跳到正文
原文
Nature 健康护理主题·· 9 天前精选AI 评分70

利用常规数据评估机器学习预测妊娠并发症和不良出生结果

Evaluation of machine learning-based prediction of pregnancy complications and adverse birth outcomes using routine data

AI 导读

该研究利用大规模电子健康记录数据,评估机器学习模型在预测妊娠并发症和不良出生结果方面的潜力。结果显示,某些并发症在早期妊娠阶段具有较好的预测性能,但整体性能仍需提高。

推荐理由

研究利用大规模电子健康记录数据,评估机器学习模型在预测妊娠并发症和不良出生结果方面的潜力。

正文

Introduction

Maternal mortality rates in the United States are higher than that of any other high-income nation, with a disproportionate burden on historically marginalized groups, including Black and Indigenous women, women in rural areas, and those with limited access to healthcare1,2,3. A leading contributor to maternal mortality is underlying pregnancy and adverse birth outcomes4,5,6, which can lead to severe maternal harm, neonatal complications, and long-term morbidity5,6,7. Improving early identification of these complications offers an opportunity for timely interventions8,9,10 that could reduce mortality and health inequities.

Data-driven machine learning (ML) modeling in health has demonstrated strong early prediction performances in a range of clinical tasks11. However, work in early diagnostic prediction for pregnancy complications has been limited to cohorts evaluating risk factors after disease diagnosis12, non-representative study populations13,14, or used specialized data that is not routinely available (e.g., retinal vascular features15, polygenic risk scores16, or specialized ultrasound biomarkers17). Such models have limited utility in practice for practical early prediction18; However, this is precisely where there is the most need.

Currently, standard clinical workflows for perinatal risk stratification rely primarily on episodic screenings and qualitative history-based checklists. For gestational diabetes mellitus (GDM), while some risk-based screening is conducted in the first trimester, universal screening in the United States is typically anchored between 24 and 28 weeks of gestation19,20, a timeframe that may limit the efficacy of early lifestyle-based prevention. For hypertensive disorders such as pre-eclampsia (PE), the U.S. Preventive Services Task Force guideline-based checklists are used to identify candidates for early aspirin prophylaxis21; however, these tools can flag nearly 90% of a cohort as being at increased risk22, resulting in low specificity and a failure to account for dynamic, longitudinal physiological changes.

Our study is a pragmatic evaluation of the potential that ML models have in predicting pregnancy complications. Instead of demonstrating that a specific model can offer improvement on a specific single or set of targets23,24, we conduct a thorough audit of the predictability (or lack thereof) of different complications. Importantly, we use only routinely collected electronic health record (EHR) data, demonstrate the potential benefits of current ML tools to improve pregnancy outcomes through available treatments, and underscore current limitations in the field. We evaluate models using two large samples of patients. First, we use pre- and during- pregnancy data from a population-based cohort of 442,760 pregnancies from Kaiser Permanente Northern California (KPNC), to train models every six weeks of gestation for prediction of seven pregnancy and adverse birth outcomes: PE, gestational hypertension (GHTN), GDM, preterm birth, stillbirth, placental abruption (PA), and small for gestational age (SGA). We then validate these results on an external cohort of 6105 pregnancies from the national All of Us dataset25. We a priori focus on outcomes with a variety of negative effects, and leading contributions to maternal and neonatal morbidity and mortality26,27,28,29,30,31,32,33. Beyond model performance, we evaluate whether early prediction translates to actionable clinical benefit through cost-effectiveness analyses that pair model-based risk stratification with evidence-based interventions. To the best of our knowledge, there has been no other holistic evaluation of data-driven pregnancy and birth complication prediction across large, heterogeneous populations.

First, we evaluate the potential of in-distribution models in the KPNC cohort. Across 49 evaluations (seven early prediction windows for seven outcomes), we find that only three outcomes (PE, GHTN, GDM) have sufficiently strong performance, calibration and treatment cost-effectiveness analyses to support practical screening. We find that strong performance is driven by different data features over time and per condition, emphasizing the necessity of longitudinal, temporally-aware modeling, which has been previously limited34,35,36. We perform subgroup robustness checks in all cases, for instance, evaluating whether models over-predict risk for PE in Non-Hispanic (NH) Black patients as compared to their ground truth diagnoses. This analysis is important as some causes, e.g., missing early blood pressure measurements, indicates model gaps are addressable with future data collection improvement37. Second, we examine model transferability, evaluating models in an external cohort of pregnancies. We find that pregnancy models must be tuned to individual settings: even KPNC models finetuned with up to 50% of the All of Us data do not consistently perform as well as models trained from scratch in that sample. Finally, we survey prior modelling literature with reported higher screening performance23,24,38,39,40. We find that a majority of studies are either not population based or utilize specialized features, which lead to limited potential applicability18.

Our findings indicate that risk-based ML models are plausible high-precision, continuous risk stratification systems during the first and early second trimesters for PE, GHTN, GDM. Our work also highlights that there is a critical gap in modelling capacity for other conditions. Importantly, since our models rely solely on routinely collected EHR features, our findings may be applicable, with further external validation, to other clinical settings where timely intervention can improve maternal and neonatal health8,9,10, especially for vulnerable populations who may face barriers to comprehensive prenatal care41.

Methods

Dataset description

This population-based cohort study utilized data from KPNC, a large integrated, not-for-profit, prepaid health plan and delivery care system that serves over 4.6 million members across 14 counties, including approximately 45,000 deliveries annually. KPNC provides comprehensive inpatient, emergency, and outpatient care through its 21 hospitals and more than 255 outpatient clinics. The members are highly representative of the general population residing in the served geographic area in terms of age, sex, race and ethnicity, neighborhood-level income, education, and social vulnerability42. Routine prenatal care within KPNC includes standardized collection of vital signs (including blood pressure) and universal gestational diabetes screening using a glucose challenge test typically performed between 24 and 28 weeks, which supports high completeness and consistency of key predictors and outcome ascertainment in the EHR. This study was approved by the KPNC Institutional Review Board (IRB# 00001045) and waived the requirement for informed consent from participants for the use of de-identified data.

There were 448,953 live and stillbirth deliveries who received prenatal care at KPNC between January 1, 2011 and October 31, 2021. Data on physician diagnosis (International Classification of Diseases-9/10 codes), vital signs (e.g., anthropometrics, heart rate, blood pressure), lab tests, and medications were extracted from the EHRs system and the Virtual Data Warehouse. The KPNC Institutional Review Board approved this project and waived the requirement for informed consent from participants for use of de-identified data.

In the KPNC cohort, self-identified race and ethnicity, as social rather than biological constructs, were obtained from the EHRs data. Individuals with Hispanic origin were classified as Hispanic, regardless of reported race. Non-Hispanic (NH) individuals were categorized as NH Asian/Pacific Islander, NH Black, NH White, and Other/Unknown (see Supplementary Table 1). Given the small percentage of individuals in the multiracial (3.2%) group who self-identified as more than one race, we merged them into the Other group, which included American Indian and Alaska Native individuals.

External validation was performed using data from the All of Us Research Program, a nationwide cohort study launched by the National Institutes of Health in May 2018 to advance precision medicine and health equity research. The program enrolls a diverse group of participants across the United States through a network of over 340 recruitment sites, including health care provider organizations and direct volunteer enrollment locations. As of this analysis, the program had enrolled over 800,000 participants43 with 77% from communities historically underrepresented in biomedical research and 46% from racial and ethnic minorities. Data elements include EHRs, standardized physical measurements, health questionnaires, biospecimens, and digital health technology data. EHR data are harmonized across sites and vendors using the Observational Medical Outcomes Partnership (OMOP) common data model to facilitate interoperability and large-scale research applications.

For this study, we identified patients from the All of Us cohort with live or stillbirth deliveries and available prenatal care records. The analytic cohort included 6105 pregnancy episodes with sufficient vital signs and laboratory data for model evaluation. The cohort was racially and ethnically diverse, with 42.9% Hispanic, 32.6% non-Hispanic White, 16.2% non-Hispanic Black, 3.4% non-Hispanic Asian, and 5.0% other racial/ethnic groups. Data on physician diagnoses (International Classification of Diseases-10 codes), vital signs, laboratory tests, and medications were extracted from harmonized EHR data. Use of de-identified All of Us data is approved under the program’s central institutional review board and data use agreement25.

Outcome definitions

We selected three common pregnancy complications to predict, based on clinical significance and data availability. Three of these complications were “pregnancy” complications, defined as complications that happen during the gestational period. These outcomes were PE, GHTN, and GDM. They are further defined as follows, based on the guidance from the American College of Obstetricians and Gynecologists:

  1. 1.

    Pre-eclampsia: Preeclampsia/eclampsia (PE) was defined after 20 weeks’ gestation among individuals without chronic hypertension as any of: (1) ≥ 1 inpatient diagnosis code (ICD-9: 642.4–642.6; ICD-10: O14, O15); or (2) hypertension (≥ 2 outpatient SBP/DBP ≥ 140/90 mmHg at least 4 h apart, or severe-range SBP/DBP ≥ 160/110 mmHg confirmed within 1 h) accompanied by evidence of severe features based on diagnosis codes (e.g., proteinuria, thrombocytopenia, pulmonary edema) and/or abnormal laboratory results (e.g., creatinine >1.1 mg/dL or liver transaminases ≥2× the laboratory upper limit of normal)44.

