Deep learning-based prognosis prediction among preeclamptic pregnancies using electronic health record data

Background Preeclampsia (PE) is one of the leading factors in maternal and perinatal mortality and morbidity worldwide. The only cure for PE to date is to deliver the placenta and stop gestation. However, the timing of delivery among PE patients is essential to minimize the risk of severe maternal morbidities, and at the same time ensure the survival of the baby.

Methods In this study, we constructed a series of deep learning-based models to predict the prognosis, or the time to delivery, since the initial diagnosis of PE using electronic health record (EHR) data. We extracted and processed 1578 pregnancies in Michigan Medicine at the University of Michigan in 2015-2021 as the discovery cohort. Using the Cox-nnet v2 algorithm, we built the baseline model with EHR information prior to diagnosis, as well as the full model including baseline information and lab testing results and vital signs at the time of diagnosis. We evaluated the models using the C-index and log-rank p-values in KM survival curves, using both 20% testing data of the Michigan cohort, as well as 1177 PE pregnancy EHR data from the Medical Center of the University of Florida.

Results The baseline prognosis model for time to delivery since PE diagnosis achieved C-index values of 0.75 and 0.72 on the training and testing set respectively. While the full model reached C-indices of 0.77 and 0.74 in the same training and testing sets. Both models performed better than their Cox-PH model counterparts. The seven most important features in the baseline model in descending order were diagnosis gestational age, severe PE, past PE, age, parity, gravidity, and uncomplicated diabetes. Meanwhile, 14 most important features were selected and interpreted in the full model, including diagnosis gestational age, parity, severe PE, past PE, features in lab tests (white blood cell, platelet, and red blood cell counts, AST value), min respiratory rate, and features measuring blood pressure (minimum, mean and standard deviation of systolic blood pressure, and maximum and standard deviation of diastolic blood pressure).

Conclusion The time to delivery predicting models provide clinicians valuable tools and options to quantify the delivery risks and make better decisions on the optimal delivery time of PE patients at the time of diagnosis. Implementation of these actionable models into PE clinical care practice is expected to significantly improve the management of PE patients.