Deep learning-based prognosis models accurately predict the time to delivery among preeclamptic pregnancies using electronic health record

Background Preeclampsia (PE) is one of the leading factors in maternal and perinatal mortality and morbidity worldwide. Delivery timing is key to balancing the risk between severe maternal and neonatal morbidities in pregnancies complicated by PE. Method In this study, we constructed and validated first-of-their-kind deep learning models that can forecast the time to delivery among patients with PE using electronic health records (EHR) data. The discovery cohort consisted of 1,533 preeclamptic pregnancies, including 374 cases of early-onset preeclampsia (EOPE), that were delivered at University of Michigan Health System (UM) between 2015 and 2021. The validation cohort contained 2,172 preeclamptic pregnancies (including 547 EOPE) from University of Florida Health System (UF) in the same period. Using Cox-nnet, a neural network-based prognosis prediction algorithm, we built baseline models of all PE patients and of the subset of EOPE patients, using 47 features on demographics, medical history, comorbidities, the severity of PE, and gestational age of initial PE diagnosis. We also built full models using 62 features, combining those in baseline models and additional features on lab tests and vital signs, on the same PE patients and EOPE subset. The models were re-trained and re-validated using reduced sets of the most important features, to improve their interpretability and clinical applicability. Findings The 7-feature baseline models on all PE patients reached C-indices of 0·73, 0·74 and 0·73 on UM training, hold-out testing and UF validation dataset respectively, whereas the 12-feature full model had improved C-indices of 0·78, 0·79 and 0·74 on the same datasets. For the EOPE cases, the 6-feature baseline model achieved C-indices of 0·67, 0·68 and 0·63 on the training, hold-out testing and UF validation dataset respectively, while its 13-feature full model counterpart reached C-indices of 0·74, 0·76 and 0·67 in the same datasets. Besides confirming the prognostic importance of gestational age at the time of diagnosis and of sPE status, all four models identified parity and PE in prior pregnancies as important features, which are not in the current guidelines for PE delivery timing. Laboratory results and vital signs such as platelet count, the standard deviation of respiratory rate within a 5-day observation window, and mean diastolic blood pressure are critical to increase the accuracy of predicting time to delivery, in addition to testing aspartate aminotransferase and creatinine levels. For EOPE time to delivery prediction, comorbidities such as pulmonary circulation disorders and coagulopathy as defined in Elixhauser Comorbidity Index are important to consider. Interpretation We set up a user-friendly web interface to allow personalized PE time to delivery prediction. The app is available at These actionable models may help providers to plan antepartum care in these pregnancies and significantly improve the management/clinical outcomes of pregnancies affected by PE. Funding This study is funded by the National Institutes of Health Evidence before this study Determining the optimal delivery time is essential in preeclampsia management to balance the risk of maternal and neonatal morbidities. Current clinical guidelines for delivery timing in preeclampsia, according to the American College of Obstetricians and Gynecologists (ACOG), mainly depend on the gestational age at diagnosis and the severity of PE. However, the current knowledge doesn’t provide a quantitative prediction of patients’ risk of delivery, nor does it discuss the effect of some important phenotypic factors (eg. patients’ demographics, lifestyles and comorbidities) on delivery time. Rather, according to a systematic review published in 2021, 18 prior studies predicted the timing of delivery for preeclampsia using biomarkers, which are yet to be implemented in routine checkups in pregnancy. On the other hand, EHR data are routinely collected but often overlooked information, with huge potential to predict challenging time to delivery problems such as those in PE. Added value of this study To our knowledge, these are the first deep-learning-based time to delivery prediction models for PE and EOPE patients using routine clinical and demographic variables. We enlist the quantitative values of critical EHR features informative of delivery time among PE patients, many of which are newly reported clinical features. We disseminate these models by the web tool “PE time to delivery Predictor”. Implications of all the available evidence All models are externally validated with a large EHR dataset from the University of Florida Health System. Adopting these models may provide clinicians and patients with valuable management plans to predict and prepare for the best delivery times of pregnancies complicated by PE, especially for EOPE cases in which consequences of early delivery are more significant. Further prospective investigation of these models’ performance is necessary to provide feedback and potential improvement of this model. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement LXG was supported by grants K01ES025434 awarded by NIEHS through funds provided by the trans-NIH Big Data to Knowledge (BD2K) initiative ([www.bd2k.nih.gov][1]), R01 LM012373 and LM012907 awarded by NLM, R01 HD084633 awarded by NICHD. DJL was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (K01DK115632) and the University of Florida Clinical and Translational Science Institute (UL1TR001427). AM is supported by the National Center for Advancing Translational Science (5TL1TR001428). ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: IRB of the University of Michigan Medical School gave ethical approval for this work(HUM#00168171). I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines and uploaded the relevant EQUATOR Network research reporting checklist(s) and other pertinent material as supplementary files, if applicable. Yes All data produced in the present study are available upon reasonable request to the authors * PE : preeclampsia EOPE : early-onset preeclampsia LOPE : late-onset preeclampsia EHR : electronic health record SBP : systolic blood pressure DBP : diastolic blood pressure RR : respiratory rate HELLP : hemolysis, elevated liver enzymes, low platelet count AST : aspartate transaminase PI : prognosis score UM : University of Michigan UF : University of Florida ICD-10 : The International Classification of Diseases, Tenth Revision MAP : mean arterial pressure UtA-PI : uterine artery pulsatility index PLGF : placental growth factor ACOG : American College of Obstetricians and Gynecologists [1]: http://www.bd2k.nih.gov