Prediction of clinical outcomes in women with placenta accreta spectrum using machine learning models: an international multicenter study
Sherif A Shazly,Ismet Hortu,Jin-Chung Shih,Rauf Melekoglu,Shangrong Fan,Farhat Ul Ain Ahmed,Erbil Karaman,Ildar Fatkullin,Pedro V Pinto,Setyorini Irianti,Joel Noutakdie Tochie,Amr S Abdelbadie,Ahmet M Ergenoglu,Ahmet O Yeniel,Sermet Sagol,Ismail M Itil,Jessica Kang,Kuan-Ying Huang,Ercan Yilmaz,Yiheng Liang,Hijab Aziz,Tayyiba Akhter,Afshan Ambreen,Çağrı Ateş,Yasemin Karaman,Albir Khasanov,Fatkullina Larisa,Nariman Akhmadeev,Adelina Vatanina,Ana Paula Machado,Nuno Montenegro,Jusuf S Effendi,Dodi Suardi,Ahmad Y Pramatirta,Muhamad A Aziz,Amilia Siddiq,Ingrid Ofakem,Julius Sama Dohbit,Mohamed S Fahmy,Mohamed A Anan,and Middle East Obstetrics and Gynecology Graduate Education (MOGGE) foundation – Artificial intelligence (AI) unit
DOI: https://doi.org/10.1080/14767058.2021.1918670
Abstract:Introduction: Placenta accreta spectrum is a major obstetric disorder that is associated with significant morbidity and mortality. The objective of this study is to establish a prediction model of clinical outcomes in these women. Materials and methods: PAS-ID is an international multicenter study that comprises 11 centers from 9 countries. Women who were diagnosed with PAS and were managed in the recruiting centers between 1 January 2010 and 31 December 2019 were included. Data were reanalyzed using machine learning (ML) models, and 2 models were created to predict outcomes using antepartum and perioperative features. ML model was conducted using python® programing language. The primary outcome was massive PAS-associated perioperative blood loss (intraoperative blood loss ≥2500 ml, triggering massive transfusion protocol, or complicated by disseminated intravascular coagulopathy). Other outcomes include prolonged hospitalization >7 days and admission to the intensive care unit (ICU). Results: 727 women with PAS were included. The area under curve (AUC) for ML antepartum prediction model was 0.84, 0.81, and 0.82 for massive blood loss, prolonged hospitalization, and admission to ICU, respectively. Significant contributors to this model were parity, placental site, method of diagnosis, and antepartum hemoglobin. Combining baseline and perioperative variables, the ML model performed at 0.86, 0.90, and 0.86 for study outcomes, respectively. Ethnicity, pelvic invasion, and uterine incision were the most predictive factors in this model. Discussion: ML models can be used to calculate the individualized risk of morbidity in women with PAS. Model-based risk assessment facilitates a priori delineation of management.