Ethical, multimodal AI tool for prediction of preeclampsia risk in early pregnancy

Project Summary/Abstract We aim to develop an ethically-centered, data-driven multimodal AI tool to predict preeclampsia risk in early pregnancy. As preeclampsia is a major contributor to the rising maternal and fetal morbidity and mortality, and there are significant racial disparities in care, our approach can revolutionize pregnancy care. We will use data from the electronic health records, encompassing structured (demographics, comorbidities, vital signs, labs, and medications) data, ultrasound images, and unstructured text (physician notes) of 202,485 individuals from Mass General Brigham. We will add polygenic risk scores for a subset of those individuals, which are available through a biobank linked to the medical records. We propose a novel, time-aware embedding to develop our multimodal AI tool for decision support in early pregnancy. Subsequently, we will utilize transparent techniques and bias mitigation to ensure equitable care. We will rigorously evaluate the performance of the novel model in external datasets from NuMom2b, AllOfUs, BCC- Preg, and University of Washington cohorts, consisting of more than 50,000 patients. We will work with ethics and human factors research experts throughout the model development stages and implement an ethics-informed framework to ensure we consider all relevant stakeholders’ needs. We have assembled an expert, cross-disciplinary team with all available state-of-the-art resources within the Harvard Medical School, Boston hospitals, and external collaborators, making us ideally suited to carry this work forward. Ultimately, we aim to validate and implement these tools in clinical practice, leading to a greatly enhanced ability to identify high-risk patients early and improve patient safety.