Abstract
Purpose
This study aims to train a novel explainable machine learning method (QLattice) to predict successful vaginal birth after cesarean and compare the performance of these models with other known machine learning- and logistic regression models.
Methods
A Danish cohort study including 11 017 women with a prior cesarean giving birth during year 2004–2016 was used to train and evaluate three machine learning algorithms (LASSO, Random Forest, and QLattice). Grobmans logistic regression model was used as baseline. Two models were developed (antenatal and prelabor).
Results
Overall, 4 897 (44,4%) women had a trial of labor, with 3 441 (70.3%) women delivering vaginally. In the antenatal model predictive variables were epidural (OR = 0.53), breech presentation (OR = 1.70), mothers height (OR = 1.03), pre-pregnancy BMI (OR = 0.95), any vaginal birth (OR = 7.74), and vaginal birth before cesarean (0.24). In the prelabor model induction of labor (OR = 0.59), primary rupture of membranes (OR = 0.52), and infant weight (OR = 0.52) were additional predictors. For the antenatal prediction models, the area under the curve (AUC) for the different methods were QLattice 0.73 (0.70–0.76), LASSO with 62 features 0.75 (0.72–0.78), Random Forest 0.74 (0.71–0.77), and Grobman 0.68 (0.65–0.71). For the Prelabor model, AUC were for QLattice 0.77 (0.74–0.80), LASSO with 67 features 0.77 (0.74–0.80), Random Forest 0.75 (0.72–0.78), and Grobman 0.70 (0.66–0.73).
Conclusions
QLattice performs equal to other machine learning algorithms contributing further by giving explainable insight with fewer variables. Future prospective studies will reveal whether individual decision support tools, based on these models can improve outcomes in women with previous cesarean.