Using Multimodal Clinical and SDOH Data to Develop Risk Models for Predicting Severe Maternal Morbidity - PROJECT ABSTRACT Severe maternal morbidity (SMM) is defined as an unexpected adverse outcome during pregnancy or postpartum (PP) with significant short- or long-term negative consequences to a woman’s health. Rates of SMM have risen substantially in the past decades and are >50% higher for non-Hispanic Black women than non-Hispanic White women. Leading causes of SMM and associated deaths are preventable and result, in part, from a systems-level failure to recognize and manage the co-occurring physical, mental, and social risk factors that women experience across the perinatal, intrapartum, and PP periods. Adequate prenatal care and appropriate management of chronic conditions during pregnancy may mitigate rising maternal risk. However, important knowledge gaps limit efforts to improve and optimize care. First, clinicians and health systems are limited in predicting patients who will be most likely to have complications such as SMM. Second, social determinants of health (SDOH), defined as, conditions in the places where people live, learn, work, and play that affect a wide range of health and quality-of life-risks and outcomes, may be an important determinant of outcome disparities and inadequate prenatal care. However, SDOH may be sub-optimally captured by using routine screening questions during patient encounters and population-level exposures. We propose addressing these critical knowledge gaps by leveraging highly granular, longitudinal clinical data linked with SDOH data derived from the Patient-Centered Outcomes Research Institute (PCORI)-funded INSIGHT Clinical Research Network (Co-PI Pathak) on >365,000 women across New York City (NYC). In Aim 1, we will link longitudinal electronic health records (EHRs) and claims data and characterize the study cohort. We will also ascertain community- and individual-level measures of SDOH by linking population-level SDOH data for the study cohort using natural language processing (NLP) and machine learning (ML) methods from clinical note narratives. In Aim 2, we will predict the risk of SMM, as well as assess fairness and bias through ML models. Our rich multi- level data, measurement, and analytical approach will apply ML models to analyze EHRs and community/neighborhood- and individual-level SDOH data, and ascertain key drivers and predictors of SMM risk. In Aim 3, we will collaborate with the recently established NICHD U54 center to study the clinical interpretability and actionability of ML-based risk models for SMM. Findings from this aim will provide actionable recommendations in future risk score implementation efforts using point-of-care clinical decision support tools.
