Computational Approaches for Predicting Preterm Birth and Newborn Outcomes

Introduction

Globally, preterm birth (PTB) is the leading cause of infant morbidity and mortality among children under the age of five¹ and its rates have not decreased significantly in the past decade (~10%).² However, the proportion of deaths caused by PTB complications has increased from 14.5% in 2000 to 17.7% in 2019.¹ Those who survive PTB are at greater risk of long-term morbidities, including brain injury,³ cognitive impairments,⁴ cardiovascular diseases, and more. Identifying PTB outcomes and early interventions are critical to mitigating these morbidities.

Traditionally, single risk factors, such as maternal age or medical history, or simple rules-based calculators incorporating these clinical variables, were used to estimate the likelihood of adverse outcomes or to identify high-risk pregnancies. One of the oldest and frequently used metrics is the APGAR score that focuses on various aspects of the neonate to determine health status.⁵ More recently developed risk scores such as the Extremely Preterm Birth Outcomes Tool developed by the Eunice Kennedy Shriver National Institute of Child Health and Development (NICHD) in the early 2000s uses both maternal and fetal measurements combined with logistic regression to generate predictive values for infant morbidity and mortality.⁶ While such parameters and simple scoring methods provide clinicians with an estimated risk for neonatal morbidity and mortality, there is growing acknowledgment that these data are relatively ineffective in predicting such outcomes.^7,8 For example, around 60–70% of PTB newborns survived without any major morbidities compared with those born extremely preterm despite being technically classified as being PTB.^9,10 This paradox highlights the potential contribution of confounding factors as well as modifying factors that result in individual heterogeneity for neonatal outcomes and the limitations of simple gestational age (GA)-based metrics.^11,12

A multimodal approach may be required to fully grasp the clinical progression of the pathogenesis of PTB due to its high complexity. With artificial intelligence (AI), large multimodal datasets can be seamlessly processed, integrated, and digested. For example, the increasing popularity and availability of electronic health records (EHRs) has created a rich source of data for AI-based predictive models. Unlike traditional risk scores that utilize very limited clinical data, EHRs are a much more complete representation of the mother’s medical history, including longitudinal information throughout the course of her pregnancy. These advantages of using EHRs have already begun to show superior predictive ability compared with traditional GA-based risk scores.^13–19 Second, recent advancements in high-throughput omics assays have decreased the cost, time, and sample size needed for generating biological datasets. These datasets range from multi-dimensional flow cytometry-based immunomics to mass spectrometry-based metabolomics.

Incorporating biological data into AI models can capture the complex dynamics involved in PTB and therefore potentially identify key biological pathways or biomarkers¹⁴ that may ultimately lead to novel therapeutic targets and approaches. Finally, social determinants of health such as race and socioeconomic status have been shown to impact adverse outcomes of pregnancy, including PTB.²⁰ For example, Black women in the United States have a 2-fold greater risk for PTB compared with White women.²¹ Being able to integrate social determinants of health into predictive models may not only improve their accuracy; but also, identify nonpharmacological strategies capable of modulating PTB morbidities.

In this review, we provide an overview of different data modalities that can be used to predict, understand, and derive a more accurate taxonomy for PTB and associated adverse outcomes. These include phenotypical, biological, and social determinant data (Table 1). We then conclude with an overview on ethical considerations of using AI models for PTB-associated morbidities. To develop more accurate and socially equitable predictive models for PTB and its associated morbidities, it is necessary to understand how to leverage AI appropriately and ethically.

Table 1.