  2. 2.

    Gestational hypertension (GHTN): GHTN was defined after 20 weeks’ gestation among individuals without chronic hypertension33 as any of: (1) ≥ 1 inpatient diagnosis code (ICD-9: 642.3, 642.9; ICD-10: O13, O16; omitting O13.1, O16.1); (2) ≥ 2 outpatient diagnosis codes; (3) ≥ 1 outpatient diagnosis plus ≥1 antihypertensive medication dispensing; or (4) ≥ 2 outpatient blood pressure readings with SBP/DBP ≥ 140/90 mmHg at least 4 h apart44. This is often a precursor to PE33.

  3. 3.

    GDM: Defined as any degree of glucose intolerance with onset or first recognition during pregnancy45. This affects around 8% of pregnancies in the USA, with a variety of possible complications, including preterm labor and developing type 2 diabetes postpartum28.

    The other four outcomes all occur during delivery or have a direct effect on the fetus. We refer to those as adverse “birth” outcomes. These are PA, preterm birth, SGA, and stillbirth. They are further defined as follows:

  4. 4.

    Preterm birth: Defined as an obstetric estimate-based gestational age at delivery <37 weeks of gestation. Incidence rate is around 8.6% of deliveries globally and accounts for approximately 17.8% of deaths in children under five years old26.

  5. 5.

    Stillbirth: Defined as the death of a fetus >20 weeks into gestation. Definitions vary across countries, but in the U.S., stillbirth occurs in approximately 0.6% of pregnancies. The majority of stillbirths occur in the uterus during gestation, but they can also happen during labor32.

  6. 6.

    PA: Defined as the separation of the placenta from the uterus prior to delivery. Incidence rate is between 0.4 and 1.0% of global deliveries with severe complications including hypovolemic shock of the mother and possible hypoxia, ischemia, acidosis, and death of the fetus29.

  7. 7.

    SGA: Defined as a birth weight <10th percentile based on the 2017 United States national reference46 for sex- and gestational age-specific singleton birth weight distributions. Additional downstream complications can include reduced growth and development and increased risk for metabolic diseases34.

Model definitions

There were a total of 49 models created across the 7 different pregnancy complications. For each complication, there were models created for seven distinct timepoints: 0, 6, 12, 18, 24, 30, and 36 weeks of gestation. At each timepoint cutoff, only data collected prior to that date was utilized. For example, in the 0-week model, only data collected prior to pregnancy would be included, but for the 36-week model, data prior to pregnancy and through 36 weeks of gestation would be used.

For the complications that arise during gestation, patients were excluded in the model if they were diagnosed at or before the cutoff date. As an example, a patient diagnosed at 28 weeks for gestational diabetes would be excluded from the 30- and 36-week models but included in the rest of the gestational diabetes models.

Data preprocessing

Last menstrual period (LMP) was back calculated using the date of delivery and gestational age at birth. Gestational age at delivery was defined using the obstetric estimate from the EHR, based on the LMP if confirmed by early ultrasound, or on early ultrasound dating if the two differed by more than 7 days, following the American College of Obstetricians and Gynecologists reVITALize Initiative47 and previously validated against early ultrasonographic examinations48. We retained all lab tests during the study period. For vital signs, we took the average per day for blood pressures and mode per day for heart rates. If there were multiple diagnoses of the same condition, the first date was retained.

We constructed input features based on EHR data availability and historically important features with a focus on regularly obtained values. The final feature set included demographic information (k = 3), lifestyle factors (k = 2), medical history (k = 35), lab tests (k = 19), vital signs (k = 3), and early pregnancy markers for second trimester and later predictions (k = 3). For the Adverse birth outcomes, any during pregnancy complications were included as a feature but were only marked as positive if they were diagnosed before the cutoff date (n = 4) (Supplementary Note 3). For example, if a patient was diagnosed with gestational diabetes at 32 weeks, gestational diabetes would be included as a positive feature for the 36-week stillbirth model but as a negative feature in the 30-week model. Early pregnancy markers were also only included in models after which that marker would have been identified. Additionally, certain diagnoses excluded patients from models. For instance, we exclude patients who have a code for pre-pregnancy or early pregnancy chronic hypertension for the GHTN models. Similarly, we exclude all patients who have a diagnosis of diabetes, for the GDM models.

We focused our study using data up to two years prior to the LMP and during the gestational period. Three buckets were created for the lab test and vital sign features: prior to LMP, during the first 24 weeks of gestation, and after 24 weeks through 36 weeks of gestation. Within each bucket, the maximum, minimum, average, and standard deviation was taken across the available feature set. Any columns that only consisted of values were dropped prior to model training. Pregnancies were excluded if they did not have any vital data and had no lab data during the modeling period. This was the only exclusion criteria based on record completeness.

For the models that had cutoffs between buckets, the bucket closest to the cutoff was adjusted. For instance, in the 30-week model, the buckets would be prior to LMP, the first 24 weeks of gestation, and after 24 weeks through 30 weeks.

Cohort selection

For each of the seven pregnancy complications or adverse birth outcomes, a pregnancy was marked as either part of the affected cohort or the control group. Each model included a selected cutoff where only features before that cutoff were selected. If a person was diagnosed before the cutoff date in a during pregnancy outcome, they were excluded from the model. For instance, if a person was diagnosed at 18 weeks with gestational diabetes, that pregnancy would be excluded from the 24-week gestational diabetes model.

Modeling

Deployment-constrained modeling pipeline

After the literature search, each top performing study was analyzed for timepoints used and cohort restrictions. Where applicable, cohort restrictions were mimicked. For GDM, this included not dropping patients who were diagnosed before 20 weeks. For preterm birth, this included excluding all multiparous patients. For all comparisons, we used the same time range as the top-performing study. This included using data up until delivery for the stillbirth model.

Data was split 80/20 train/test. XGBoost models were trained via the training set and then evaluated on the testing set. Hyperparameters were selected via fivefold cross validation in a grid search for each model with the “auc_roc” metric optimized over the training set. The grid parameters were: lambda [0.1, 1, 1.5], subsample [0.75, 1], max_depth [2, 4, 6, 8], eta [0.01, 0.05, 0.1], and gamma [0.25, 0.5]. Uncertainty was measured via 95% confidence intervals that were obtained over 500 boot-strapped intervals on the testing set. Performance was primarily measured via area under the receiver operating curve (AUC) with a secondary metric of area under the precision recall curve (PRC).

General modeling pipeline

Data for all 49 models went through the same preprocessing pipeline. Data was split 80/20 train/test. XGBoost models were trained via the training set and then evaluated on the testing set. XGBoost models were selected after preliminary tests against a set of supervised models, including Logistic Regression and Random Forest. Hyperparameters were selected via fivefold cross validation in a grid search for each model with the “auc_roc” metric optimized over the training set. The grid parameters were: lambda [0.1, 1, 1.5], subsample [0.75, 1], max_depth [2, 4, 6, 8], eta [0.01, 0.05, 0.1], and gamma [0.25, 0.5]. MinMax scaling was done using parameters from the train set. Missing values and class imbalance were naively handled with the XGBoost model.

After predictions were generated, overall testing set performance and subgroup performance was evaluated. Subgroups were generated based on age (under 25, 25–30, 30–35, 35+), race (Hispanic, Non-Hispanic White, Non-Hispanic Black, Non-Hispanic Asian, and Other), and parity (first pregnancy versus not first pregnancy). Uncertainty was measured via 95% confidence intervals that were obtained over 500 boot-strapped intervals on the testing set. Performance was primarily measured via AUC with a secondary metric of area under the PRC. Feature importances for each model were recorded. All modeling and analyses were conducted with Python.

Early prediction models

Early identification models were selected for additional analysis based on clinical relevance. For Pregnancy models, the 12-week model was selected as it would be able to generate a prediction before the completion of the first trimester. Additionally, it would be at least 12 additional weeks before any of the Pregnancy complications could be clinically diagnosed, leaving room for early intervention. For Birth models, 24 weeks was used as an intervention timepoint as it left approximately one month prior to delivery. These thresholds were also chosen based on decisions made in prior literature.

GDM baseline with NIH guidelines

Based on NIH’s published GDM risk checklist49, we built a baseline model similar to ref. 23. The following are the questions established by the checklist and how we define them as binary features.

  1. 1.

    Are you overweight or very overweight? BMI ≥ 25.0.

  2. 2.

    Are you related to anyone who has diabetes now or had diabetes in his or her lifetime? Family history of diabetes.

  3. 3.

    Are you older than 25? Prepregnancy maternal age > 25.

  4. 4.

    Are you Hispanic/Latina, African American, American Indian, Alaska Native, Asian American, or Pacific Islander? Not in the Non-hispanic White subgroup of our dataset.

  5. 5.

    In a previous pregnancy, did you have any of the following: Gestational diabetes, stillbirth or miscarriage, large baby. History of gestational diabetes, history of an abortive outcome, history of macrosomia. Any one of these being true was marked as a positive feature.

