Dynamic risk prediction and phenotype discovery in preeclampsia and gestational hypertension

The National Institute of Child Health and Human Development awarded Emory University $426,948 on September 1, 2026, under the Child Health and Human Development Extramural Research program (CFDA 93.865) to develop and validate artificial intelligence models that integrate electronic health record data to identify clinically meaningful subtypes of hypertensive disorders of pregnancy and predict disease onset before clinical diagnosis. The project will model longitudinal physiological changes and temporal clinical documentation during pregnancy to reveal early deviations and subtype-specific patterns predictive of preeclampsia and gestational hypertension development, with the long-term goal of enabling early, individualized risk prediction and management through scalable, data-driven tools implementable within existing clinical systems. Work is performed in Atlanta, Georgia, with a period of performance from September 1, 2026, through August 31, 2028. This is a Project Grant assistance type.

Abstract

Hypertensive disorders of pregnancy (HDP) are among the leading causes of maternal morbidity and mortality in the united states, and their incidence continues to rise. Current screening methods based on baseline risk factors or static models fail to capture the dynamic physiological and clinical changes that precede disease onset. The inability to detect early signs of preeclampsia or gestational hypertension contributes to delayed diagnosis and preventable complications. Moreover, current clinical diagnostic categories for HDP do not capture the underlying heterogeneity of the disease. These categories group together conditions with distinct etiologic mechanisms and trajectories, limiting the ability to identify early, mechanism-specific risk patterns. By uncovering data-driven subtypes that align with these underlying mechanisms, this study will provide new insight into the pathophysiology of HDP and inform tailored surveillance and treatment approaches. The long- term goal of this research is to enable early, individualized risk prediction and management of HDP through scalable, data-driven tools that can be implemented within existing clinical systems. The overall objective of this proposal is to develop and validate artificial intelligence (AI) models that integrate structured and unstructured electronic health record (EHR) data to identify clinically meaningful HDP subtypes and predict disease onset before clinical diagnosis. Our central hypothesis is that modeling longitudinal physiological changes and temporal clinical documentation during pregnancy will reveal early deviations and subtype- specific patterns predictive of HDP development. In aim 1, we will identify and characterize HDP subtypes by applying unsupervised methods to gestational age-aligned EHR representations that include physiological measures widely available in ehrs and different clinical settings (e.g., blood pressure, weight gain). By integration of a large language model (LLM)-derived note embeddings capturing symptoms, clinician impressions, and management actions, we will validate and interpret the subgroups. In aim 2, we will develop and evaluate early-pregnancy multimodal prediction models using temporal-domain analysis and fusion of structured and unstructured data to detect subtle physiological and clinical shifts preceding onset, then perform knowledge distillation to create structured-only models suitable for deployment and validation in the national epic cosmos dataset. Successful completion of this project will yield a validated, interpretable, and scalable risk prediction framework capable of detecting HDP and its subtypes before clinical onset. In addition, IT will produce a robust, reusable framework for processing and integrating clinical notes and structured data from perinatal ehrs. This proposal is innovative because IT integrates longitudinal temporal modeling with LLM- based extraction of clinical insights to bridge structured and unstructured EHR data for HDP subtype discovery and risk assessment. The results will establish a new paradigm for dynamic, data-driven maternal risk stratification, directly supporting NIH priorities to reduce maternal morbidity and mortality.