The objectives were to develop and validate prediction models for OASI in different risk scenarios and identify predictor variables with relevant contributions to the performance of the models using nationwide databases. Further, to define the prerequisites for obtaining reliable models based on available predictors and to construct a user-friendly, graphic, interactive, and web-based tool for obstetricians, pregnant women and for educational use.

This was a prospective, register-based, multicentre, national, cohort study. The setting was all maternity units in Sweden (n=45) reporting to the Medical Birth Register (MBR) (1). The MBR was the primary data source, supplemented with information from The Total Population Register, The Swedish National Inpatient Register, and the Swedish Longitudinal Integrated Database for Health Insurance and Labor Market Studies. All 1st and 2nd births with a singleton pregnancy in gestational week ≥37+0 with a cephalic presentation were included from January 1, 2009, to December 31, 2018. OASI was defined as third- and fourth-degree injuries presented as one single group and identified by codes O70.2 and O70.3 (the International Classification of Diseases, 10th revision) and by the surgical code MBC33. 
The design strategy created three birth scenarios according to the risk of OASI: 1. The 1st vaginal delivery in nulliparous women (~5%), 2. The 1st vaginal delivery after a prior cesarean delivery (VBAC) (~11%), and 3. The 2nd vaginal delivery (~2%). Secondly, to order candidate predictors in five domains according to the timeline and availability of predictors: A. Maternal biometrics and characteristics; B. Obstetric information from a previous delivery; C. Maternal morbidity; D. Labor events and interventions (current); E. Infant biometrics. 
We adhered to the procedures for developing and validating prediction models outlined in the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis: the TRIPOD statement (2). A protocol for the study was published at ClinicalTrials.gov on February 1, 2022, NCT 05218837. The latest update was posted on February 17, 2022.
Possible predictors were predetermined before the analyses. The selection was guided by subject matter knowledge and our previous works, a search of systematic reviews in the literature, and retrievable information in the registers. 47 potential predictors were considered and ordered into the five domains. Seven predictors were excluded before or during the analysis. The number of candidate predictors for the selection procedure to obtain the final models of scenarios 1-3 was 28, 34, and 40, respectively (Table).  The reliability of predictions was deemed excellent, with a total number of 609,916 births and 25,245 OASIs. The birth/OASI ratio for the 1st vaginal birth was 17 (332,457/19,723), 10 (22,829/2188) for VBAC, and 76 (254,630/3334) for the 2nd birth. The number of Events (OASIs) per candidate Predictor Variable (EPV) for the final models was 704 (19,723/28), 64 (2188/34), and 83 (3334/40) for scenarios 1-3, respectively.

The statistical technique used was multiple logistic regression with OASI as the binary outcome and several continuous and categorical predictor variables, which were fitted to the complete set of candidate predictors of each risk scenario. The selection of predictors for the final models among candidate predictors was based on backward elimination with minimization of the Bayesian Information Criteria (BIC) as stopping rule. Non-linear effects of continuous predictors were evaluated using natural cubic splines with three knots at the 10th, 50th, and 90th percentiles. Linear and spline effects were included simultaneously in the modelling procedure, enabling a data-driven selection between linear and non-linear trends. Interaction terms that minimized BIC further were added to the model. For internal validation the same procedure was applied using 200 bootstrap samples, which permitted counting the number of times a predictor was selected in each model. For the near external validation, the total study cohort was divided by a non-random temporal split on December 31, 2011, into a more recent dataset, including births in 2012-2018 (n=407,198, ~70%), and an earlier dataset with births in 2009-2011 (n=202,348, ~30%). The performance was evaluated with c-statistics regarding discriminative ability and calibration plots (intercept, slope, and calibration-in-the-large) regarding the agreement between predicted and observed outcomes. For each variable in the respective final model, the β-estimate, standardized β, odds ratio with 95% confidence interval, Wald chi-square, and p-value were presented. To estimate each variable's relative contribution, the proportion of the sum of all standardized β was calculated and presented as a percentage. Statistical analyses were performed using SAS 9.4 (SAS Inc.).

Infant birth weight (IBW), in the current delivery measured postpartum, was the dominant predictor in both high- and low-risk births, accounting for ~30% of the total predictive capacity of the prediction models. The accuracy of the predictions was reduced and disputable (c-statistic 0.59 and 0.59 in Scenario 1 and 2) if IBW was absent. This effect was less pronounced in scenario 3 but the c-statistic still decreased to 0.70 from 0.79 due to the substantial contribution of “Obstetric information from the previous delivery” (previous IBW, OASI, and vacuum extraction; the sum = 35% of the total standardized ß). Together with maternal biometrics, these three predictors contained about 70% of the total prognostic information of the models. The predictors in “Labor events and interventions” in the current delivery accounted for 31%, 31%, and 16% of the total standardized ß in scenario 1-3, respectively. Information from the previous birth was important for women facing their second vaginal delivery, contributing 35% of the overall predictive strength of the final model.

Provided that fetal biometrics are accessed antenatally, all models were accurate and useful in predicting OASI. The results can be used to guide obstetricians and pregnant women regarding the management of delivery. The Domain “Labor events and interventions” predictors are only available during delivery but unobtainable antenatally. This information may, therefore, be added as conditional questions, e.g., “If I were to need a vacuum delivery including a medio-lateral episiotomy – how would that influence the predicted risk of OASI?” The result can be read from the resultant change in the individual predicted probability, with and without the event.

Since information about fetal birth weight is an imperative requirement for an accurate and useful prediction model of OASI, the findings of this study support the ongoing and promising efforts of accessing fetal biometrics antenatally by ultrasound and MRI (3).  An online calculator will be provided for clinical use and as evidence-based support for women's autonomous decision-making.

Cnattingius S, Källén K, Sandström A, et al. The Swedish medical birth register during five decades: documentation of the content and quality of the register. Eur J Epidemiol 2023;38:109-120.
Moons KGM, Altman DG, Reitsma JB, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD): Explanation and Elaboration. Ann Intern Med 2015;162:W1-W73.
Kadji C, Cannie MM, Kang X, et al. Fetal magnetic resonance imaging at 36 weeks predicts neonatal macrosomia: the PREMACRO study. Am J Obstet Gynecol 2022;226:238.e1-12.

Continence 7S1 (2023) 100956
DOI: 10.1016/j.cont.2023.100956