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

BackgroundPreeclampsia (PE) is one of the leading factors in maternal and perinatal mortality and morbidity worldwide with no known cure. Delivery timing is key to balancing maternal and fetal risk in pregnancies complicated by PE. Delivery timing of PE patients is traditionally determined by closely monitoring over a prolonged time. We developed and externally validated a deep learning models that can predict the time to delivery of PE patients, based on electronic health records (EHR) data by the time of the initial diagnosis, in the hope of reducing the need for close monitoring. MethodUsing the deep-learning survival model (Cox-nnet), we constructed time-to-delivery prediction models for all PE patients and early-onset preeclampsia (EOPE) patients. The discovery cohort consisted of 1,533 PE cases, including 374 EOPE, that were delivered at the University of Michigan Health System (UM) between 2015 and 2021. The validation cohort contained 2,172 PE cases (547 EOPE) from the University of Florida Health System (UF) in the same time period. We built clinically informative baseline models from 45 pre-diagnosis clinical variables that include demographics, medical history, comorbidity, PE severity, and initial diagnosis gestational age features. We also built full models from 60 clinical variables that include additional 15 lab tests and vital signs features around the time of diagnosis. ResultsThe 7-feature baseline models on all PE patients reached C-indices of 0.74 and 0.73 on UM hold-out testing and UF validation dataset respectively, whereas the 12-feature full model had improved C-indices of 0.79 and 0.74 on the same datasets. For the more urgent EOPE cases, the 6-feature baseline model achieved C-indices of 0.68 and 0.63, and its 13-feature full model counterpart reached C-indices of 0.76 and 0.67 in the same datasets. ConclusionsWe successfully developed and externally validated an accurate deep-learning model for time-to-delivery prediction among PE patients at the time of diagnosis, which helps to prepare clinicians and patients for expected deliveries.