Key papers by data type and computational methodology

Data Type	Methodology	Goal	Citations
Electronic Health Record	Use EHR data from NICU patients and classification/regression tree models	Predict probabilities of severe neonatal morbidities	Hamilton, E. F. et al. Estimating risk of severe neonatal morbidity in preterm births under 32 weeks of gestation. J Matern Fetal Neonatal Med 33, 73–80 (2020) [25].
	Use EHR data from NICU patients and 8 different classification models optimized using a nested cross-validation method	Predict occurrence of sepsis at least 4 hours before using EHR records	Masino, A. J. et al. Machine learning models for early sepsis recognition in the neonatal intensive care unit using readily available electronic health record data. PLoS One 14, e0212665 (2019). [26].
	Use EHR data of prenatal visits and elastic net/gradient boosting algorithms	Predict preeclampsia from early pregnancy information	Marić, I. et al. Early prediction of preeclampsia via machine learning. Am J Obstet Gynecol MFM 2, 100100 (2020). [20]
Biological	Use metabolomic, proteomic, and immune data to build a single stacked generalization model	Predict labor onset independent of gestational age	Stelzer, I. A. et al. Integrated trajectories of the maternal metabolome, proteome, and immunome predict labor onset. Sci Transl Med 13, eabd9898 (2021). [36]
	Use urinary metabolites with random forest algorithm	Predict gestational age using urinary metabolites in both healthy and pathological pregnancies	Contrepois, K. et al. Prediction of gestational age using urinary metabolites in term and preterm pregnancies. Sci Rep 12, 8033 (2022). [37]
	Use mass cytometry-based immune data and modified elastic net algorithm	Model immunological events during pregnancy	Aghaeepour, N. et al. An immune clock of human pregnancy. Sci Immunol 2, eaan2946 (2017). [38]
Social Determinants of Health	Use both maternal and newborn EHR data including SDOH features in a long short-term memory (LSTM) neural network	Predict 24 different neonatal outcomes	De Francesco, D. et al. Data-driven longitudinal characterization of neonatal health and morbidity. Sci Transl Med 15, eadc9854 (2023). [17]
	Use maternal plasma analytes and covariates (including SDOHs) with cross-validated gradient-boosted tree model (XGBoost)	Profile biological and epidemiological signatures of pregnancy	Espinosa, C. A. et al. Multiomic signals associated with maternal epidemiological factors contributing to preterm birth in low- and middle-income countries. Sci Adv 9, eade7692 (2023). [51]

Discussion

Use of EHRs to predict maternal and neonatal health

Clinical decision support systems for predicting maternal and neonatal health outcomes have evolved significantly over the past decade. To diagnose PTB, the oldest methods used a few parameters, such as GA and/or birthweight. Newer clinical risk scores followed, derived from phenotypical (i.e., lab measurements, medical history) and epidemiological (i.e., maternal age, race) analyses of PTB to better delineate the risk of associated morbidities, further advancing our understanding of PTB. These advances were made by using concurrent advances in AI that allowed clinicians and scientists to fully leverage the large and increasingly available phenotypic data available in EHRs to predict maternal and neonatal complications.

With the increasing availability of EHR data through efforts such as the Meaningful Use initiative,²² the field has incorporated more sophisticated risk calculators and machine-learning-based systems that leverage comprehensive phenotypic data. For instance, predictive models incorporating EHR data into regularized linear regressions have been shown to be predictive of early preeclampsia, a major cause of PTB and a maternal morbidity that has a complex clinical etiology.¹⁸ A second study used EHR data and machine learning to predict severe neonatal morbidities, including intraventricular hemorrhage (IVH) and necrotizing enterocolitis (NEC), and were demonstrated to be more accurate compared with traditional multivariable analyses.²³ Such studies show the effectiveness of leveraging the enormity of EHR data to increase the predictive accuracy for clinically multifaceted phenomena such as PTB.

Secondly, machine-learning algorithms can also aid in determining the optimal timing of treatment of PTB-associated morbidities. A machine-learning model based on EHR data was able to identify septic infants at least 4 hours prior to clinical presentation and confirmation, enabling rapid and timely treatment, such as administration of antibiotics.²⁴ Other machine-learning-based tools such as the Bhutani nomogram and the subsequent development of the interactive BiliTool application have enabled the identification of treatment thresholds for infants with hyperbilirubinemia, recommending when phototherapy or more intensive interventions should be initiated.²⁵ Predictive AI models will become increasingly accurate as pipelines to incorporate EHR data continue to evolve, providing clinicians with rapid and useful real-time recommendations.

