Point of Care Prediction of Maternal Admission to the Intensive Care Unit Using Machine Learning

Over 90% of maternal mortality occurs in wards/outpatient areas without Intensive Care Unit (ICU) utilization; delays in assessing maternal critical illness cause underutilization of the ICU. Despite this, few advanced prediction models exist to prognose ICU admission risk and existing models have low accuracy/recall. Flagging high risk patients, ahead of critical complications, signals healthcare providers to alter clinical management, improve provider coordination, and allow proper utilization of ICU services when indicated, improving patient outcomes. We develop machine learning (ML) models, trained on data available at the point of care, like patient demographics and clinical history, to predict maternal ICU admission risk. Patients from 2018-2020 in the U.S. Vital Statistics dataset without missing data met inclusion criteria for the study. Any variables with over 50% missing values and any clinical variables accessible during/after the intrapartum period were discarded to allow the model to predict solely on point-of-care variables. An extreme gradient boosting ML model was developed on the data, and a Bayesian Tree-Structured Parzen Estimator algorithm was utilized to optimize hyperparameters. The model was trained on patients from 2018- 2019 and was tested on a hold-out set of patients from 2020. 51 clinical parameters for 5,352,142 obstetric patients from the 2018-2019 years were included in the training data, of which 7,351 were admitted to the ICU. The model was tested on 2,574,457 patients from the 2020 year, of which 3,346 patients were admitted to the ICU. The extreme gradient boosting model had an AUC ROC of 0.78 (95% CI: 0.77-0.79), an accuracy of 99.89% (95% CI: 99.88%-99.89%) with 71.4% recall after threshold tuning. Hyperparameter optimized ML methods can predict maternal ICU admission with high accuracy discrimination ability. Our model paves the way for translating obstetrical data into a clinical platform with personalized risk scores to guide clinical management and potentially improve outcomes at the point of care.