Evaluating Postpartum Hemorrhage Transfusion Risk with a Machine Learning Model for Informed Consent: Retrospective Cohort Study (Preprint)

<sec> <title>BACKGROUND</title> BACKGROUND Postpartum hemorrhage requiring a blood transfusion is a concern for patients and communities. Prediction of an individual’s transfusion risk prior to delivery can be evaluated with a machine learning for shared decision making. </sec> <sec> <title>OBJECTIVE</title> OBJECTIVES To develop a machine learning model for clinical decision support use, it is proposed that internal, cleaned datasets will improve model performance over models trained with uncleaned datasets. To evaluate the role of delivery mode and indications with an individual’s postpartum transfusion risk, a machine learning model was developed for the purpose of antepartum clinical decision support and quality improvement. </sec> <sec> <title>METHODS</title> METHODS A 10-year retrospective cohort (n=12,634) was analyzed to determine the probability of transfusion based on delivery mode and delivery indications in a community-based health system. Datasets were evaluated with the best performing machine learning model. Prototype desktop clinical support applications for individual transfusion risk assessment were developed using the best performing model. </sec> <sec> <title>RESULTS</title> RESULTS Physician communication alone decreased the number of elective Cesarean sections compared to 10-year historical data, (p value =.02 by paired t-test, and p= &lt;.001). The XGBoost model on the uncleaned dataset (n=1734) performance was an AUC =.71, PR-ROC =.82, and F1 score =.64). The cleaned dataset (n=1734) performance was AUC= .71, PR-ROC =.78., and F1 score= .80. A synthetic dataset derived from the cleaned dataset (n=12,000) was used for validation with an AUC =.71, PR-ROC =.79, and F1 score =. 66. </sec> <sec> <title>CONCLUSIONS</title> CONCLUSION Machine learning models are useful to determine an individual’s postpartum transfusion risk based on delivery mode and indications for delivery. Model performance on uncleaned data was improved in this study, but did not reflect clinical decision-making for elective deliveries. Clinically relevant datasets are important for model training for use in clinical decision support tools development. </sec> <sec> <title>CLINICALTRIAL</title> N/A </sec>