Finally, generalizability is an important characteristic for any EHR-based AI model because these models often perform variably across different health systems. In one example, a model trained to predict PTB on a cohort of patients from Vanderbilt performed nearly as accurately on a similar-sized cohort from the University of California, San Francisco.²⁶ However, it is important to consider that this generalizability may not always hold, especially across two settings with extremely different patient populations. Therefore, other studies have begun to examine methods of fine-tuning models to fit a particular context. A study on the early prediction of COVID-19 infections showed that techniques such as transfer learning were effective at increasing the predictive power of EHR-based AI models.²⁷ However, transfer learning requires the availability of model architecture and pretrained weights from the original developing institution. But the authors also discovered that fine-tuning models using site-specific thresholds, which does not require any prior knowledge of the model, was reasonably effective at discriminating COVID-19 infections. As similar studies are now being conducted for PTB, the generalizability of EHR-based models will further improve, making them a cost-effective and highly portable tool for predicting PTB and its associated morbidities.

Despite the accuracy and advantages of using EHR-derived AI prediction models, one drawback is the lack of “interpretability” compared with traditional rules-based systems. Interpretability, the degree to which a human can understand how the model works, is important in the clinical setting where clinicians and hospital leadership need to trust a model’s outputs before deployment. In addition, interpretability is beneficial during model development and maintenance for identifying errors and/or areas for improvement. Finally, interpretability is an important first step to “explainability” for clinicians, where they must be able to explain their rationale for treatment to patients and other physicians before beginning treatment. Traditional scoring metrics incorporate small numbers of clinical risk factors with clear associations to the disease of interest, making their output scores more interpretable for clinicians. This is in contrast to EHR-based models, which sacrifice interpretability for accuracy in leveraging the full complexity of the EHR data. For example, one technique uses sub-clustering of patients by specific features and/or risk factors for post hoc analyses and reasoning. This technique was used in the cardiovascular field to predict drug-induced QT prolongation where the interpretable model revealed specific anti-arrhythmic medications associated with higher risk while antiemetics decreased risk.²⁸ While much more research is needed in this area for EHR-based AI models in PTB, achieving interpretable models will be transformational for the field.

The shift from rules-based systems to machine learning models is beneficial for the field, as machine-learning models often outperform rules-based systems that often fail to capture more subtle clinical nuances.^29,30 Despite their current imperfections, machine-learning models utilizing EHR data have shown incredible promise in providing more accurate predictions, clinically useful metrics, and cost-effective generalizability, which were not previously possible at this scale using traditional PTB-scoring methods.

Utilization of high-dimensional biological data in predictive models

In addition to EHRs, biological data are a second major source useful in deriving machine-learning models for PTB and its morbidities. Technological advancements in high-throughput multi-dimensional analytical techniques are cheaper, faster, and easier to perform on a smaller volume of sample than ever before, and have created an explosion of rich biological datasets. These include flow and mass cytometry, transcriptomics, genomics, proteomics, and metabolomics datasets. These omics data are at the single-cell resolution with some techniques capable of preserving spatial information or combining multiple omics analysis within a single cell. The distinguishing advantage of such biological data is its potential to reveal detailed molecular mechanisms of health and disease, especially in complex clinical states, such as PTB, likely driven by the interaction of multiple biological components rather than a single molecule or cell type.

A better understanding of how metabolites, for example, change throughout a healthy pregnancy may provide foundations for downstream identification of pathways and molecules that are part of the pathogeneses of various pregnancy-related diseases. One study used a combination of mass spectrometry, mass cytometry, and aptamer-based technology to measure the metabolites, immune cells, and proteins, respectively, in the blood of healthy pregnant women.³¹ Integrating these biological datasets using an AI-based approach, the authors were able to predict time to labor and precisely track changes in metabolites, immune cells, and proteins, laying the foundation for future studies to identify disruptions in this “normal” cadence or profile. Another study analyzed urinary metabolites using a mass cytometry-based technique to predict GA while also establishing a reference for steroid hormones.³² Such studies lay the groundwork for establishing a reference for multiple types of biological data during pregnancy, thereby allowing the identification of abnormal deviations from these references that may be exploited for therapeutic benefit.