  6. 6.

    Do you have polycystic ovary syndrome (PCOS) or another health condition linked to problems with insulin? History of PCOS.

  7. 7.

    Have you ever had problems with insulin or blood sugar, such as insulin resistance, glucose intolerance, or “prediabetes”? Average pre-pregnancy fasting glucose feature as > 100.050.

  8. 8.

    Do you have high blood pressure, high cholesterol, and/or heart disease? History of chronic hypertension pre-pregnancy or hypercholesterol (heart disease was not in our feature set). Either being true was marked as a positive feature.

Initial performance was assessed as “high risk” from the checklist (2 or more items flagged) on our test set. The checklist items were then all fed in as features to our XGBoost pipeline for additional performance comparisons.

PE baseline from Society for Maternal-Fetal Medicine checklist

Based on the Society for Maternal-Fetal Medicine Checklist for PE riskt51, we built a comparator baseline model19. The following factors were labeled high risk and transformed into binary features: PE in a previous pregnancy and pre-pregnancy chronic hypertension. The following factors were labeled moderate risk and transformed into binary features: nulliparity, obesity (BMI of 30 or higher), self-reported African ancestry, and maternal age 35 years or older.

The following checklist items were excluded from the model due to lack of availability in the dataset or lack of relevance in our early 12-week model comparison: multifetal pregnancy, kidney disease, autoimmune disorder, antiphospholipid or anticardiolipin syndrome, mother or sister with preeclampsia, low socioeconomic status, patient born with low birth weight or small-for-gestational age, previous pregnancy with small-for-gestational age or other adverse outcome, interpregnancy interval more than 10 years.

Initial performance was assessed as “high risk” from the checklist (1 or more items in high risk or 2 or more items from moderate risk) on our test set. The checklist items were then all fed in as features to our XGBoost pipeline for additional performance comparisons.

External validation

XGBoost models developed with the KPNC dataset were saved and then used on an external cohort from the All of Us platform. 6940 pregnancies were identified, with 6105 having at least one lab or vital sign necessary for prediction within 36 weeks of estimated LMP date. The All of Us cohort was derived using delivery concept codes to identify pregnancy. Because the date of LMP was not available, the delivery date minus 40 weeks was utilized as a proxy for LMP. This means that the windows for certain outcomes could be shifted, i.e., early pregnancy data treated as pre-pregnancy data.

All features in KPNC were able to be derived in the All of Us platform, except family history of diabetes, smoking history, alcohol use, insurance information, initiation of prenatal care, and abortion history. Data preprocessing was done with the same pipeline in the KPNC dataset, where OMOP codes were initially used to find the corresponding feature. Validation was performed with 0% tuning (the whole All of Us cohort was used in the validation set), 20% tuning, and 50% tuning. The same features were also used to develop and test XGBoost using only All of Us data as a comparison point. Confidence intervals were derived with 500 bootstrapped samples on the validation set. The same performance metrics used for KPNC evaluation were replicated. For the dissimilarity analyses, two-sided Pearson correlation tests were used to obtain p-values and determine statistical significance.

Historical study search

To contextualize our results against the existing literature, we performed a structured search for prior studies predicting the seven pregnancy complications and adverse birth outcomes that we consider. We searched PubMed for articles published from 1 January 2020 onward, combining each outcome name and its alternate names/abbreviations (restricted to the title field) with terms for ML (“ML,” “deep learning,” “artificial intelligence”) and prediction (“prediction,” “predictive”). This identified 864 records across the seven outcomes. After removing 3 duplicate records, titles and abstracts were screened to remove irrelevant records, and the remaining reports were assessed for eligibility against four criteria: use of (i) EHR or otherwise routinely collected data, (ii) a population-based design, (iii) sufficient methodological clarity to interpret the model, and (iv) a reported performance metric. This yielded 8 eligible studies for PE, 3 for GHTN, 12 for gestational diabetes, 7 for preterm birth, 4 for stillbirth, 0 for PA, and 0 for small-for-gestational-age (see Supplementary Fig. 5). For each complication with at least one eligible study, we identified the best-performing reported model and compared it to our model at the prediction time point closest to that used in the original study, replicating the corresponding gestational window and any cohort restrictions where possible to enable a direct comparison of reported test-set AUROC (Table 2).

Results

Cohort analysis and data pipeline

The KPNC dataset consisted of 448,953 pregnancies from 332,834 individuals, and the All of Us dataset consisted of 6105 pregnancies from 5347 individuals. A full breakdown of demographics and medical history information for the cohort, prior to any exclusions, and each complication are presented in Supplementary Table 1.

We selected seven common pregnancy complications as prediction tasks, based on clinical significance and data availability. Three of these complications were pregnancy complications occurring during the gestational period: PE, GHTN, and GDM. The four adverse birth outcomes occur during delivery or have a direct effect on the fetus: PA, preterm birth, SGA, and stillbirth. Many of these outcomes are intercorrelated (Supplementary Fig. 1). Outcomes are further defined in “Methods” by a combination of ICD-9/10 codes, lab values, and/or medications (Supplementary Table 2).

Data was extracted from the KPNC EHR, including demographic information (k = 3), lifestyle factors (k = 2), medical history (k = 35), lab tests (k = 19), vital signs (k = 3), early pregnancy markers for second trimester and later predictions (k = 3), and pregnancy complications for birth outcomes (k = 4) (see Supplementary Table 3 for all features).

We define three aggregation periods based on a pregnant individual’s LMP: pre-LMP, early pregnancy, and late pregnancy. We refer to all pre-LMP features obtained from 2 years prior to LMP until LMP as “pp” (pre-pregnancy), all early pregnancy features (0–24 weeks of gestation) as “ep,” and late pregnancy features as “lp” (24–36 weeks of gestation). All features were aggregated over each period (pp, ep, and lp) using simple statistical operators, i.e., minimum (min), maximum (max), average (avg), and standard deviation (std). We note that this aggregation technique could obscure temporal trends. Throughout this text, we will denote features using first the time bucket, second the type of feature, and third the statistical aggregation. For instance, the maximum systolic blood pressure measurement taken between 0 and 24 weeks of gestation is “ep_sbp_max.” Feature aggregation and modeling pipeline are shown in Fig. 1.

Fig. 1: Data aggregation and modeling pipeline.

a Data preprocessing pipeline for labs and vital time series data. Time series data (labs and vitals) is aggregated using statistical functions over buckets based on the cutoff timepoint. The time-specific aggregates are used with other static patient information as features to train models for different pregnancy check points and complications. b Feature availability over time in each of the three buckets, aggregated by having at least one lab or vital sign available.

We trained 49 XGBoost models to cover seven possible complications at seven different prediction cutoff timepoints: 0, 6, 12, 18, 24, 30, and 36 weeks of gestation. XGBoost models were selected for their ability to capture nonlinearity and natively handle missing values52. Multitask neural network (NN) models did not improve scores (Supplementary Table 4). For any given cutoff point, only features from two years prior to the LMP up to the selected cutoff point were used in order to avoid label leakage. For instance, for the 36-week cutoff, the labs and vitals signs are used up until 36 weeks and 0 days. The outcomes have to be diagnosed at least one day after that cutoff, greater than or equal to 36w1d.

We use an 80/20 train/test split, stratified so that no individual could have a pregnancy in both the training and testing set. Following model evaluation, we further trained XGBoost52 models using similar cohort restrictions to simulate performance in prior work with deployment-constrained EHR-based models.

We included all KPNC pregnancies with at least one lab or vital sign before the cutoff point in a model; this led to 442,760 pregnancies being included in the training or testing set of at least one model. An analysis on the demographic makeup of the excluded patients can be found in Supplementary Table 11. Final models were also externally validated on a cohort of 6,105 pregnancies with sufficient data in the All of Us dataset. A summary of the inclusion and exclusion criteria for both cohorts is shown in Supplementary Fig. 2.

We used area under the receiver operating characteristic curve (AUC)53 and area under the PRC to assess performance54. Confidence intervals were calculated via 500 bootstrapped intervals over the test set55. We use SHAP values as a signed estimate of model output change for features, i.e., a higher value for a feature indicates that the model predicted probability is higher, but note that this is not a causal analysis due to the lack of global concordance56.

Model performance varies by complication and improves through gestational period

We evaluate the ability of models to predict each outcome through different intervals of pregnancy, starting with pre-pregnancy data and adding data in subsequent six-week intervals until 36 weeks of gestation. At each timepoint, models are only given data that occurred prior to condition diagnosis timing. In preterm birth the final timepoint is 35 weeks of gestation to allow for sufficient sample size as pregnancies are excluded from later timepoints if delivery has already occurred.

We find that model performance (both AUC and PRC) tends to improve as gestation progresses, though this varies by complication (Fig. 2). All pregnancy complications had an AUC of 0.8 or higher for at least one timepoint. As in prior work, we find that SGA and PA, in particular, performed poorly even close to delivery13,57. Stillbirth is also difficult to forecast, but performance improves significantly when data up through delivery is used, i.e., results with no early prediction window (Table 2). This difference is largely due to sample size drop off when excluding births that have already occurred for later timepoints.

