Fig. 6

Preterm birth prediction accuracy is influenced by clinical context. A Preterm birth prediction models trained and evaluated only on cesarean section (C-section) deliveries perform better (PR-AUC = 0.47) than those trained only on vaginal delivery (PR-AUC = 0.23). ROC-AUC patterns were similar (Additional file 1: Fig. S8). Billing codes (ICD-9 and CPT) present before 28 weeks of gestation were used to train a model to distinguish preterm from non-preterm birth for either C-sections (n = 5475) or vaginal deliveries (n=15,487). B Recurrent preterm birth can be accurately predicted from billing codes. We trained the models to predict preterm birth for the second delivery in a cohort of 1416 high-risk women with a prior preterm birth documented in their EHR. Three models were trained using data available 10 days, 30 days, and 60 days before the date of the second delivery. Models accurately predict the birth type in this cohort of women with a history of preterm birth (PR-AUC ≥ 0.75). ROC-AUC varied from 0.82 at 10 days to 0.77 at 60 days before the second delivery (Additional file 1: Fig. S9). Expected performance by chance is the preterm birth prevalence in each cohort (dashed lines)