Abstract Body:
Introduction. Early identification of pregnant women at risk of hypertensive disorders of pregnancy (HDP) is critical for improving pregnancy outcomes. The USPSTF risk stratification approach (hereafter called USPSTF checklist), typically used to assess preeclampsia risk for prophylactic aspirin use, does not account for nuanced relationships between risk factors and does not account for how a patient’s clinical profile evolves during pregnancy. Our objectives were to determine: 1) if a machine learning model to predict HDP could outperform the current checklist used in clinical practice, and 2) whether incorporating longitudinal blood pressure (BP) measurements further enhances predictive performance.

Hypothesis. We hypothesize that 1) a baseline prediction model using machine learning will provide improved identification of HDP risk compared with the USPSTF checklist, and 2) a dynamic prediction model that additionally incorporates longitudinal BP measures after the first prenatal visit will further boost model performance.

Methods. Pregnant individuals prospectively enrolled in a digital health platform from 2022-2024 were included. Using electronic health records and platform-collected data, we trained a baseline prediction model for HDP based on the USPSTF checklist risk factors using a machine learning approach (ridge logistic regression). We then incorporated longitudinal BP values taken during pregnancy at-home with a remote patient monitoring device and in-clinic. BP features included in the model were computed as a set of 13 summary statistics characterizing centrality, spread, and trends. Performance of both models was assessed using area under the curve (AUC).

Results. Among 568 assessed women, the frequency of HDP was 18.1%. The AUC for the baseline model was 0.75 (95% CI: 0.68, 0.82) while the AUC for the dynamic model was 0.88 (95% CI: 0.83,0.93). The USPSTF checklist had a specificity of 0.62 and a sensitivity of 0.58. In contrast at this specificity level, the baseline and dynamic models had a sensitivity of 0.76 and 0.92, respectively.

Conclusions. Both prediction models using machine learning and dynamic longitudinal BP data outperformed the current standard USPSTF checklist for identification of patients at risk of HDP. Dynamic incorporation of longitudinal BP measures after the first prenatal care visit improves the ability to identify patients at risk of HDP compared to a model that only uses baseline inputs.