Dynamic Prediction of 30-Day Mortality in Patients With Trauma Using a Hybrid Neural Network Model: Model Development and Evaluation Study

The model can output 30-day mortality risk at any point along the trauma care trajectory, rather than only at a fixed admission time point. Conventional trauma risk assessment systems are described as simple and data-sparse, whereas this model unifies demographic and comorbidity variables as tabular data and vitals, laboratory results, and medications as sequences with temporally determined dynamic bin sizes, processed through multiple transformer encoder layers before binary classification. Developed on 9,496 patients; holdout evaluation (20%, 1,829 patients) with full-length trajectories yielded AUROC 0.962 (95% CI 0.930-0.994) and AUPRC 0.655 (95% CI 0.545-0.765).

The model demonstrates early discriminative capability from the prehospital phase onward. Active-cohort evaluation shows risk estimation as early as 1 hour from first patient contact, moving the prediction window into prehospital care. Active-cohort AUROC 0.905 (95% CI 0.856-0.953) at 1 hour from first patient contact.

The model showed higher discrimination than two commonly used trauma scores at most time points. Mean ΔAUROC +0.297 versus the Revised Trauma Score, with 50 of 54 time points significant; mean ΔAUROC +0.169 versus the Trauma and Injury Severity Score, with 46 of 54 time points significant. Time-point comparisons on the holdout set, reporting significance counts and mean differences.

The study supports the feasibility of sequential modeling for trauma risk prediction using automatically extracted electronic health record data. Data came from pre- and in-hospital care records in the Capital Region of Denmark between 2017 and 2024; scaling and normalization parameters for continuous inputs were fitted on observed values only, missing values were handled through zero-imputation with auxiliary indicator channels, and missing categorical values were mapped to a reserved token with a learned embedding. Model development and evaluation study reporting AUROC and AUPRC with confidence intervals.

The model targets 30-day all-cause mortality risk prediction across the full trauma care trajectory from prehospital care to discharge, in settings with automatically extractable electronic health records containing time-series data such as vitals, laboratory results, and medications. Its intended uses include triage decision support, patient deterioration alerts, bedside decision-making, and family counseling. The study developed and evaluated the model on data from the Capital Region of Denmark for 2017 to 2024, so results apply to the population and care system represented by that data source.

This is a model development and evaluation study reporting discrimination performance on a holdout set and an active cohort; calibration, clinical utility, and prospective deployment results are not presented in the text. Performance in populations and care systems outside the Capital Region of Denmark, and robustness under different missingness patterns and data quality, remain questions for further research. In addition, the currently loaded text is abstract-level content without figures or supplementary materials, so information on model architecture details, hyperparameters, and subgroup analyses is limited.