Fig. 2: AUC and PRC values over time for each pregnancy complication.

a Plots illustrate changes in AUC scores from pre-pregnancy (0 weeks) through 36 weeks of gestation for each complication. Numbers on bars represent the number of positive test cases at each timepoint. Bars represent performance on the full test set; error bars indicate 95% confidence intervals on n = 500 bootstraps. b Plots illustrate PRC over time. Bars represent performance on the full test set; error bars indicate 95% confidence intervals on n = 500 bootstraps. c Early predictor AUC curves with 95% confidence intervals. Early predictors are at clinically relevant prediction points: 12 weeks of gestation for complications occurring during pregnancy (pre-eclampsia, gestational hypertension, gestational diabetes) and 24 weeks of gestation for complications identified at birth (preterm birth, stillbirth, placental abruption, small for gestational age). Dark lines represent performance on the full test set; shaded intervals represent performance on 95% confidence intervals on n = 500 bootstraps. d Early predictors PRC models with 95% confidence intervals plotted. Dark lines represent performance on the full test set; shaded intervals represent performance on 95% confidence intervals on n = 500 bootstraps. Data for Fig. 2 is in Supplementary Data Table 1.

Complications prediction generally improves as pregnancy data availability increases, with some sharp improvements in AUC as specific tests are collected, e.g., between 24 and 30 weeks of gestation for GDM (Fig. 3), a trend we examine in Supplementary Note 1. We note that only OGTT tests prior to censoring were used, meaning no diagnostic test was used. We also include an experiment where all OGTT values are removed and find that early detection performance minimally changes (AUC 0.800 [0.794, 0.806] versus AUC 0.782 [0.777, 0.788], see Supplementary Table 5). Importantly, some improvements in model performance align with specific gestational periods during which the predicted risk scores experience a rise for positive cases, indicating increased discriminative power at these points, while risk scores for negative cases remain consistently low (Supplementary Fig. 4).

Fig. 3: Feature Importance.

a The three SHAP values for top features in each model. Each point represents an individual pregnancy’s SHAP value for a given feature, with positive SHAP values indicating that the feature increases the predicted risk, and negative SHAP values indicating it reduces the risk. The color is the value of the feature, and the x-axis is the SHAP value. The abbreviations used are as follows: Glucose tolerance test 2 h (OGTT-2), Glucose Tolerance Test 1 h post-meal (OGTT-1h), body mass index (BMI), systolic blood pressure (sbp), diastolic blood pressure (dbp), pre-pregnancy (pp), early pregnancy (ep), and late pregnancy (lp). b The most important lab or vital in the final model from Fig. 4a and its availability (per percentage of pregnancies that have that complication) over time through 36 weeks. For instance, at the 15 week mark, around 90% of pregnancies have an early blood pressure measurement and about 80% have a pre-pregnancy blood pressure measurement.

Feature importance varies across time and complication

We use SHapley Additive exPlanations (SHAP) to assign each feature an importance value for a particular prediction model and timepoint, and plot the top three features determined by the sum of the absolute SHAP values per feature (Fig. 3)58. Although SHAP values are useful in offering the above intuitive explanations, they can exhibit instability and may lack global concordance due to correlation between features, which can complicate the interpretation of absolute feature rankings in complex clinical scenarios such as pregnancy complications56.

We find that blood pressure and parity features are important across time and complication. The importance of historical (i.e., pre-pregnancy) and current (i.e., early or late pregnancy) data also changes over time, with markers of physiological state, e.g., current blood pressure, eventually surpassing pre-pregnancy condition. For instance, all models highly weight a historical (e.g., parity) or demographic variable (e.g., race/ethnicity) at the 0 and 12 weeks marks, but vital and lab data become more important over time. Models also tended to weight the latest metrics of blood pressure higher than the earlier ones. This aligns with clinical prioritization of more recent physiological state in maternal and fetal health59,60. A full version of Fig. 3 with all feature abbreviations can be found in Supplementary Fig. 5.

Model performance and clinical utility for screening

We assess the potential utility of routine EHR models for screening (early onset prediction) using two specific timepoints: 12 weeks for pregnancy complications and 24 weeks for adverse birth outcomes (Fig. 2B). We note that 12 weeks generates a prediction within the first trimester, which is important to enable timely clinical interventions61, and 24 weeks showcases models that allow for intervention prior to delivery62. The best 12-week pregnancy models have AUCs of 0.800, 0.789, and 0.740, respectively for GDM, GHTN, and PE. We find that these pregnancy models are well-calibrated across key subgroups (Fig. 4B, C and Supplementary Fig. 7).

Fig. 4: Clinical utility of early GDM model.

a Calibration plots GDM across subgroups. b Recall versus risk percentile curves with 50 random samples of positive cases from the best and worst performing subgroups for GDM.

Following ref. 23, we evaluate whether model performances in comparison to opportunistic screening models, such as NIH’s GDM risk checklist49. We find that the checklist had an AUC of 0.581 [0.577, 0.584] at 12 weeks and a PRC of 0.098 [0.095, 0.100], which increased to 0.68 [0.673, 0.686] and 0.184 [0.176, 0.192], respectively, when using checklist items as features in our XGBoost pipeline. Following the same procedure for the Society for Maternal-Fetal Medicine’s PE checklist51, we had a naive AUC of 0.500 [0.492, 0.507] and PRC of 0.052 [0.050, 0.053], and an XGBoost AUC of 0.660 [0.652, 0.668] and PRC of 0.131 [0.123, 0.138].

These all underperform compared to our early screening GDM and PE models at all time points, underscoring the competitive performance of our models. In contrast, all 24-week birth models have lower performance, ranging between 0.622 and 0.683 for AUCs, indicating the difficulty of the screening task without other forms of data.

To assess whether early prediction translates to actionable clinical benefit, we evaluated whether 12-week models could enable cost-effective interventions during the window when prevention is still possible. For instance, the 12-week GDM model enables clinically actionable risk stratification well before standard glucose tolerance testing at 24–28 weeks63. At the top 10% screening threshold, the model identifies over 40% of women who will develop GDM with a positive predictive value of 34.1%, representing a fourfold enrichment over the baseline prevalence. Threshold-based performance metrics for all outcomes are reported in Supplementary Table 11.

Early identification of GDM risk may support a window for structured lifestyle intervention (diet and exercise counseling), which has demonstrated a 23% relative risk reduction at a cost of approximately $199 per woman64. Using a threshold-based cost-effectiveness framework65, model-based screening followed by lifestyle intervention is cost-effective at every threshold evaluated, with net savings exceeding $12,000 per case prevented even at the broadest screening threshold of top 20%, corresponding to approximately $10.8 million in total savings across the KPNC test cohort (Table 1). Additional threshold-based and intervention-based values are reported in Supplementary Table 12.

Table 1 We show net savings per case prevented for the 12-week GDM model under lifestyle counseling as a primary intervention with relative risk (RR) of 0.7718 and an incremental cost of $199 per woman

Full size table

Analogous analyses for preeclampsia and spontaneous preterm birth, as reported in Supplementary Table 12, demonstrate that aspirin prophylaxis for PE66 is similarly cost-effective at every threshold, while the two-step cervical length triage pathway for PTB does not achieve cost-effectiveness even under optimistic assumptions with a cervical length triage pathway67. Together, these findings underscore that prediction model performance alone is insufficient to ensure clinical utility. The availability, efficacy, and cost structure of downstream interventions are equally critical.

Age and race/ethnicity subgroup performances for early prediction models

We analyze our model results across race/ethnicity, age, and parity subgroups to ensure that no groups are under-diagnosed68. We show the best and worst performing subgroups for early intervention models (Fig. 5, full evaluation in Supplementary Table 6). We also include BMI groups as a potential subgroup with prior evidence of high risk69, and find they do not have large variance (Supplementary Table 7). We recognize that prevalence rates differ by subgroup70 and report rates within our cohort in Supplementary Table 8.

Fig. 5: Subgroup differences for early prediction models.

a Subgroup AUC differences for early identifier models from Fig. 2c. The best and worst performing subgroup in each model are plotted along with overall performance. The number in each label represents the number of positive test set cases in each subgroup. Dots represent performance on the full test set; error bars represent 95% confidence intervals from n = 500 bootstraps. b Calibration plots for GHTN and GDM across subgroups. c Recall versus risk percentile curves with 50 random samples of positive cases from the best and worst performing subgroups. d GHTN risk for patient subgroups over time, which show that Black patients have elevated risk in the negative cases, i.e., an increase in false positives. Data for this figure is in Supplementary Data Table 1.

We find that models perform worse in predicting PE and GHTN in NH Black patients, and GDM and SGA in Asian patients. One potential reason for this gap is data availability; key variables like early pregnancy systolic blood pressure “ep_sbp” are recorded earlier and more often in White patients than Black patients (Race/Ethnicity in Supplementary Fig. 9). Otherwise, overall recall versus risk percentile for early intervention is well-calibrated (Fig. 5b, c). We hypothesize that performance difference can partially be attributed Black patients having elevated risk in negative cases (false positives) meaning that models predominantly overestimate risk as opposed to missing patients (Fig. 5d). Models also underperform in younger patients, potentially due to older patients having important features like SBP recorded earlier and more often; this gap persists both pre-pregnancy and early pregnancy (Age in Supplementary Fig. 9).

