Risk stratification and prediction of emergency delivery in early-onset preeclampsia using machine learning

Figures

• 

  Fig. 1. Discrimination evaluation of the predictive models. (A-B) Training and testing sets performance metrics. Heatmap displayed sensitivity, specificity, accuracy, precision, F1-score, AUROC, and AUPRC values for eight machine learning models. (C-D) Training and testing sets ROC curves. ROC curves showed true positive rate (sensitivity) versus false positive rate (1-specificity) for all models, with the diagonal representing random chance. Curves closer to the top-left corner indicate better discrimination

• 

  Fig. 2. Calibration curves of predictive models. Calibration plots comparing predicted probabilities (x-axis) against observed frequencies (y-axis) for eight machine learning models. The diagonal dashed line represents perfect calibration. Each model’s curve was labeled with its corresponding Hosmer-Lemeshow test P-value, where P > 0.05 indicated good calibration. Points above/below the diagonal represent over/under-prediction respectively

• 

  Fig. 3. SHAP analysis of the XGBoost model. (A) Waterfall plot. It illustrated how individual features contribute to shifting the base prediction value (left) to the final model output (right) for a single case. Features were ordered vertically by contribution magnitude, with red/blue bars indicating positive/negative impacts respectively. (B) Feature importance plot. It displayed global feature importance ranked by mean absolute SHAP values (x-axis). Each point represents a case, colored by feature value (blue = low, red = high), showing how feature values correlate with impact direction. (C) Decision plot. It showed how feature values collectively influence predictions across multiple cases. Each line represents a case’s prediction path from the population baseline (left) to its final prediction (right). Line convergence patterns reveal consistent influential features