Predicting severe maternal morbidity at admission for delivery using intelligible machine learning

To develop an externally validated Explainable Boosting Machine (EBM) to predict severe maternal morbidity (SMM) at the time of hospital admission for delivery. This retrospective study used clinical data for 155,935 births (24-43 weeks) at 20 U.S. hospitals (2016 - 2021). An Explainable Boosting Machine (EBM) was trained to predict SMM (hysterectomy, blood transfusion, DIC, amniotic fluid embolism, thromboembolism, or eclampsia) using 50 features known at admission (including demographic characteristics, obstetric and other medical factors). The model was trained using data from 106,903 births at 14 hospitals and externally validated with data from 49,032 births at 6 different hospitals. A logistic regression (LR) model was trained on the same data for comparison. SMM occurred in 2262 cases (1.5%) (Table 1). The features that were most predictive of SMM included preeclampsia/gestational hypertension, economic distress quintile (based on the area where the patient lived), self-reported race, age, height, nulliparity, labor type (induction versus spontaneous labor) and initial cervical exam on admission. Risk for SMM was lowest in pregnant people aged 21-37, and rose significantly for those < 20 years and >37 years old. SMM risk decreased as height increased and with increasing cervical dilation on the initial exam in hospital. External validation gave an AUC of 0.70 (CI 0.68 – 0.71) for EBMs, and 0.69 (CI 0.67 – 0.71) for LR. The ROC plot (Figure 1) identified 59% of births with SMM for a 25% screen positive rate. Using only the top 6 features, EBMs yielded an AUC of 0.66 (95% CI 0.65 – 0.67). Our study showed that EBMs can provide clinically useful classification of SMM risk at the time of admission for delivery and highlighted hypertensive disorders of pregnancy and sociodemographic factors that result in societal disadvantage, as important risk factors for SMM. EBM interpretability also gave insights into risk factors for SMM not traditionally considered such as maternal height.View Large Image Figure ViewerDownload Hi-res image Download (PPT)