Much effort has also been spent on using biological data and AI models to predict PTB-associated morbidities. In some cases, such models may additionally identify novel and/or potential disease-relevant pathways. For example, one study used cytometry by time of flight (CyTOF) data combined with a modified elastic net model to predict immune events during pregnancy.³³ Importantly, the study identified a novel IL-2-dependent signaling mechanism in T-cell subsets that maintains progesterone levels during pregnancy. A second study that also used a modified elastic net model integrated blood transcriptomics, proteomics, metabolomics and lipidomics, and vaginal microbiomics datasets to predict preeclampsia.³⁴ In addition to predicting preeclampsia, the model and subsequent downstream analyses revealed biological pathways important in the pathogenesis of preeclampsia, such as tryptophan, caffeine, and arachidonic acid metabolism. Finally, many studies have focused on analyzing the microbiome in relation to PTB. The microbiome is inherently complex since it is composed of an incredibly diverse set of bacterial species that act in concert rather than individually. This complexity makes AI an ideal tool for analyzing microbiome data. Indeed, one study that utilized nine publicly-available vaginal microbiome datasets was able to identify three metrics encompassing species diversity and bacterial community states that were particularly helpful in predicting PTB.³⁵ Increasing use of biological data in AI models will lead to new scientific discoveries that may be therapeutically beneficial. While such results will require individual follow-up studies for verification, they do provide a rich resource of new biological questions that may otherwise have gone undetected.

Despite the exciting advantages of using biological data in AI models for predicting PTB-associated morbidities, an important obstacle to overcome are the often small sizes of biological data sets. While the technological advances have certainly increased the ease of generating high-throughput multi-dimensional data, these data are still not as readily available as routine laboratory tests or use in clinical exams that are part of EHR data. The small sample sizes present challenges for traditional AI models that require large sample sizes. However, some studies have already begun to optimize models with limited numbers of sample. For example, one study used a deep-learning model and cytometry data to synthetically simulate a patient’s immune responses to various stimuli in a cell-type agnostic manner.³⁶ Their results showed that the in silico-generated cell responses were highly similar to ground-truth responses. In addition, studies like those mentioned above have already begun to successfully integrate multiple kinds of biological data (i.e., proteomics, transcriptomics, and lipodomics³⁴) for each patient sample, thereby increasing the depth of analysis per sample rather than the number of samples itself. The utilization of additional forms of biological data such as epigenetics or spatial transcriptomics, which have thus far not been extensively integrated into machine-learning models for PTB, may further increase the predictive power of AI models for PTB despite limited sample sizes. Such techniques allow for a more profound understanding of biological datasets, even when working with limited sample sizes.

As multidimensional biological data become increasingly available, AI models for predicting PTB-associated morbidities will be able to make better predictions. In addition to showing effective predictive power, these models have also created new biological avenues for investigation. Even with the limitations discussed, the use of biological data in AI models for PTB will transform the bench-to-bedside process.

Importance of considering social determinants of health in predictive models

While EHR and biological data encompass a wide range of measurable features for building machine-learning models, they do not account for the social context of the patient. Social determinants of health (SDOH) are defined as conditions in the environments of individuals that affect a wide range of health and quality of life issues.³⁷ The United States Department of Health and Human Services has largely grouped social determinants of health into 5 categories: access to healthcare, education, environment/neighborhood, economic stability, and social/community context.³⁷ Each of these categories can directly or indirectly impact biological and consequently, clinical outcomes in a patient. Unsurprisingly, one of the major drivers of PTB in addition to phenotypic and biological data are SDOH.^38,39 Thus, understanding the role of SDOH in PTB is not only necessary for a complete understanding of disease pathogenesis but also presents an opportunity for alternative “therapies” such as utilizing policies or social health initiatives to target PTB rather than pharmacological therapies. Therefore, understanding the effect of SDOH on the universally important medical conditions of prenatal and perinatal-related complications is important for early treatment and intervention.

