Deep learning for early prediction of 30-day in-hospital mortality in emergency cardiovascular patients: a retrospective model-development study

Early identification of emergency cardiovascular patients at high risk of death may support monitoring and timely escalation of care. We developed a deep-learning model for 30-day all-cause in-hospital mortality using routinely available admission data. This retrospective single-center study included 241 adults admitted to an emergency department with cardiovascular disease in 2021; 30 died in hospital within 30 days. Nineteen variables recorded within 6 h were evaluated using penalized logistic regression, random forest, gradient boosting, k-nearest neighbors, eXtreme Gradient Boosting, a deep neural network (DNN), a residual multilayer perceptron, and an FT-Transformer. Stratified five-fold cross-validation produced one out-of-fold probability per patient. Scaling and Synthetic Minority Over-sampling Technique were confined to each training fold. The DNN achieved a receiver operating characteristic area under the curve (ROC AUC) of 0.903 (95% confidence interval [CI] 0.822–0.969), a precision-recall AUC of 0.795 (0.669–0.906), and a Brier score of 0.050 (0.033–0.069). The DNN had the highest precision-recall AUC and lowest Brier score, and gradient boosting had the highest ROC AUC. Across the seven comparator models, ROC AUC ranged from 0.769 to 0.927, precision-recall AUC from 0.420 to 0.709, and Brier score from 0.059 to 0.096. Integrated gradients identified peripheral oxygen saturation, lactate, white blood cell count, and systolic blood pressure as the largest contributors to DNN predictions. The DNN showed promising performance for early mortality prediction in this small single-center cohort. Its relative performance varied by evaluation metric. External validation and recalibration are required before clinical use.