Generally, performance differences can be attributed to a variety of complex factors, and similar feature availability can shrink these gaps (Supplementary Table 13). These differences point to addressable, structural differences in care between groups, which is consistent with prior research documenting variations in prenatal care utilization and quality across race/ethnicity and age groups71. While the performance gaps observed in our work are small, we note that our findings support prior work noting systemic gaps in prenatal care for Black patients68.

Data availability is also crucial within the parity subgroups; for instance, multiparous patients have more pre-pregnancy features, but the data gaps do not persist in pregnancy. However, despite this data asymmetry only existing pre-pregnancy, models still perform better in multiparous patients than nulliparous patients in all complications (Parity in Fig. 5a). This aligns with prior work that has found multiparous patients are easier to identify as high risk72.

Model transferability in national health sample

We investigate whether performant models trained on a large cohort of data in the US generalize to a nation-wide sample of pregnancies using the All of Us dataset. We evaluate KPNC models at 12 weeks for pregnancy models and 24 weeks for birth models using finetuning to adapt models to new data distributions73,74, via the XGBoost package52. For “No Tuning,” we apply models trained on KPNC data (without any modification) to the entire All of Us sample and report performance. We also evaluate KPNC models that are finetuned with 20 or 50% All of Us data (1125/1231 and 2799/3062 samples for pregnancy complications, adverse birth outcomes, respectively), before the model is tested on the remaining sample (test set ranges from [2795–4469] for pregnancy complications and [3043–4874] for birth). SGA was excluded from this analysis as it is not consistently coded in All of Us. We find that in-distribution models (trained and tested on All Of Us) outperform transferred models, and that fine-tuning transferred models improves performance across all settings (Fig. 6a). With no tuning, pregnancy complications in All of Us had an average of a 12.9 point performance gap as compared to the 12 week KPNC test data; this decreased to 8.8 with 20% tuning, and 8.5 with 50% tuning. This confirms work in other settings that finds even models trained with large amounts of health data must be locally tuned in order to perform well in a new sample75,76.

Fig. 6: Model Transferability of the KPNC gestational hypertension prediction model across All of Us recruitment sites.

a External validation of KPNC early models in the All of Us cohort with varying degrees of fine tuning (0, 20, 50%) as compared to a model fully retrained and on All of Us data (AoU Model). Bars represent performance on the full test set; error bars indicate 95% confidence intervals on n = 500 bootstraps. b California (KPNC training site location) is highlighted in red. The blue dashed line indicates overall mean AUC (0.641); the gray dotted line represents chance performance (0.50). States are ordered by descending AUC. Bars represent performance on the full test set; error bars indicate 95% confidence intervals on n = 500 bootstraps. c Feature completeness dissimilarity from California, measured as the mean absolute difference in per-feature missingness rates across all 147 model features (r = −0.281, p = 0.432) using a two-sided Pearson correlation test, unadjusted for multiple comparisons, n = 10 states. Point size scales with sample size; color indicates AUC. Data for this figure is in Supplementary Data Table 3.

Importantly, model performance is relatively stable across All of Us recruitment sites with substantial samples (mean AUC = 0.641). Feature missingness dissimilarity from the training site showed a weak negative association with performance (r = −0.281, p = 0.432; Fig. 6c), while the availability of the top GHTN features (blood pressure, BMI, and heart rate) showed the strongest association (r = 0.547, p = 0.102; Supplementary Fig. 9c). Neither geographic distance (r = 0.303, p = 0.395; Supplementary Fig. 6a) nor demographic dissimilarity (r = −0.002, p = 0.995; Supplementary Fig. 9a) were associated with performance variation. Performance differences across sites may also reflect differences in coding practices, data density, and prenatal care patterns in addition to model transportability issues. For instance, in Supplementary Table 9, we note differences in feature completeness, especially across blood pressure measurements.

Comparing practical models to historical studies

We conduct a thorough literature search of 864 papers to identify studies that target our selected pregnancy complications (Supplementary Fig. 6). We further filter to work that used EHR data, performed a population-based study, had methodological clarity, and defined performance metrics to obtain 8/42 studies for PE, 3/10 for GHTN, 12/88 for GDM, 7/92 for preterm birth, 4/16 for stillbirth, 0/4 for PA, and 0/22 for SGA. Many studies required either non-routine data collection (e.g., specialized biomarkers) or strict inclusion criteria that limited generalizability to routine care settings. We also found that many data-driven population-based studies did not attempt to forecast complication onset early, instead targeting diagnosis of established conditions or post-diagnosis outcome prediction.

For each complication where population-based ML studies existed, we note the best performing published ML model for comparison and match the stated time window of data collection for a direct reported test set AUCs comparison (Table 2). Some studies include cohort restrictions that are impractical deployment constraints, and we replicated these restrictions when possible to provide an apt comparison to our models. We note that two outcomes are absent from Table 2 because, to the best of our knowledge, there is no work predicting future SGA or PA using EHR data. Studies we excluded used features like fetal biometrics and admissions data or were not population based14,77.

Table 2 Comparison to historical models

Full size table

We find that XGBoost models trained with routine EHR data often perform similarly to, or better than, prior work; this is even true when prior work used specialized cohorts or data23,24,38,39,40. Our models provide a more realistic evaluation of performance compared to previously published preterm birth prediction models that are single hospital-based, include only nulliparous pregnancies rather than the general pregnant population, or include non-routine data, such as fetal and cervical abnormalities.

Discussion

Our work is an exploration of model utility in a large cohort of patients with validations across a wide range of pregnancy complications and adverse birth outcomes over the gestational period. We benchmark practical screening of these complications using only routinely available EHR data in order to facilitate generalizability. We find that model performance between complications varies, but generally improves further into gestation as shown in past work23,24. While all pregnancy outcomes had an AUC of between 0.74 and 0.80 during the first trimester, birth outcomes performed more poorly with AUCs ranging from 0.63 to 0.67. This highlights the difficulty of birth complication predictions with routine data, consistent with prior literature63,13. We hypothesize that pregnancy complications are easier prediction targets due to a combination of earlier onset, larger sample sizes, and direct ties to feature measurements such as blood pressure78. We find that the clinical utility of our early-intervention models also varies by outcome, with PE and GDM showing both strong calibration and cost-benefits of implementation, whereas preterm birth struggles to see a cost-benefit, even with strong calibration performance.

We find that blood pressure and parity are consistently important features, aligning with clinical knowledge on hypertensive disorders’ impact on pregnancy79. PE alone affects an estimated 4.6% of worldwide deliveries and causes the deaths of >70,000 women and 500,000 babies annually26,33. Our work also concords with prior findings that nulliparous pregnancy as a clinical risk factor, likely in part due to high-risk people choosing to not have a subsequent pregnancy72. Importantly, we find that non-static features such as labs and vitals became more important as patients approached delivery, indicating that data obtained later in pregnancy may be valuable for predicting future complications. Models also performed consistently worse in younger age and nulliparous patients, potentially corresponding to younger patients receiving delayed prenatal care68 and fewer data points for nulliparous women80.

Our race/ethnicity fairness analyses also identified gaps that could be driven by feature availability. We note that there are few historical studies that analyze race/ethnicity subgroup performance20. underscoring the need for fairness analyses in this space41. Among the five comparator studies we reviewed (Table 2) four incorporated race/ethnicity as a predictive variable23,38,39,40, but only one study conducted a formal fairness analysis24. Specifically, prior work on preterm birth prediction found performance gaps between Asian and Hispanic patients versus Black and White patients24. While we find that inclusion of race/ethnicity improves performance in select models (Supplementary Table 10), we emphasize that demographic and socially-constructed features may improve or harm subgroup fairness depending on the setting81,82.

Our test of model transferability in the All of Us cohort found an averaged 0.08 difference in model test AUCs even with 50% fine tuning. While this may be partially due to limited sample size, we note that prior external validation work with preterm birth has also underscored how differences in prevalence in conditions can affect transfer learning performance24. This is an important consideration for any risk score used in population screening, and should be carefully considered as models are validated.

Our study has several limitations. First, we note that our goal was to benchmark the potential of models built using only routinely-collected EHR data, and we find performance is similar to other data-driven models in this space. However, the AUC and PRC values remain low for many complications early in pregnancy. All complications had less than a 10% incidence rate across our population, with two complications having lower than a 2% rate. This rare event prediction remains a challenge in the maternal morbidity space34,35. Moreover, we limited the scope of our study to tabular data that is easily de-identified across sites, and while we included a broad range of pregnancy complications, our study does not include all pregnancy-related risks. Additionally, our external validation is limited to a US-based cohort (All of Us), and the generalizability of these models to healthcare systems and populations outside the United States remains untested. Within the All of Us cohort, gestational age is approximated as delivery date minus 40 weeks due to the absence of consistently coded LMP dates, which introduces imprecision in the feature selection windows and may attenuate external validation performance. Finally, we also cannot exclude minimal overlap between KPNC participants and the All of Us cohort given its nationwide recruitment, though California does not outperform other states in our geographic analysis (Supplementary Fig. 6a), suggesting any such overlap does not meaningfully inflate validation results.

