Development and internal validation of an intrapartum caesarean risk prediction model to guide rural obstetric transfer decisions: a population-based study using BORN Ontario data.

Abstract

To develop and internally validate a clinical prediction model estimating the probability of emergency caesarean delivery among low-risk pregnancies using routinely available intrapartum variables. The model is intended to support clinicians in recommending transfer decisions in rural obstetric settings without onsite caesarean capacity. Retrospective cohort study with multivariable logistic regression and internal validation using bootstrap resampling. Province-wide birth registry in Ontario, Canada. Singleton, hospital births from low-risk pregnancies between 1 April 2012 and 31 March 2020, captured in the Better Outcomes Registry and Network Ontario. Exclusions included previous caesarean, major comorbidities and other high-risk conditions. Emergency caesarean delivery during labour. Among 611 644 low-risk pregnancies, 66 482 (10.9%) resulted in emergency caesarean delivery. Key predictors included abnormal fetal health surveillance, nulliparity, gestational hypertension, polyhydramnios area under the receiver operating characteristic curve (AUC) and oxytocin augmentation. The optimism-corrected AUC was 0.86, indicating strong discrimination. Calibration was generally good, though the model underestimated caesarean risk in patients first admitted to level I hospitals. This study presents a proof-of-concept prototype with strong discrimination, though calibration of the full model was suboptimal in level I hospitals, the intended setting for application. While not clinically deployable in its current form, the model lays the groundwork for a translational pathway that will require local recalibration, external validation and contextual adaptation before implementation. Threshold-based clinical utility (eg, positive/negative predictive value or net benefit) was not assessed and will be evaluated during external validation and local recalibration.