Discussion
PAS-ID is a multicenter international database that includes data of 797 patients from 11 centers that present 9 countries from Europe, Asia and Africa. The current study was conducted to establish prediction models of different critical outcomes of PAS using emerging ML models. Two prediction models were created to determine risk of adverse maternal outcomes, namely massive PAS-associated peripartum blood loss, admission to ICU and prolonged hospitalization, using antepartum and peripartum inputs. Diagnostic performance of all models ranged between 0.80 and 0.90, which is defined as ‘excellent’ for a diagnostic test (9). Internal validity was demonstrated by consistency of diagnostic performance between train and test sets in all models.
Although PAS has been one of the most concerningly mounting obstetric complications in contemporary practice, evidence-based recommendations on management of PAS are limited and mostly represent level 3, 4 or good practice point recommendations (10). Cesarean hysterectomy seems to be the most acceptable approach whenever possible (3, 10). However, if cesarean hysterectomy is rejected by the patient, practice recommendations are generally less determinate. Of conservative options, leaving the placenta in situ was considered by the Royal College of Obstetricians and Gynaecologists (RCOG) in women who are highly motivated to preserve the uteri (10). This approach was considered investigational by the American College of Obstetricians and Gynaecologists (ACOG) and no recommendations were made in its regard (3). This practice does not seem to be prevalent though (4). Other uterine preserving techniques were less endorsed.
Nevertheless, management of PAS is globally diverse and is not enclosed by these recommendations (4). This may be attributed to paucity of data. Also, most data come from case reports and case studies that lack a study design and selection criteria, and likely reflect a single surgeon or team experience. Conduction of prospective studies, pilot studies, and clinical trials on management of PAS may be restricted by ethical considerations and recruitment difficulties and therefore, conclusions from large retrospective studies may present the first step to support future prospective studies and enhance evidence on current widely adopted management strategies.
To our knowledge, PAS-ID may present the first international multicenter database that investigates clinical outcomes of PAS in centers that offer both cesarean hysterectomy and conservative management. The database is one of the largest databases available on PAS in the literature and it conveys a wide range of practice. The current study applied ML algorithms, which tends to provide accurate prediction and enclose complex and hidden relations between studied variables and outcomes. Although ML is generally used with large databases to permit model learning, current prediction models seem to perform better that traditional statistics and to yield consistent performance on untested data and excellent internal validity. Clinically, MOGGE PAR-A score can be used to determine high risk group, who may benefit from additional interventions (e.g. prophylactic IR procedures). Similarly, MOGGE PAR-P score may determine women in whom significant bleeding is anticipated. The score may be used to outline intraoperative management in priori, by calculating risk using different scenarios. Specifically, it may help to avoid unnecessary steps, which may not seem to lower risk of these outcomes (e.g. internal iliac artery ligation), to determine whether certain measurements may be helpful (e.g. preoperative or intraoperative ultrasound), and whether some intraoperative steps would be safe to do (e.g. delayed cord clamping).
The current study is limited by the retrospective nature of the study. Although the study was based on an international database, generalizability to some regions of the world, that were not represented in this data, cannot be determined. Although PAS-related research is associated with inherited risk of reflecting a specific team practice, our data were representative of management approaches that are widely recognizable in the literature. Complexity of ML models may present a barrier to their applicability. However, these calculations can be programmed into an application to facilitate their use in clinical practice. For this purpose, a simple tool (MOGGE PAR score, version 1.0) was created to enable use of these models and is available at (https://www.mogge-obgyn.com/clinical-studies) for research purposes (Figure S1).
In conclusion, utilization of ML algorithms may provide an individualized tool to determine women with PAS who are at high risk of significant morbidity and to optimize management plan in priori based on available information. Prospective validation of these scores may permit robust evidence-based recommendations on management of PAS, converge current treatment options, and determine required training and skills that would be deemed satisfactory to manage PAS if uterine preservation is considered.
Acknowledgements: none
Disclosure of Interests: The authors have no conflicts of interest. No financial disclosure to declare