Results
Baseline, disease and management characteristics A cohort of 797 patients was recruited in PAS-ID. Of those, 727 women were considered eligible for this current study based on adequate documentation of patient outcomes. Mean age of participants was 33.15 ± 4.93 years, mean parity was 2.20 ± 1.37, and mean BMI was 27.89 ± 4.50 kg/m2. Prevalence of twin pregnancy was 2.6%. Pregnancy was complicated by gestational diabetes and hypertension in 8.6% and 7.1%, respectively. Participants reported history of obstetric dilation and curettage (D & C) in 26.92% and history of gynaecologic D & C in 14.96% of cases. The placenta was most commonly located centrally over the internal os (35.76%). Mean antepartum hemoglobin level was 11.04 ± 1.43 g/dl. Preoperative ultrasound was performed in 95.32% of cases and 4.13% underwent intraoperative ultrasound. Planned cesarean hysterectomy was performed in 18.7% of patients. Placenta accreta was encountered in 41.68%, followed by placenta increta (31.22%) and percreta (27.10%). Bladder invasion was present in 13.07% of cases. The most commonly performed uterine incision was low transverse (56.03%), followed by high transverse incision (27.88%). Delayed cord clamping was done in 16.69% of all deliveries. Incising through the placenta to deliver the fetus occurred in 28.75% of cases. Among women who were conservatively managed, the most commonly performed procedure was compression sutures (32.88%). Different modalities of IR were used in 9.42% of cases. Incidence of unintentional cystotomy was 10.18%. Patient characteristics, PAS characteristic and management details are summarized in Table 1. Primary and secondary outcomes Massive PAS-associated blood loss was reported in 17.74% of all cases. Mean estimated blood loss was 1786.33 ±1707.74 ml. Patients received a mean of 2.66 ± 4.91 of packed RBCs units, 1.37 ± 2.69 of fresh frozen plasma units, 1.42 ± 6.37 of cryoprecipitate and 0.81 ± 4.45 of platelet unites. Peripartum DIC manifested in 5.78% of patients. After delivery, 26% of all patients were admitted to the ICU. Mean length of hospital admission was 6.16 ± 6.36 days. Mean postoperative hemoglobin was 9.33 ± 1.75 g/dl (Table 1). Prediction model using conventional statistics Among prenatally-determined variables, maternal age (aOR 1.06; 95% CI, 1.001 – 1.12), ethnicity (aOR 0.09; 95% CI, 0.04 – 0.23), previous CS (aOR 5.65; 95% CI, 1.91 – 16.73 for previous 2 CSs), prior gynaecologic D & C (aOR 2.83, 95% CI, 1.37 – 5.80), antepartum hemoglobin level (aOR 0.75; 95% CI, 0.62 – 0.90), and intrauterine fetal death (aOR 6.40; 95% CI, 1.04 – 39.48) were significantly associated with risk of massive blood loss. AUC of antepartum model performance was 0.84 and 0.81 among development and validation groups, respectively (Table S1). Adding peripartum variables to the model, variables that exhibited significant association with massive blood loss included ethnicity (aOR 0.18; 95% CI, 0.05 – 0.67), Previous 2 CS (aOR 4.92; 95% CI, 1.34 – 18.05), prior gynaecologic D & C (aOR 4.58; 95% CI, 1.76 – 11.93), intrauterine fetal death (aOR 18.36; 95% CI, 1.68 – 200.73) and preoperative hemoglobin (aOR 0.75; 95% CI, 0.61 – 0.93). Significant intraoperative variables were placental bed sewing (aOR 0.17; 95% CI, 0.04 – 0.70), incising through the placenta (aOR 0.31, 95% CI, 0.14 – 0.69), IR (aOR 3.48; 95% CI, 1.23 – 9.78), complete placental invasion (aOR 3.92; 95% CI, 1.71 - 8.99), and bladder invasion (aOR 5.33; 95% CI, 2.27 – 12.50). AUC for this model was 0.91 for development group and 0.82 for validation group (Table S1). ML prediction modelsAntepartum prediction model (MOGGE PAR-A score)For PAS-associated massive blood loss, diagnostic accuracy of antepartum ML model was 0.84 for both train and test sets (Figure 1A). Model evaluation is summarized in Table 2. The most contributing factors to this model were parity (12%), previous CSs (12%), Asian ethnicity (12%), and centrally located placenta (9%). Size of contribution of baseline variables in this model is illustrated in Figure 2A. Median and interquartile range (IQR) of calculated probability was 13.0 (9.0 – 19.1) in women who did not have massive blood loss and 22.5 (16.3 – 32.5) in women with massive blood loss (Figure 3A). Antepartum ML model for prediction of prolonged hospitalization was associated with an AUC of 0.80 and 0.81 for train and test sets, respectively (Figure 1B). Asian ethnicity was the most influential variable to this model (14%), followed by central placenta (12%), anterior low placenta (10%), antepartum diagnosis using abdominal sonography and Doppler assessment (9%), and antepartum hemoglobin level (8%) (Figure 2B). Women who were hospitalized for > 7 days had a median probability of 12.1 (8.6 – 18.6) compared to women who were hospitalized for a shorter duration (7.1 [4.9 - 10.0]) (Figure 3B). Diagnostic accuracy of antepartum ML prediction of maternal ICU admission was 0.85 (train set) and 0.82 (test set) (Figure 1C). Major contributors to the model include European and Asian races (15% and 13%, respectively), central placenta (12%), posterior and anterior low placentas (7%), parity (5%), BMI ≥ 30 (5%), previous CSs (5%), antepartum hemoglobin (5%) (Figure 2C). Women who were and were not admitted to ICU had a median probability of 27.6 (17.9 – 39.8) and 11.3 (7.1 – 17.6), respectively (Figure 3C).Peripartum prediction model (MOGGE PAR-P score) Peripartum prediction model of massive blood loss yielded an AUC of 0.88 and 0.86 for train and test tests, respectively (Figure 1D). Method of diagnosis of PAS presented a major contribution to this model, including whether ultrasound was combined with Doppler assessment (12%) or magnetic resonance imaging (MRI) (10%), or if diagnosis was first made intrapartum (7%). Other major variables are parametrial invasion (8%), and bladder invasion (5%) (Figure 2D). Median calculated probability among women who had massive blood loss was 27.4 (17.8 – 39.8) and 12.8 (9.4 – 18.3) in women with no massive bleeding (Figure 3D). Regarding prolonged hospitalization, AUC of peripartum model was 0.86 for the train and 0.90 for the test set (Figure 1E). Parametrial invasion contributed the most to this model (12%), followed by high transverse uterine incision (8%), intraoperative ultrasound (7%), bladder invasion (6%) (Figure 2E). Median probability in women who were admitted for longer than 7 days was 10.5 (7.4 – 16.4), while women who were admitted for shorter duration had a median of 6.3 (4.8 – 8.6) (Figure 3E). AUC peripartum prediction model of ICU admission was 0.88 and 0.86 for the train and test sets, respectively. The largest contribution comes from ethnicity (Asian [17%] and European [9%]), delayed cord clamping (7%), complete placental invasion (4%), and internal iliac artery ligation (4%) (Figure 2F). Calculated probability of admission to ICU in women who were or were not admitted to ICU was 11.4 (5.2 – 21.5) and 1.6 (0.5 – 5.2), respectively (Figure 3F).