Using machine learning to predict the risk of developing hypertensive disorders of pregnancy using a contemporary nulliparous cohort

• Published in: AJOG global reports Vol. 4; no. 4; p. 100386
• Main Authors: Schor, Jonathan S., Kadambi, Adesh, Fulcher, Isabel, Venkatesh, Kartik K., Clapp, Mark A., Ebrahim, Senan, Ebrahim, Ali, Wen, Timothy
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.11.2024; Elsevier
• Subjects: Hypertensive disorders of pregnancy; Machine learning; Obstetrics and Gynecology; Original Research; Risk prediction; Risk prediction; Hypertensive disorders of pregnancy; Machine learning; machine learning; risk prediction
• ISSN: 2666-5778; 2666-5778
• DOI: 10.1016/j.xagr.2024.100386

Hypertensive disorders of pregnancy (HDP) are significant drivers of maternal and neonatal morbidity and mortality. Current management strategies include early identification and initiation of risk mitigating interventions facilitated by a rules-based checklist. Advanced analytic techniques, such as machine learning, can potentially offer improved and refined predictive capabilities. To develop and internally validate a machine learning prediction model for hypertensive disorders of pregnancy (HDP) when initiating prenatal care. We developed a prediction model using data from the prospective multisite cohort Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be (nuMoM2b) among low-risk individuals without a prior history of aspirin utilization for preeclampsia prevention. The primary outcome was the development of HDP. Random forest modeling was utilized to develop predictive models. Recursive feature elimination (RFE) was employed to create a reduced model for each outcome. Area under the curve (AUC), 95% confidence intervals (CI), and calibration curves were utilized to assess discrimination and accuracy. Sensitivity analyses were conducted to compare the sensitivity and specificity of the reduced model compared to existing risk factor-based algorithms. Of 9,124 assessed low risk nulliparous individuals, 21% (n=1,927) developed HDP. The prediction model for HDP had satisfactory discrimination with an AUC of 0.73 (95% CI: 0.70, 0.75). After RFE, a parsimonious reduced model with 30 features was created with an AUC of 0.71 (95% CI: 0.68, 0.74). Variables included in the model after RFE included body mass index at the first study visit, pre-pregnancy weight, first trimester complete blood count results, and maximum systolic blood pressure at the first visit. Calibration curves for all models revealed relatively stable agreement between predicted and observed probabilities. Sensitivity analysis noted superior sensitivity (AUC 0.80 vs 0.65) and specificity (0.65 vs 0.53) of the model compared to traditional risk factor-based algorithms. In cohort of low-risk nulliparous pregnant individuals, a prediction model may accurately predict HDP diagnosis at the time of initiating prenatal care and aid employment of close interval monitoring and prophylactic measures earlier in pregnancy.