Aim: Identify women at risk of severe post-partum hemorrhage (PPH) by building a prediction model based on clinical variables available at PPH diagnosis.
Methods: We analyzed data on a cohort of 7236 women with PPH after vaginal delivery from 106 maternity units. Severe PPH was defined as the loss of more than 2000 mL of blood, peripartum drop in hemoglobin of 4 g/dL or more, transfusion of at least four packed red blood cells, embolization, hemostasis surgery, transfer to an intensive care unit or death. The Akaike criterion helped selecting the covariates of a multivariate logistic regression model. The performance of the model was studied through building a receiver-operator curve (ROC). The relative utility of the final model was used to determine the importance of the model in decision-making.
Results: Among all PPH, the prevalence of severe cases was 18.5%. Several clinical variables were significantly associated with severe PPH (e.g. parity, multiple pregnancy, labor induction, instrumental delivery). The multivariate prediction model was built. The area under the ROC for prediction of severe cases was 0.63 (95% confidence interval, 0.62-0.65). Nevertheless, the sensitivity and specificity of the prediction model were 0.49 and 0.70, respectively, for a threshold at 0.20 (near prevalence). The relative utility was 0.19 for a threshold near prevalence (20%).
Conclusion: Because of important misclassifications, even the best model we could build with the available clinical data cannot be reasonably recommended for routine use. Every patient with PPH should receive most optimal management. Other types of information, possibly laboratory data, are probably needed.
Keywords: expected utility; morbidity; post-partum hemorrhage; prediction.
© 2014 The Authors. Journal of Obstetrics and Gynaecology Research © 2014 Japan Society of Obstetrics and Gynecology.