Several directions could extend this work. Incorporating clinical notes and imaging data could enhance model performance83,84, particularly for outcomes such as SGA and PA where ultrasound measurements are known to be informative13,14,57,77,85. A targeted follow-up linking structured ultrasound measurements with routine EHR data would quantify the incremental predictive value above the baseline established here. Future work could also target downstream severe maternal morbidity criteria86, and external validation in international cohorts with confirmed gestational dating would strengthen the generalizability of these findings. More broadly, our findings indicate that early prediction models for pregnancy complications may have the potential to enhance prenatal care by enabling risk-based referrals to specialist care87,88. Our use of routinely available EHR data is particularly valuable in health systems with limited resources, where targeted interventions can have the greatest impact, and specialized additional tests might not be available89. Such systems could expand early risk screening to patients for preeclampsia, gestational diabetes and GHTN without specialized testing, if deployed in populations similar to model training data (“in-distribution”)42. The dramatic reduction in maternal mortality in Sierra Leone, nearly 75% since 2000, demonstrates that systematic, data-driven approaches to risk identification and resource allocation can lead to major improvements90. In these settings, predictive models can serve as a practical tool for health systems and public health agencies to reduce disparities and improve outcomes.

Ethics statement

The KPNC Institutional Review Board approved this project (IRB# 00001045) and waived the requirement for informed consent from participants for use of de-identified data. Use of de-identified All of Us data is approved under the program’s central Institutional Review Board and data use agreement. There was no patient or public involved during this study.

Data availability

A de-identified analytic dataset may be requested from the corresponding author with a scientifically sound proposal. Data access is restricted to requests that are approved by the study investigators and the Kaiser Foundation Research Institute Human Subjects Committee for which an appropriate data use agreement has been executed. Approved data may be used only for the purposes specified in the approved proposal and data use agreement and may not be redistributed. We anticipate responding to requests within 4–6 weeks (contact yeyi.zhu@kp.org). The All of Us dataset is available to all registered research users on their platform (https://www.researchallofus.org/). The data to replicate the cohort in All of Us is available in the manuscript’s GitHub (see “Code Availability”).

Code availability

All code for this study is available at the following GitHub link: https://github.com/MLforHealth/Pregnancy-complications-and-adverse-birth-outcomes. A synthetic sample of data was generated to enable the running of Kaiser Permanente Northern California files, and the All of Us files can be run on their platform (https://www.researchallofus.org/).

References

  1. Gunja, M. et al. Insights into the U.S. maternal mortality crisis: an international comparison. Commonwealth Fund https://doi.org/10.26099/cthn-st75 (2024).

  2. Pregnancy Mortality Surveillance System. CDC. https://www.cdc.gov/maternal-mortality/php/pregnancy-mortality-surveillance/?CDC_AAref_Val=https://www.cdc.gov/reproductivehealth/maternal-mortality/pregnancy-mortality-surveillance-system.htm (2021).

  3. Daw J. R., MacCallum-Bridges C. L., Admon L. K. Trends and disparities in maternal self-reported mental and physical health. JAMA Intern. Med. https://doi.org/10.1001/jamainternmed.2025.1260 (2025).

  4. Four in 5 pregnancy-related deaths in the U.S. are preventable. CDC. https://www.cdc.gov/media/releases/2022/p0919-pregnancy-related-deaths.html (2022).

  5. Pregnancy Complications. Cleveland Clinic. https://my.clevelandclinic.org/health/articles/24442-pregnancy-complications (2022).

  6. Fink, D. A. et al. Trends in maternal mortality and severe maternal morbidity during delivery-related hospitalizations in the United States, 2008 to 2021. JAMA Netw. Open 6, e2317641 https://doi.org/10.1001/jamanetworkopen.2023.17641 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  7. Crump, C., Sundquist, J. & Sundquist, K. Adverse pregnancy outcomes and long-term mortality in women. JAMA Intern. Med. 184, 631–640 https://doi.org/10.1001/jamainternmed.2024.0276 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  8. Takele, W. W. et al. Effective interventions in preventing gestational diabetes mellitus: a systematic review and meta-analysis. Commun. Med. 4, 75 https://doi.org/10.1038/s43856-024-00491-1 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  9. Martínez-Vizcaíno, V. et al. Exercise during pregnancy for preventing gestational diabetes mellitus and hypertensive disorders: an umbrella review of randomised controlled trials and an updated meta-analysis. BJOG 130, 264–275 https://doi.org/10.1111/1471-0528.17304 (2023).

    Article  PubMed  Google Scholar 

  10. Roberge, S., Bujold, E. & Nicolaides, K. H. Meta-analysis on the effect of aspirin use for prevention of preeclampsia on placental abruption and antepartum hemorrhage. Am. J. Obstet. Gynecol. 218, 483–489 https://doi.org/10.1016/j.ajog.2017.12.238 (2018).

    Article  CAS  PubMed  Google Scholar 

  11. Tang, A. S. et al. Harnessing EHR data for health research. Nat. Med 30, 1847–1855 https://doi.org/10.1038/s41591-024-03074-8 (2024).

    Article  CAS  PubMed  Google Scholar 

  12. van Doorn, P. et al. A. Consecutive prediction of adverse maternal outcomes of preeclampsia, using the PIERS-ML and fullPIERS models: a multicountry prospective observational study. PLoS Med. 22, e1004509 https://doi.org/10.1371/journal.pmed.1004509 (2025).

    Article  Google Scholar 

  13. Asadpour, V., Puttock, E. J., Getahun, D., Fassett, M. J. & Xie, F. Automated placental abruption identification using semantic segmentation, quantitative features, SVM, ensemble and multi-path CNN. Heliyon 9, e13577 https://doi.org/10.1016/j.heliyon.2023.e13577 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  14. Huang, K. et al. A risk factor analysis and prediction model of placental abruption. Int J. Clin. Experim. Med 12, 13023–13031 (2019).

    Google Scholar 

  15. Wu, Y. et al. Noninvasive early prediction of preeclampsia in pregnancy using retinal vascular features. npj Digit. Med. 8, 188 https://doi.org/10.1038/s41746-025-01582-6 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  16. Kovacheva, V. P. et al. Prediction of preeclampsia from clinical and genetic risk factors in early and late pregnancy using machine learning and polygenic risk scores. Hypertension 81, 264–272 https://doi.org/10.1161/HYPERTENSIONAHA.123.21053 (2024).

    Article  CAS  PubMed  Google Scholar 

  17. Antwi, E. et al. Machine-learning-based prediction of pre-eclampsia using first-trimester maternal characteristics and biomarkers. Ultrasound Obstet. Gynecol. 61, 413–421 https://doi.org/10.1002/uog.26134 (2023).

    Article  Google Scholar 

  18. Jones, M. C. et al. Maternal early warning scores shown to be methodologically weak and at high risk of bias. J. Clin. Epidemiol. https://doi.org/10.1016/j.jclinepi.2025.111833 (2025).

    Article  PubMed  Google Scholar 

  19. ACOG clinical practice update: screening for gestational and pregestational diabetes in pregnancy and postpartum. Obstet. Gynecol. 144, e20–e23 https://doi.org/10.1097/AOG.0000000000005612 (2024).

  20. American Diabetes Association Professional Practice Committee for Diabetes. 15. Management of diabetes in pregnancy: standards of care in diabetes-2026. Diab. Care 49, S321–S338 https://doi.org/10.2337/dc26-S015 (2026).

  21. U. S. Preventive Services Task Force et al. Aspirin use to prevent preeclampsia and related morbidity and mortality: US preventive services task force recommendation statement. JAMA, 326 1186–1191 https://doi.org/10.1001/jama.2021.14781 (2021).

  22. McElrath, T. F. et al. Utility of the US preventive services task force for preeclampsia risk assessment and aspirin prophylaxis. JAMA Netw. Open 8, e2521792 https://doi.org/10.1001/jamanetworkopen.2025.21792 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  23. Artzi, N. S. et al. Prediction of gestational diabetes based on nationwide electronic health records. Nat. Med. 26, 71–76 https://doi.org/10.1038/s41591-019-0724-8 (2020).

    Article  CAS  PubMed  Google Scholar 

  24. Abraham, A. et al. Dense phenotyping from electronic health records enables machine learning-based prediction of preterm birth. BMC Med. 20, 333 https://doi.org/10.1186/s12916-022-02522-x (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  25. The All of Us Research Program Investigators. The ‘all of us’ research program. N. Engl. J. Med. 381, 668–676 (2019).

  26. Dimitriadis, E. et al. Pre-eclampsia. Nat. Rev. Dis. Prim. 9, 8 https://doi.org/10.1038/s41572-023-00417-6 (2023).

    Article  PubMed  Google Scholar 

  27. Bello, N. A. et al. Prevalence of hypertension among pregnant women when using the 2017 American College of Cardiology/American Heart Association blood pressure guidelines and association with maternal and fetal outcomes. JAMA Netw. Open 4, e213808 https://doi.org/10.1001/jamanetworkopen.2021.3808 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  28. Nakshine, V. S. & Jogdand, S. D. A comprehensive review of gestational diabetes mellitus: impacts on maternal health, fetal development, childhood outcomes, and long-term treatment strategies. Cureus 15, e47500 https://doi.org/10.7759/cureus.47500 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  29. Tweddell, S. M. et al. Placental abruption and neonatal anemia. J. Perinatol. 43, 782–786 https://doi.org/10.1038/s41372-023-01603-w (2023).

