Prediction of postpartum hemorrhage using traditional statistical analysis and a machine learning approachAJOG Global Reports at a Glance

• Published in: AJOG global reports Vol. 3; no. 2; p. 100185
• Main Authors: Vahid Mehrnoush, MD, Amene Ranjbar, MD, Mohammadsadegh Vahidi Farashah, PhD, Fatemeh Darsareh, PhD, Mitra Shekari, MSc, Malihe Shirzadfard Jahromi, MSc
• Format: Journal Article
• Language: English
• Published: Elsevier 01.05.2023
• Subjects: analysis; machine learning; postpartum hemorrhage; risk factors
• ISSN: 2666-5778; 2666-5778

BACKGROUND: Early detection of postpartum hemorrhage risk factors by healthcare providers during pregnancy and the postpartum period may allow healthcare providers to act to prevent it. Developing a prediction model that incorporates several risk factors and accurately calculates the overall risk for postpartum hemorrhage is necessary. OBJECTIVE: This study used a traditional analytical approach and a machine learning model to predict postpartum hemorrhage. STUDY DESIGN: Women who gave birth at the Khaleej-e-Fars Hospital in Bandar Abbas, Iran, were evaluated retrospectively between January 1, 2020, and January 1, 2022. These pregnant women were divided into 2 groups, namely those who had postpartum hemorrhage and those who did not. We used 2 approaches for the analysis. At the first level, we used the traditional analysis methods. Demographic factors, maternal comorbidities, and obstetrical factors were compared between the 2 groups. A bivariate logistic regression analysis of the risk factors for postpartum hemorrhage was done to estimate the crude odds ratios and their 95% confidence intervals. In the second level, we used machine learning approaches to predict postpartum hemorrhage. RESULTS: Of the 8888 deliveries, we identified 163 women with recorded postpartum hemorrhage, giving a frequency of 1.8%. According to a traditional analysis, factors associated with an increased risk for postpartum hemorrhage in a bivariate logistic regression analysis were living in a rural area (odds ratio, 1.41; 95% confidence interval, 1.08–1.98); primiparity (odds ratio, 3.16; 95% confidence interval, 1.90–4.75); mild to moderate anemia (odds ratio, 5.94; 95% confidence interval 2.81–8.34); severe anemia (odds ratio, 6.01; 95% confidence interval 3.89–11.09); abnormal placentation (odds ratio, 7.66; 95% confidence interval, 2.81–17.34); fetal macrosomia (odds ratio, 8.14; 95% confidence interval, 1.02–14.47); shoulder dystocia (odds ratio, 7.88; 95% confidence interval, 1.07–13.99); vacuum delivery (odds ratio, 2.01; 95% confidence interval, 1.15–5.98); cesarean delivery (odds ratio, 1.86; 95% confidence interval, 1.12–3.79); and general anesthesia during cesarean delivery (odds ratio, 7.66; 95 % confidence interval, 3.11–9.36). According to machine learning analysis, the top 5 algorithms were XGBoost regression (area under the receiver operating characteristic curve of 99%), XGBoost classification (area under the receiver operating characteristic curve of 98%), LightGBM (area under the receiver operating characteristic curve of 94%), random forest regression (area under the receiver operating characteristic curve of 86%), and linear regression (area under the receiver operating characteristic curve of 78%). However, after considering all performance parameters, the XGBoost classification was found to be the best model to predict postpartum hemorrhage. The importance of the variables in the linear regression model, similar to traditional analysis methods, revealed that macrosomia, general anesthesia, anemia, shoulder dystocia, and abnormal placentation were considered to be weighted factors, whereas XGBoost classification considered living residency, parity, cesarean delivery, education, and induced labor to be weighted factors. CONCLUSION: Risk factors for postpartum hemorrhage can be identified using traditional statistical analysis and a machine learning model. Machine learning models were a credible approach for improving postpartum hemorrhage prediction with high accuracy. More research should be conducted to analyze appropriate variables and prepare big data to determine the best model.