Materials and methods
Study Population The “Placenta Accreta Spectrum International Database (PAS-ID)” is an international database that was launched by Middle-East Obstetrics and Gynaecology Graduate Education (MOGGE) Foundation to conduct the current study (ClinicalTrials.gov identifier: NCT04384510). The database was created on January 21st, 2020 and received contribution from a consortium of 11 tertiary centers located in 9 countries that represent 3 continents. These centers are referral centers for complex PAS cases and they all offer both cesarean hysterectomy and uterine preservation procedures. Data of all patients with PAS who were managed in these centers between January 1st, 2010 and December 31st, 2019 were retrospectively collected. Patients were considered eligible if they received clinical and histopathological diagnosis of PAS and were managed, delivered, and followed-up for 6 weeks postpartum by their respective study site. Exclusion of candidates was made if relevant documented information and follow-up was deficient (e.g. single antenatal visit) or if no authorization to use anonymous patient data was provided for research purposes. Data were collected using a standardized spreadsheet, which included 57 variables that comprise patient baseline information (e.g. age, parity, body mass index “BMI”, ethnicity, smoking status), obstetric and gynaecologic data(e.g. obstetric complications, previous CS, prior gynaecologic surgeries), medical history, antepartum and intrapartum disease characteristics (e.g. PAS type, complete versus focal uterine wall invasion, bladder invasion, parametrial invasion, placental location), diagnosis (antepartum versus intrapartum diagnosis, imaging modality, and gestational age at diagnosis), antepartum hemoglobin level, intraoperative details (e.g. hysterectomy versus uterine preservation, uterus preserving procedures used either surgical or IR-related, success of uterine preservation, use of preoperative or intraoperative sonographic assessment, type of uterine incision and its relation to the placenta, intraoperative blood loss, transfused blood products, surgical complications), maternal outcomes (success of uterine preservation, length of hospital stay, admission to intensive care unit [ICU], postoperative complications), and neonatal outcome(APGAR score at 1 and 5 minutes, admission to NICU, need for respiratory support, neonatal morbidity and mortality). Data collection was completed on June 15th, 2020. Institutional review board (IRB) approval was obtained from all participating centers. Study Outcomes Primary outcome of this study was massive PAS-associated blood loss, which we defined as intraoperative blood loss ≥ 2500 ml, blood loss that required massive blood transfusion (transfusion of ≥ 10 units of packed red blood cells [RBCs] within 24 hours), or blood loss that was complicated by intraoperative disseminated intravascular coagulopathy (DIC). Secondary outcomes included maternal admission to ICU and prolonged hospital stay (postpartum hospital stay for more than 7 days). Prediction models PAS-ID was used to establish an antepartum prediction model to calculate a score that presents probability of peripartum massive PAS-associated blood loss, admission to ICU and prolonged hospital stay. “MOGGE placenta accreta risk-antepartum score” or “MOGGE PAR-A score” aims at predicting these outcomes once PAS diagnosis is made antenatally. “MOGGE placenta accreta risk-peripartum score” or “MOGGE PAR-P score” is a second scoring system that was created to predict the same outcomes using baseline features in conjugation with disease- and surgery-related peripartum variables. This score is designed to calculate probability of unfavorable outcomes of a management strategy and clinical scenario(s) in priori, and would, thereby, assist designation of management. Statistical analysis Conventional statistics Variables were described as means and standard deviations for continuous variables, and numbers (percentages) for categorical variables. Missing data were generally less than 5% in all variables. For reason of comparison, a prediction model of the primary outcome was created using conventional statistics. Data were randomly split into a model development group and model validation group in a 4:1 ratio. Within model development group, each independent variable was tested using univariable logistic regression. Results were expressed in unadjusted odds ratio (OR) and 95% confidence interval (CI). Variables that exhibited a p-value of less than 0.2 on univariable logistic regression were included in a multivariable logistic regression model and adjusted ORs (aORs) were calculated. The diagnostic performance of prediction model was evaluated using receiver operating characteristic (ROC) curve, which was applied to both model development and validation groups. Statistical analysis for this part was performed using STATA 16 software (StataCorp, College Station, TX).ML prediction model ML model was applied using python® programing language (Spyder 3.3.6) with Scikit‐learn (ML library package) through Anaconda 3.0 platform. For purpose of training and validation, data were randomly assigned to a train set (0.8) and test set (0.2). The model was developed using the train set and was applied to the test set to assess internal validation. A ‘train/test split’ technique was considered over k-fold cross-validity because it is associated with unbiased performance regardless of sample size (8). A logistic regression algorithm with gradient descent was performed on a train set using L-BFGS solver with a maximum iteration set to 1000. Algorithms were all successfully converged at less than 10 iterations in all models. Each model was evaluated using Jaccard index, confusion matrix, weighted precision, recall, F1 score, and log loss were calculated. A ROC curve was used to assess diagnostic performance of each model through the test set to assess model validity. Intercept value and coefficients of each model were used to calculate probability of the specific outcome. Range of calculated probability of each outcome among women who did and did not develop this outcome was graphed using a “box and whisker” plot. The graph was created to provide a reference to facilitate interpretation of calculated probabilities in clinical setting.