses this gap by using locally relevant predictors and contemporary machine learning approaches to improve precision risk stratification in routine antenatal care.

The multicentre prospective cohort design strengthens the validity and generalizability of the planned model across Lagos State’s diverse obstetric population. The use of REDCap for standardised data collection, rigorous handling of missing data, and the evaluation of multiple machine learning algorithms, including RF, XGBoost, MLP, LightGBM, and glmnet, will enable the selection of the best-performing model based on discrimination, calibration, and clinical usefulness. Independent external validation will further enhance reliability and reduce the risk of overfitting. The inclusion of SHAP values will provide interpretable insights into predictor importance, facilitating clinical acceptance and supporting transparent model deployment.

The proposed study may be limited by the reliance on clinical and epidemiologic predictors without laboratory biomarkers, which may reduce predictive accuracy compared with biomarker-enhanced models. However, this approach prioritises feasibility, scalability, and cost-effectiveness in low-resource settings. By converting the final model into a user-friendly online application and a practical scoring system, the PreSev Study aims to support frontline clinicians in making timely decisions regarding surveillance, prophylaxis, and referral. Ultimately, this study may contribute to reducing Nigeria’s high burden of preeclampsia-related complications and complement ongoing efforts to strengthen maternal health in the country.

Authors' contributions

BOO and KSO made substantial contributions to the concept or design of the work in this paper. BOO and KSO drafted the article. All the authors () revised the manuscript critically for important intellectual content and approved the version to be published.

Funding

The authors received funding for this work from Grand Challenges Nigeria (GCNg). The content of this paper is solely the responsibility of the authors. It does not necessarily represent the official views of GCNg. The funder had no role in the conceptualisation, decision to publish, or preparation of this manuscript.

Data availability

No data are associated with this article. The authors intend to grant public access to the full protocol, participant-level dataset, and statistical code associated with this study.

Declarations

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.