    Article  CAS  PubMed  Google Scholar 

  30. Thanh, B. Y. L. et al. Mode of delivery and pregnancy outcomes in preterm birth: a secondary analysis of the WHO Global and Multi-country Surveys. Sci. Rep. 9, 15556 https://doi.org/10.1038/s41598-019-52015-w (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  31. Saenger, P., Czernichow, P., Hughes, I. & Reiter, E. O. Small for gestational age: short stature and beyond. Endocr. Rev. 28, 219–251 10.1210/er.2006-0039 (2007).

  32. Stillbirth. Cleveland Clinic. https://my.clevelandclinic.org/health/diseases/9685-stillbirth (2023).

  33. ACOG Practice Bulletin No. 202. Gestational hypertension and preeclampsia. Obstet. Gynecol. 133, 1 https://doi.org/10.1097/AOG.0000000000003018 (2019).

  34. Ward, Z. J. et al. Simulation-based estimates and projections of global, regional and country-level maternal mortality by cause, 1990–2050. Nat. Med. 29, 1253–1261 https://doi.org/10.1038/s41591-023-02310-x (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  35. Tanner, D., Lavista Ferres, J. M. & Mitchell, E. A. Improved estimation of the relationship between fetal growth and late stillbirth in the United States, 2014–15. Sci. Rep. 14, 6002 https://doi.org/10.1038/s41598-024-56572-7 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  36. Espinosa, C. et al. Data-driven modeling of pregnancy-related complications. Trends Mol. Med. 27, 762–776 https://doi.org/10.1016/j.molmed.2021.01.007 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  37. Albarqi, M. N. The impact of prenatal care on the prevention of neonatal outcomes: a systematic review and meta-analysis of global health interventions. Healthcare 13, 1076 https://doi.org/10.3390/healthcare13091076 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  38. Chen, Y. et al. Machine-learning predictive model of pregnancy-induced hypertension in the first trimester. Hypertens. Res. 46, 2135–2144 https://doi.org/10.1038/s41440-023-01298-8 (2023).

    Article  PubMed  Google Scholar 

  39. Marić, I. et al. Early prediction of preeclampsia via machine learning. Am. J. Obstet. Gynecol. MFM 2, 100100 https://doi.org/10.1016/j.ajogmf.2020.100100 (2020).

    Article  PubMed  Google Scholar 

  40. Malacova, E. et al. Stillbirth risk prediction using machine learning for a large cohort of births from Western Australia, 1980–2015. Sci. Rep. 10, 5354 https://doi.org/10.1038/s41598-020-62210-9 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  41. Osterman, M. J. K. & Martin, J. A. Timing and adequacy of prenatal care in the United States, 2016. National vital statistics reports: from the Centers for Disease Control and Prevention, National Center for Health Statistics. Natl. Vital-. Stat. Syst. 67, 1–14 (2018).

    Google Scholar 

  42. Davis, A. C., Voelkel, J. L., Remmers, C. L., Adams, J. L. & McGlynn, E. A. Comparing Kaiser Permanente members to the general population: implications for generalizability of research. Perm. J. 27, 87–98 https://doi.org/10.7812/TPP/22.172 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  43. National Institutes of Health. All of Us Adds Data from 50% More Participants in Largest Data Expansion to Date https://allofus.nih.gov/article/announcement-all-of-us-adds-data-from-50-more-participants-in-largest-data-expansion-to-date (All of Us Research Program, 2025).

  44. Zhu, Y., Wang, E. Z., Ngo, A. N., Greenberg, M. B. & Ferrara, A. Development and validation of case-ascertainment algorithms for hypertensive disorders of pregnancy using longitudinal electronic health records data. J. Clin. Epidemiol. 188, 112007 https://doi.org/10.1016/j.jclinepi.2025.112007 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  45. American Diabetes Association Professional Practice Committee. 2. Diagnosis and classification of diabetes: standards of care in diabetes-2024. Diabetes Care 47, S20–S42 10.2337/dc24-S002 (2024).

  46. Aris, I. M., Kleinman, K. P., Belfort, M. B., Kaimal, A. & Oken, E. A 2017 US reference for singleton birth weight percentiles using obstetric estimates of gestation. Pediatrics 144, e20190076 https://doi.org/10.1542/peds.2019-0076 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  47. Menard, M. K., Main, E. K. & Currigan, S. M. Executive summary of the reVITALize initiative: standardizing obstetric data definitions. Obstet. Gynecol. 124, 150–153 https://doi.org/10.1097/AOG.0000000000000344 (2014).

    Article  PubMed  Google Scholar 

  48. Hedderson, M. M., Ferrara, A. & Sacks, D. A. Gestational diabetes mellitus and lesser degrees of pregnancy hyperglycemia: association with increased risk of spontaneous preterm birth. Obstet. Gynecol. 102, 850–856 https://doi.org/10.1016/S0029-7844(03)00697-1 (2003).

    Article  PubMed  Google Scholar 

  49. US Department of Health and Human Services, National Institutes of Health & Eunice Kennedy Shriver National Institute of Child Health and Human Development. Am I at Risk for Gestational Diabetes? https://www.nichd.nih.gov/sites/default/files/publications/pubs/Documents/gestational_diabetes_2012.pdf (2012).

  50. Prediabetes. Mayo Clinic. https://www.mayoclinic.org/diseases-conditions/prediabetes/diagnosis-treatment/drc-20355284 (2023).

  51. SMFM Patient Safety and Quality Committee. Electronic address: smfm@smfm.org, Combs, C. A., & Montgomery, D. M. Society for maternal-fetal medicine special statement: checklists for preeclampsia risk-factor screening to guide recommendations for prophylactic low-dose aspirin. Am. J. Obstet. Gynecol. 223, B7–B11 https://doi.org/10.1016/j.ajog.2020.06.003 (2020).

  52. Chen, T. & Guestrin, C. XGBoost: a scalable tree boosting system. In KDD '16: Proc. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, https://doi.org/10.48550/arXiv.1603.02754 (2016).

  53. Melo, F. Area under the ROC curve. In Encyclopedia of Systems Biology (eds Dubitzky, W. et al.) https://doi.org/10.1007/978-1-4419-9863-7_209 (Springer, New York, NY, 2013).

  54. Sofaer, H. & Hoeting, J. & Jarnevich, C. The area under the precision-recall curve as a performance metric for rare binary events. Methods Ecol. Evol. https://doi.org/10.1111/2041-210X.13140 (2018).

  55. Mokhtar, S., Yusof, Z. & Sapiri, H. Confidence intervals by bootstrapping approach: a significance review. Malays. J. Fundam. Appl. Sci. 19, 30–42 10.11113/mjfas.v19n1.2660 (2023).

  56. Kumar, I. E., Venkatasubramanian, S., Scheidegger, C. & Friedler, S. A. Problems with Shapley-value-based explanations as feature importance measures. In Proceedings of the 37th International Conference on Machine Learning (ICML'20), Vol. 119, 5491–5500 (2020).

  57. Saw, S. N. et al. Machine learning improves early prediction of small-for-gestational-age births and reveals nuchal fold thickness as unexpected predictor. Obstet. Gynaecol. https://doi.org/10.1002/pd.5903 (2021).

  58. Lundberg, S. M. & Lee, S.-I. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17) 4768–4777 (Curran Associates Inc., Red Hook, NY, USA, 2017).

  59. Bohn, M. K. & Adeli, K. Physiological and metabolic adaptations in pregnancy: importance of trimester-specific reference intervals to investigate maternal health and complications. Crit. Rev. Clin. Lab Sci. 59, 76–92 https://doi.org/10.1080/10408363.2021.1978923 (2022). Epub 2021 Sep 29. PMID: 34587857.

    Article  PubMed  Google Scholar 

  60. Kepley J. M., Bates K., Mohiuddin S. S. Physiology, maternal changes. In StatPearls [Internet] https://www.ncbi.nlm.nih.gov/books/NBK539766/ (Treasure Island (FL), StatPearls Publishing, 2023).

  61. Baschat, A. A. et al. Prediction of preeclampsia utilizing the first trimester screening examination. Am. J. Obstet. Gynecol. 211, 514.e1–514.e5147 https://doi.org/10.1016/j.ajog.2014.04.018 (2014).

    Article  PubMed  Google Scholar 

  62. Romero, R. et al. Vaginal progesterone in women with an asymptomatic sonographic short cervix in the midtrimester decreases preterm delivery and neonatal morbidity: a systematic review and metaanalysis of individual patient data. Am. J. Obstet. Gynecol. 206, 124.e1–124.19 https://doi.org/10.1016/j.ajog.2011.12.003 (2012).

    Article  CAS  PubMed  Google Scholar 

  63. American College of Obstetricians and Gynecologists. Gestational diabetes mellitus. Practice Bulletin No. 190. Obstet. Gynecol. 131, e49–e64 (2018).

