Predictive factors of advanced interventional procedures in a multicentersevere postpartum hemorrhage study

Electronic Supplementary Material # 2

Details on the statistical analysis

Results are expressed as mean and standard deviation (sd), median and first and third quartiles (Q_1 -Q_3 ) or counts and percent. The study outcome was defined as parturients undergoing at least one interventional procedure defined as either uterine artery embolisation or intra-abdominal packing or arterial ligation or hysterectomy. Marginal association between single variables and outcome was assessed by a Wilcoxon rank-sum test for quantitative variables and Fisher’s exact test. First analyses and model building were performed on the 2004-2005 database.

Multiple logistic regression was used to determine a set of variables independently associated with each outcome. Variables associated with intervention at a 0.15 level and with less than 5% of missing data were considered in the multiple model. Missing data were imputed by use of the multiple imputation by chained equation which resulted in an imputed dataset [1][2]. Of note, for clinical application, the following continuous covariates were categorized: systolic blood pressure (SBP) < 90 mmHg, diastolic blood pressure (DBP) < 55 mmHg and heart rate (HR) > 115 bpm (as described previously [3]), prothrombin time (PT) < 50% and fibrinogen < 2 g.l-1 (the latter 2 covariates both constituting the threshold for FFP transfusion in Lariboisière Hospital center and also described by Charbit et al [4] as severity factor in PPH), troponin detectable or not (enabling the use of semi-quantitative measure). The model selection involved the following steps: 1000 bootstrap samples were drawn with replacement, a backward stepwise variable selection algorithm was performed on each bootstrap sample (stopping rule based on p-value cut-off at 0.05) and the variables selected in more than 50% of selections were kept for the final model. The validity of the final logistic regression model was checked using le Cessie and van Houwelingen goodness-of-fit test [5]. Discriminative ability of the final model was evaluated by the c- index (identical to the area under the receiver operating characteristics (ROC) curve) [6]. A corrected c-index was estimated using a new bootstrap procedure : differences between the performance on the bootstrap sample and the original sample were taken as a measure of the overoptimism of the selected model[7][8]. The model was finally used to define a simple clinical prognostic score within the range [0,5] as a linear predictor with a unit coefficient associated to each the five final selected variables. In terms of c-index, the loss of that score as compared to the fitted model was less than 0.01. The linear predictor was then split in three categories namely 0, 1, or more than 2 with still a loss of 0.01 in terms of c-index.

To assess its external validity, the score was evaluated in the multicenter validation cohort using a logistic regression model. Odds ratios and their 95% CI were calculated and potential interaction between score and center was tested. ROC curves were also presented with their estimated AUC, for the external validation cohort (patients from other centers), the internal validation cohort (patients from LariboisièreHospital) and the whole validation cohort comprising of both previous samples

All tests were two-sided at the 0.05 significance level. Analyses were performed using R statistical package (online at

References

1.Brand JPL (1999) Development, implementation and evaluation of multiple imputation strategies for the statistical analysis of incomplete data sets. Leiden and Erasmus University, Rotterdam

2.van Buuren S, Boshuizen HC, Knook DL, (1999) Multiple imputation of missing blood pressure covariates in survival analysis. Stat Med 18: 681-694

3.Karpati PC, Rossignol M, Pirot M, Cholley B, Vicaut E, Henry P, Kevorkian JP, Schurando P, Peynet J, Jacob D, Payen D, Mebazaa A, (2004) High incidence of myocardial ischemia during postpartum hemorrhage. Anesthesiology 100: 30-36; discussion 35A

4.Charbit B, Mandelbrot L, Samain E, Baron G, Haddaoui B, Keita H, Sibony O, Mahieu-Caputo D, Hurtaud-Roux MF, Huisse MG, Denninger MH, de Prost D, (2007) The decrease of fibrinogen is an early predictor of the severity of postpartum hemorrhage. J Thromb Haemost 5: 266-273

5.le Cessie S, HC vH, (1991) A goodness of fit test for binary regression models, based on smoothing methods. Biometrics 47: 1267-1282

6.Harrell FE, Jr., Califf RM, Pryor DB, Lee KL, Rosati RA, (1982) Evaluating the yield of medical tests. Jama 247: 2543-2546

7.Steyerberg EW, Eijkemans MJ, Harrell FE, Jr., Habbema JD, (2001) Prognostic modeling with logistic regression analysis: in search of a sensible strategy in small data sets. Med Decis Making 21: 45-56

8.Harrell FE, Jr., Lee KL, Mark DB, (1996) Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med 15: 361-387