ABODE-VAE: An Autonomous Feature Importance Learning Framework for Age-Stratified Mortality Prediction in Sepsis Patients

Sepsis is a serious disease that leads to life-threatening organ dysfunction and is associated with a high risk of mortality. Effective early in-hospital mortality prediction is essential for improving patient prognosis. However, most current predictive models suffer from two major limitations: class imbalance in clinical datasets and the insufficient interpretability of the models needed to support clinical decisions. Therefore, we propose a deep learning model named the attention-based ordinary differential equation variational autoencoder (ABODE-VAE), which autonomously assigns importance scores to features, thereby enhancing model interpretability. In addition, the model integrates the weighted representations with minority-class latent features learned by the variational autoencoder, incorporating reconstruction-error signals to enhance minority representations and thus mitigate class imbalance. This study analysed intensive care unit (ICU) patients diagnosed with sepsis from the Medical Information Mart for Intensive Care IV (MIMIC-IV), focusing on two distinct age cohorts: younger adults (18–65 years) and older adults ( \(\ge \) 65 years). The experimental results demonstrate that the proposed model achieves outstanding performance, with area under the receiver operating characteristic curves (AUROCs) of 0.8808 and 0.8643, respectively. Moreover, several key mortality predictors consistently identified across both age cohorts include the Glasgow Coma Scale score, oxygen saturation, mechanical ventilation, and body temperature.