  64. Guo, X. Y. et al. Improving the effectiveness of lifestyle interventions for gestational diabetes prevention: a meta-analysis and meta-regression. BJOG 126, 311–320 https://doi.org/10.1111/1471-0528.15467 (2019).

    Article  PubMed  Google Scholar 

  65. Liu, V. X., Bates, D. W., Wiens, J. & Shah, N. H. The number needed to benefit: estimating the value of predictive analytics in healthcare. J. Am. Med Inf. Assoc. 26, 1655–1659 https://doi.org/10.1093/jamia/ocz088 (2019).

    Article  Google Scholar 

  66. US Preventive Services Task Force. Aspirin use to prevent preeclampsia and related morbidity and mortality: preventive medication. JAMA 326, 1186–1191 https://doi.org/10.1001/jama.2021.14781 (2021).

  67. Romero, R. et al. Vaginal progesterone decreases preterm birth ≤ 34 weeks of gestation in women with a singleton pregnancy and a short cervix: an updated meta-analysis including data from the OPPTIMUM study. Ultrasound Obstet. Gynecol. 48, 308–317 https://doi.org/10.1002/uog.15895 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  68. Howell, E. A. Reducing disparities in severe maternal morbidity and mortality. Clin. Obstet. Gynecol. 61, 387–399 https://doi.org/10.1097/GRF.0000000000000349 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

  69. Abdi, F. et al. Prepregnancy overweight and obesity and the risk of adverse pregnancy outcomes. Sci. Rep. 15, 16846 https://doi.org/10.1038/s41598-025-02113-9 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  70. Bornstein, E., Eliner, Y., Chervenak, F. A. & Grünebaum, A. Racial disparity in pregnancy risks and complications in the US: temporal changes during 2007-2018. J. Clin. Med. 9, 1414 https://doi.org/10.3390/jcm9051414 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  71. Hung, P. et al. Racial and ethnic and rural variations in the use of hybrid prenatal care in the US. JAMA Netw. Open 7, e2449243 https://doi.org/10.1001/jamanetworkopen.2024.49243 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  72. Miranda, M. L., Edwards, S. E. & Myers, E. R. Adverse birth outcomes among nulliparous vs. multiparous women. Public Health Rep. 126, 797–805 https://doi.org/10.1177/003335491112600605 (2011).

    Article  PubMed  PubMed Central  Google Scholar 

  73. Heaton, J. Ian Goodfellow, Yoshua Bengio, and Aaron Courville: Deep Learning. Genet Program Evol. Mach. 19, 305–307 https://doi.org/10.1007/s10710-017-9314-z (2018).

    Article  Google Scholar 

  74. Pan, S. J. & Yang, Q. A survey on transfer learning. In IEEE Transactions on Knowledge and Data Engineering Vol. 22, 1345–1359 https://doi.org/10.1109/TKDE.2009.191 (2010).

  75. Cai, J. et al. Impact of localized fine tuning in the performance of segmentation and classification of lung nodules from computed tomography scans using deep learning. Front. Oncol. 13, 1140635 https://doi.org/10.3389/fonc.2023.1140635 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  76. Rahman, S., Jiang, L. Y., Gabriel, S., Apinyanaphongs, E. & Chunara, R. Generalization in healthcare AI: evaluation of a clinical large language model. Preprint at https://doi.org/10.48550/arXiv.2402.10965 (2024).

  77. Yang, M., Wang, M., Zhu, Q. & Li, N. Development and validation of a risk prediction model for placental abruption in patients with preeclampsia. Placenta 164, 1–9 https://doi.org/10.1016/j.placenta.2025.03.005 (2025).

    Article  CAS  PubMed  Google Scholar 

  78. Muche, A. A. et al. Prognostic prediction models for adverse birth outcomes: a systematic review. J. Glob. Health 14, 04214 https://doi.org/10.7189/jogh.14.04214 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  79. Garovic, V. D. et al. Hypertension in pregnancy: diagnosis, blood pressure goals, and pharmacotherapy: a scientific statement from the American Heart Association. Hypertension 79, e21–e41 https://doi.org/10.1161/HYP.0000000000000208 (2022).

    Article  CAS  PubMed  Google Scholar 

  80. Barton, J. R. & Sibai, B. M. Prediction and prevention of recurrent preeclampsia. Obstet. Gynecol. 112, 359–372 https://doi.org/10.1097/AOG.0b013e3181801d56 (2008).

    Article  PubMed  Google Scholar 

  81. Siddique, S. M. et al. The impact of health care algorithms on racial and ethnic disparities: a systematic review. Ann. Intern. Med. 177, 484–496 https://doi.org/10.7326/M23-2960 (2024).

    Article  PubMed  Google Scholar 

  82. Suriyakumar, V. M., Ghassemi, M. & Ustun, B. When personalization harms performance: reconsidering the use of group attributes in prediction. In Proc. 40th International Conference on Machine Learning (ICML'23), Vol. 202, 33209–33228 (2023).

  83. Kwon, Y. J. F. et al. Combining initial radiographs and clinical variables improves deep learning prognostication in patients with COVID-19 from the emergency department. Radiol. Artif. Intell. 3, e200098 https://doi.org/10.1148/ryai.2020200098 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  84. Yang, B. & Wu, L. How to leverage multimodal EHR data for better medical predictions. In Proc. 2021 Conference on Empirical Methods in Natural Language Processing, 4029–4038 (2021).

  85. Chen, X. et al. Constructing small for gestational age prediction models: a retrospective machine learning study. Eur. J. Obstet., Gynecol., Reprod. Biol. 305, 48–55 https://doi.org/10.1016/j.ejogrb.2024.11.022 (2025).

    Article  PubMed  Google Scholar 

  86. Identifying Severe Maternal Morbidity (SMM). CDC https://www.cdc.gov/maternal-infant-health/php/severe-maternal-morbidity/icd.html (2024).

  87. Knight, M. et al. Saving Lives, Improving Mothers’ Care Rapid Report 2021: Learning from SARS-CoV-2-Related and Associated Maternal Deaths in the UK June 2020-March 2021 (National Perinatal Epidemiology Unit, University of Oxford, Oxford, 2021).

  88. Yan, A. P. et al. A roadmap to implementing machine learning in healthcare: from concept to practice. Front. Digital Health 7, 1462751 https://doi.org/10.3389/fdgth.2025.1462751 (2025).

    Article  Google Scholar 

  89. American Association of Birth Centerset et al. Obstetric Care Consensus #9: Levels of Maternal Care. Am. J. Obstet. Gynecol. 221, B19–B30 https://doi.org/10.1016/j.ajog.2019.05.046 (2019).

  90. Our World in Data. Sierra Leone has reduced maternal death rates by almost 75% since 2000. https://ourworldindata.org/data-insights/sierra-leone-has-reduced-maternal-death-rates-by-almost-75-since-2000.

Download references

Funding

This work was supported by the National Heart, Lung, and Blood Institute (R01HL157666) and the National Science Foundation Graduate Research Fellowship Program.

Author information

Authors and Affiliations

  1. Harvard-MIT Health Sciences and Technology Program, Cambridge, MA, USA

    Cassandra Parent

  2. Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, USA

    Cassandra Parent, Charikleia Moraitaki & Marzyeh Ghassemi

  3. Division of Research, Kaiser Permanente Northern California, Pleasanton, CA, USA

    Amanda Ngo & Yeyi Zhu

  4. Department of Obstetrics and Gynecology, Kaiser Permanente Northern California, Oakland, CA, USA

    Mara B. Greenberg

  5. Regional Perinatal Service Center, Kaiser Permanente Northern California, Santa Clara, CA, USA

    Mara B. Greenberg

  6. Center for Upstream Prevention of Adiposity and Diabetes Mellitus (UPSTREAM), Kaiser Permanente Northern California, Pleasanton, CA, USA

    Yeyi Zhu

  7. Department of Epidemiology and Biostatistics, University of California, San Francisco, CA, USA

    Yeyi Zhu

  8. Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, USA

    Marzyeh Ghassemi

Authors

  1. Cassandra Parent
  2. Charikleia Moraitaki
  3. Amanda Ngo
  4. Mara B. Greenberg
  5. Yeyi Zhu
  6. Marzyeh Ghassemi

Contributions

C.P., Y.Z., and M.G. contributed towards the initial study design. A.N. and Y.Z. contributed to the KPNC dataset with MBG providing clinical support. C.P. and C.M. conducted all analyses. All authors contributed to the preparation and review of the manuscript.

Corresponding authors

Correspondence to Yeyi Zhu or Marzyeh Ghassemi.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Parent, C., Moraitaki, C., Ngo, A. et al. Evaluation of machine learning-based prediction of pregnancy complications and adverse birth outcomes using routine data. Commun. Health 1, 46 (2026). https://doi.org/10.1038/s44528-026-00051-0

Download citation

  • Received: 16 June 2026

  • Accepted: 01 October 2026

  • Published: 11 November 2026

  • Version of record: 11 November 2026

  • DOI: https://doi.org/10.1038/s44528-026-00051-0

来源:Nature 健康护理主题 · nature.com