Predicting interval from diagnosis to delivery in preeclampsia using electronic health records

Abstract

Preeclampsia is a major cause of maternal and perinatal mortality with no known cure. Delivery timing is critical to balancing maternal and fetal risks. We develop and externally validate PEDeliveryTime, a class of clinically informative models which resulted from deep-learning models, to predict the time from PE diagnosis to delivery using electronic health records. We build the models on 1533 PE cases from the University of Michigan and validate it on 2172 preeclampsia cases from the University of Florida. PEDeliveryTime full model contains only 12 features yet achieves high c-index of 0.79 and 0.74 on the Michigan and Florida data set respectively. For the early-onset preeclampsia subset, the full model reaches 0.76 and 0.67 on the Michigan and Florida test sets. Collectively, these models perform an early assessment of delivery urgency and might help to better prioritize medical resources.

Fig. 1. Experimental design and cohort overview of PEDeliveryTime.

Experiment Design Workflow: The discovery cohort was obtained from the University of Michigan Health System and a validation cohort of similar size and time was obtained from the University of Florida Health System. We constructed 4 predictive models: baseline and full models for all PE patients and baseline and full models for EOPE patients. The input variables in baseline models include patients’ demographics, lifestyle, comorbidities and medical history. The full models include additional laboratory tests and vital signs from within 5 days of PE diagnosis, in addition to the variables in the baseline models. We trained the Cox-nnet prognosis prediction model using 80% training from the discovery cohort, tested it on 20% hold-out data from the discovery cohort, and validated it using the validation cohort. We then built clinically informative models by reducing Cox-nnet features based on both their importance scores and significance levels. The models are examined by the importance scores of top features and stratified survival curves based on patient survival risks. We disseminated the feature-reduced, clinically informative models into a user-friendly web application for healthcare professionals to use. Created in BioRender. Garmire, L. (2025) https://BioRender.com/q52s989.

Fig. 2. PE Baseline model results, interpretation, and evaluation.

A The bar plots of C-indices from the original Cox-nnet models (red) and feature-reduced clinically informative model (blue), on the UM cross-validation(5 technical replications, error bars represent median ±SD) and UM hold-out test set and UF test set. B, E The survival curves and 95% CI (mean ±1.96SEM) of high-risk (top 25%), intermediate-risk (middle 50%) and low-risk groups (bottom 25%), categorized by predicted PI from the reduced baseline model in A on (B) hold-out UM test data and (E) UF test data. C, F ROC curves of prediction delivery time within 2 days, 7 days and 14 days using results from reduced baseline model on (C) UM hold-out test data and (F) UF test data. D The ln-transformed permutation importance scores of features in the feature-reduced baseline model. A positive sign indicates that a higher value in feature is associated with a shorter diagnosis-to-delivery time and a negative sign means an extension of diagnosis-to-delivery time. Color represents p-values of two-sided univariate Cox-PH test. G–L The distribution of diagnosis gestational age, sPE rate and PE in prior pregnancy rate, in associations with delivery gestational week (G–I) and time (days) to delivery (J–L), represented by different colors.

Fig. 3. PE Full model results, interpretation and evaluation.

A The bar plots of C-indices from the original models (red) and feature-reduced clinical informative model (blue), on the UM training cross-validation (5 technical replications, error bars represent median ±SD) and UM hold-out test data and UF test set. B, E The survival curves and 95% CI(mean ±1.96SEM) of high-risk (top 25%), intermediate-risk (middle 50%) and low-risk groups (bottom 25%), categorized by predicted PI from the reduced full model in A on (B) UM hold-out test data, (E) UF test set. C, F ROC curves of prediction delivery time within 2 days, 7 days and 14 days using results from reduced full model (A) on (C) hold-out UM test data and (F) UF test data. D The ln-transformed permutation importance scores of features in the feature-reduced baseline model. A positive sign indicates that a higher value in the feature is associated with a shorter diagnosis-to-delivery time and a negative sign means an extension of diagnosis-to-delivery time. Color represents p-values of two-sided univariate Cox-PH test. G–I The distribution of aspartate aminotransferase (AST) values, the standard deviation of diastolic blood pressure (DBP) and the standard deviation of respiratory rate (RR), in association with time (days) to delivery, represented by different colors.

Fig. 4. Results, interpretation and evaluation of baseline and full models on the EOPE patient subset.

A The bar plots of C-indices from the original Cox-nnet EOPE baseline model (red) and feature-reduced clinically informative model (blue) on the UM cross-validation (5 technical replications, error bars represent median ±SD), UM test set and UF test set. B, C The survival curves and 95% CI(mean ±1.96SEM) of high-risk (top 25%), intermediate-risk (middle 50%) and low-risk groups (bottom 25%), categorized by predicted PI from the reduced EOPE baseline model in (A) on (B) UM hold-out test set, (E) UF test data. C, F ROC curves of prediction delivery time within 2 days, 7 days and 14 days using results from reduced EOPE baseline model (A) on (C) hold-out UM test data and (F) UF test data. D The ln-transformed permutation importance score of features in the EOPE full model. Color represents p-values of two-sided univariate Cox-PH test. G The bar plots of C-indices from the original Cox-nnet EOPE full model (red) and its feature-reduced clinically informative model (blue) on the cross-validation and UM test set(5 technical replications, data are represented as mean ±1.96SEM) and UF test set. H, K The survival curves and 95% CI(mean ±1.96SEM) of high-risk (top 25%), intermediate-risk (middle 50%) and low-risk groups (bottom 25%), categorized by predicted PI from the reduced EOPE full model in (G) on (H) UM hold-out test data, (K) UF test data. I, L ROC curves of prediction delivery time within 2 days, 7 days and 14 days using results from reduced EOPE full model (G) on (I) UM hold-out test data and (L) UF test data. J The ln-transformed permutation importance scores of features in the EOPE full model. Color represents p-values of two-sided univariate Cox-PH test. M–O Analysis of creatinine values among the EOPE patients in the discovery cohort. M The dichotomized survival curves and 95% CI(mean ±1.96SEM) by the median value of creatinine. N, O Distributions of creatinine values by delivery gestational week (N) and diagnosis-to-delivery time (O). P–R Analysis of platelet counts among the EOPE patients in the discovery cohort. P The dichotomized survival curves and 95% CI(mean ±1.96SEM) by the median value of platelet counts. Q, R Distributions of creatinine values by delivery gestational week (Q) and diagnosis-to-delivery time (R).

Fig. 5. Comparison of important features among the four feature-reduced clinically informative models.

A The bubble plot of important features from PE baseline, EOPE baseline, PE full, and EOPE full models using reduced top important features. The size of the bubbles represents the permutation importance score of each feature. Color represents the sign of features in the diagnosis-to-delivery time prediction: a positive sign indicates that a higher value in the feature is associated with a shorter diagnosis-to-delivery time and a negative sign means an extension of diagnosis-to-delivery time. B Venn diagram of the important features from the four models shown in (A).