To study predictors of severe maternal morbidity (SMM) using clinical data from medical records and generalized additive models (GAMs).
This retrospective cohort study included nulliparous singleton term vertex (NTSV) births in 17 US hospitals (2016-19). Clinical data were abstracted from medical records. SMM included blood transfusion, DIC, hysterectomy, eclampsia, venous thromboembolism, and amniotic fluid embolism. Cesareans performed without labor were excluded. A multivariable GAM was trained to predict SMM from 40 demographic, pregnancy and labor characteristics. GAMs are an extension of logistic regression that allow non-linear effects of continuous variables. GAMs help identify clinical risk factors by displaying risk for each variable independently while correcting for other variables.
The cohort included 32,203 births, and the rate of SMM was 1.2% (N=392). The strongest predictors of SMM were birthweight, time from labor and delivery admission to complete dilation, maternal height, pre-eclampsia/gestational hypertension, and second stage duration. Maternal age and BMI were not strong predictors when other co-morbidities were accounted for. The association between baby weight and SMM was non-linear: SMM risk increased sharply as weight approached 4kg. The risk of SMM decreased smoothly with increasing maternal height (Figure 1). When a composite measure: birthweight / maternal height2 (the maternal baby size ratio) was substituted for height and birthweight, it became the strongest predictor in the model (Figure 2). The AUC for the final GAM model was 0.73, compared to 0.67 for a logistic regression model.
Birthweight and maternal height are key risk factors for SMM in the NTSV population. This highlights the potential to miss important risk factors when predictive models are based solely on categorical variables from administrative data. Prediction of SMM may be improved by using clinical data from medical records and non-linear statistical models that graphically represent the change in risk to provide model interpretability.
© 2020 Published by Elsevier Inc.