External validation of a machine learning model for delivery mode prediction after induction

No machine learning (ML) models for predicting delivery mode after labor induction (IOL) have been externally validated. We aimed to develop and validate one using medical records. Portuguese tertiary center data (n = 2434) were used for development and internal validation, and Consortium on Safe Labor data (n = 10,591) for external validation. Outcomes are vaginal delivery (VD) or cesarean section (CS). Internal validation employed different ML approaches, aiming for model simplification. Logistic regression performed best on internal validation: AUROC:0.793; F1-score:0.748; PPV:0.752, with good calibration and decision curve analysis (DCA), being selected for simplification. Simplified top-13 features model was selected for external validation: AUROC:0.808; F1-score:0.781; PPV:0.822, tending for VD (99.6%) while avoiding false-positives (0.5%). Calibration curves underestimated CS risk by 10-75%; DCA showed good net benefit. The model's good AUROC and DCA suggest clinical utility. Calibration curve underestimation of CS risk may result from outcome imbalance between datasets.