(1036) The Impact of Social Vulnerability Data while Utilizing Machine Learning to Predict Preeclampsia

Objective: We aimed to develop a machine learning model that utilizes social vulnerability data along with clinical information to predict the risk of preeclampsia. Additionally, we sought to analyze the impact social vulnerability data has in determining risk.

Study Design: A multi-center analysis of patients who received prenatal care at Northwell Health from January 2018 to August 2023. Medical and social vulnerability data collected at the initial prenatal visit were used for the prediction models. Social vulnerability data included both individual demographics and neighborhood-level factors based on census tracts and social vulnerability index (SVI) published by the CDC in 2022. All data was available at the initial prenatal care visit. The algorithms to predict preeclampsia were trained using a gradient boosted decision tree model.

Results: The dataset comprised of 106,626 patients, of whom 7.3% were diagnosed with preeclampsia during their pregnancy. The algorithm was fixed with a 10% false positive rate to enable comparison of predictive ability (Table 1). This predictive algorithm both with and without social vulnerability data (AUROC 0.8 vs. 0.82) exceeds the performance of current clinical guidelines (AUROC 0.68) as well as previously published models (the Mount Sinai – Sema4 consortium (AUROC = 0.70) and the Karolinska – Uppsala joint study (AUROC = 0.69)). Including social vulnerability data improves the specificity of the model by six percentage points leading to 74 fewer false positives in the test set of 1,281 patients. This difference is statistically significant with a p-value of 0.0017.

Conclusion: By using machine learning models that employ both clinical and social vulnerability data, preeclampsia can be predicted with greater accuracy than by using clinical information alone. These models could be applied to personalize risk assessment following a routine first prenatal office visit, without additional workup. Utilizing this model could allow for the targeting of high-risk patients with appropriate antepartum monitoring and counseling.