While SDOH can theoretically have positive and negative impacts on prenatal health for all, they have historically impacted racial/ethnic minority groups and the socially vulnerable population. First, minority groups have different levels of access to medical care. Compared to 80% of White women, only 50–70% of Black, Hispanic, or Native Hawaiian/Pacific Islander women receive prenatal care during their first trimester.⁴⁰ The cause for this difference is multifactorial, including healthcare coverage, the presence of physical hospitals/care facilities in the area, and transportation to care. The number of mothers on Medicaid in non-White racial groups is nearly double that of White mothers, especially in certain states where these discrepancies in numbers have further exacerbated the quality of prenatal care.^40,41 In rural underserved areas, there is an increasing loss of hospitals and care facilities for pregnant women leading to limited opportunities for prenatal care. Transportation to facilities with quality care adds to the burden of an already vulnerable population. Second, disparities in education also present implications for prenatal health. Women with lower education have been found to possess lower health literacy and knowledge of birth options throughout the various gestational stages.⁴² These women are also prone to experience psychosocial stressors, such as lack of social support, support of the father, general anxiety, and lower happiness; all of which have been shown to have deleterious effects on prenatal health and increase the risk of PTB.⁴³ Third, economic instability exerts a strong influence on prenatal health. Unsurprisingly, moms from lower socioeconomic classes are less able to obtain adequate prenatal care. But interestingly, the collective socioeconomic status of a neighborhood has been shown to influence prenatal health.³⁹ All of these SDOH factors cause delays in obtaining prenatal care or the inability to receive adequate care, resulting in greater rates of pregnancy-related mortality and complications in minority groups compared to Whites.³⁸ SDOH have also been found to influence disparities in perinatal care. The use of quantitative measures such as the Dhabhar Quick-Assessment Questionnaire for Stress and Psychosocial Factors has found that economic stress was a significant factor in assessing the risk of PTB.⁴³ A machine-learning-based interdependency analysis followed by generation of a non-linear support vector machine predictive model was conducted to find interdependencies between various features, including SDOH and PTB. The study revealed that a variety of deleterious factors associated with economic stability, including general anxiety, perceived risk of birth complications, self-rated health, and divorce play a role in the risk of PTB.⁴³ Economic instability is also tied to an increased likelihood of obesity, as well as the presence of food deserts, poor nutrition, and high blood pressure, all of which increase the risk of complications during pregnancy. These factors help index and control for other factors not addressed within the questionnaire and demonstrate that self-awareness of maternal health can itself be a strong predictor of risk assessment for PTB. The social environment also has been shown to buffer against the effects of chronic stress. Women with higher levels of support from parents, siblings, and the father show decreased risk for PTB.⁴³ Effectively assessing prematurity, PTB, and complications surrounding pregnancy will require a data-driven understanding of a patient’s phenotype, social determinants of health, and a litany of biopsychosocial factors.

While not as commonly studied as other clinical or biological characteristics as predictive features of PTB, recent studies have begun incorporating SDOH into their models. For example, one study found that specific SDOH, such as homelessness and incarceration, have distinct effects on PTB even though they may share many superficial characteristics (i.e., low socioeconomic status, stress).¹⁵ Another verified that many of the associations between SDOH and PTB that we observe in the United States are also similarly associated with PTB in low- or middle-income countries.⁴⁴ Such studies demonstrate that incorporating SDOH into AI models is not only necessary to generate accurate models of PTB, but may additionally provide new insights into the complicated connections between SDOH and PTB which were not previously evident.

Ethical model development and deployment

The increasing adoption of AI approaches in healthcare has highlighted the need for a rigorous examination of how these technologies can be developed and implemented in an ethical fashion. Broadly, concerns surrounding AI include algorithmic bias,⁴⁵ prospects of unethical data collection and usage,⁴⁶ lack of transparency or interpretability and its impact on patient safety,⁴⁷ and possibilities of exploitation.⁴⁸ International organizations such as the World Health Organization (WHO) have put forward recommendations for unified guidelines across fields, yet these have not been widely adopted. Critically, patient acceptance of the use of AI in their care is reliant on the existence of such protections,⁴⁹ emphasizing the need for these ethical frameworks.

Without a rigorous ethical framework to guide the development and deployment of AI methods in pregnancy and neonatology, these approaches can also exacerbate pre-existing health disparities.⁵⁰ Algorithmic bias, arising from inadequate training data and model implementation, can reinforce racial inequities which already exist in pregnancy outcomes.⁵¹ This can be further exacerbated by inadequate data collection with poor representation across different patient populations.⁵² The result is an AI model that solidifies, rather than mitigates, racial inequalities in healthcare. Finally, the “black box” nature of AI, characterized by lack of interpretability, hinders transparency and can compromise informed decision-making in healthcare settings for both the clinician and patient.^53,54 The lack of interpretability may particularly be harmful in marginalized communities that are already struggling with mistrust of the healthcare system. To avoid harmful clinical decisions made from flawed AI output, thorough bias mitigation strategies are essential.

