Deep learning-based prognosis models accurately predict the time to delivery among preeclampsia patients using health records at the time of diagnosis

Abstract

BackgroundPreeclampsia (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.MethodsIn 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.ResultsThe 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).ConclusionThe 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.

Discover overlooked complications after preeclampsia using electronic health records

“…Future studies incorporating causal inference or randomized controlled trials may offer more insights into the causal pathways between PE and subsequent health outcomes. Given the increasing recognized subtypes of preeclampsia, which exhibit different pathological processes, diagnosis time, symptoms, time to deliveries, as well as outcome, it may be necessary to stratify PE by subtype if the patient size is su ciently large in future studies [42][43][44]. Lastly, our ndings are purely based on EHR data.…”

Discover overlooked complications after preeclampsia from three real-world medical record datasets of over 100,000 pregnancies

“…Future studies incorporating causal inference or randomized controlled trials may offer more insights into the causal pathways between PE and subsequent health outcomes. Given the increasing recognized subtypes of preeclampsia, which exhibit different pathological processes, diagnosis time, symptoms, time to deliveries, as well as outcome, it may be necessary to stratify PE by subtype if the patient size is sufficiently large in future studies [42][43][44]. Lastly, our findings are purely based on EHR data.…”

Building and validating 5-feature models to predict preeclampsia onset time from electronic health record data

“…The Institutional Review Board of the University of Michigan Medical School (HUM#00168171) approved this usage.The data were downloaded from the Precision Health Analytic platform, a web-based interface to access de-identified EMR data. 8 All pregnant records (between years 2015 and year 2021) with at least one PE diagnosis, based on the International Classification of Diseases (ICD)-10 codes, were extracted (Supplementary Table 1 ). Patients who were diagnosed with HELLP syndrome, chronic hypertension with superimposed PE, and postpartum PE, were removed from the cohort.…”