Comprehensive assessments at multiple stages in the development and deployment process of AI methods in healthcare are necessary to minimize the possibility of harm. Firstly, it is essential that final clinical decisions be made by humans properly trained to interpret results with caution. AI models’ potential for bias must be characterized and communicated to medical providers before deployment followed by continuous monitoring post-deployment. AI models must be created by diverse groups of developers to ensure thorough inspection and mitigation of potential bias.⁵⁵ Furthermore, transparency from developers about datasets and algorithms⁵⁵ used during model development should be required, so that relevant regulatory agencies can perform effective audits to evaluate the AI models for adequate representation of patient populations and communities. Ensuring that AI models are built and implemented ethically is critical to improve maternal and neonatal care and reduce outcome disparities in marginalized populations.

When these considerations are sufficiently implemented into the research and implementation process, AI models have the potential to reduce disparities and improve health equity in pregnancy and neonatology.⁵⁶ In research settings, AI has been used effectively to predict adverse pregnancy outcomes and neonatal complications across racially and socioeconomically diverse cohorts.^57,58 Furthermore, the deployment of AI methods to understand and address the needs of under-resourced communities has shown early promise.^44,59 These examples demonstrate the beneficial potential of AI models that can extend beyond scientific or therapeutic advancement.

Conclusions and Summary

Prematurity, currently defined as birth before 37 weeks of gestation, encompasses a heterogenous population of infants. The exclusive reliance on GA and birth weight as the principal measures of prematurity fails to account for neonates who, despite being born well before the arbitrary 37-week demarcation, do not manifest any adverse health complications. Here, we have presented studies that have utilized a combination of EHR, biological, and SDOH data to generate more holistic and accurate models of PTB. As more data are incorporated from newer sources such as wearable health technologies, diet tracking, and parsed clinical notes, AIdriven approaches will continue to open new avenues not only for discovering new therapeutic targets or pathways; but also, for continuous risk assessment and real-time risk recalibration, providing clinicians with dynamic insights for timely interventions.⁶⁰ Despite the technological and organizational hurdles that still exist, machine-learning models will no doubt provide valuable clinical and biological information to physicians and scientists alike for the understanding of PTB and its associated morbidities.

Key Points:

• Use of electronic health records (EHRs) allow accurate predictive modeling of preterm birth (PTB)- associated morbidities, a clinically complex phenotype.

• Application of artificial intelligence (AI)-based models using biological data generates new knowledge regarding the pathogenesis of PTB.

• Social determinants of health are an important component of predictive models for PTB.

• Successful deployment of predictive models for PTB will require careful consideration of ethical issues including bias, generalizability, and interpretability.

Synopsis:

Preterm birth (PTB) and its associated morbidities are a leading cause of infant mortality and morbidity. Accurate predictive models and a better biological understanding of PTB-associated morbidities are critical in reducing their adverse effects. Increasing availability of multimodal high-dimensional datasets with concurrent advances in artificial intelligence (AI) have created a rich opportunity to gain novel insights into PTB, a clinically complex and multifactorial disease. Here, we review the use of AI to analyze three modes of data: electronic health records (EHRs), biological omics, and social determinants of health metrics. We conclude with important ethical considerations for development and deployment of AI-based models for PTB.

Best Practices.

What is the current practice for predicting preterm birth and associated morbidities?

Best Practice/Guideline/Care Path Objective(s):

• Improve accuracy of prediction models using multimodal data

• Generate new biological and clinical knowledge using multimodal data

• Maintain interpretability of predictive models

• Be conscious of potential biases in predictive model development and deployment

What changes in current practice are likely to improve outcomes?

• Integrate social determinants of health into predictive models

• Use strategies for bias mitigation and increase interpretability

Bibliographic Source(s):

He J, Baxter SL, Xu J, Xu J, Zhou X, Zhang K. The practical implementation of artificial intelligence technologies in medicine. Nat Med 2019;25(1):30–6.

Marić I, Contrepois K, Moufarrej MN, et al. Early prediction and longitudinal modeling of preeclampsia from multiomics. Patterns (NY) 2022;3(12):100655.

Becker M, Mayo JA, Phogat NK, et al. Deleterious and protective psychosocial and stress-related factors predict risk of spontaneous preterm birth. Am J Perinatol 2023;40(1):74–88.

Thomasian NM, Eickhoff C, Adashi EY. Advancing health equity with artificial intelligence. J Public Health Policy 2021;42(4):602–11.