Sepsis Mortality Prediction with Electronic Health Records Based on Sequential and Attention-Based Models

Sepsis is a leading cause of death in the ICU. The mortality prediction driven by medical data is vitally important for sepsis prevention and treatment. However, the task of risk prediction is particularly challenging due to the complexity and heterogeneity of medical data. The past models tend to focus only on sequential models or time embedding. In this paper, we propose a novel model architecture DiagNet, which utilizes sequential and global analysis of patient information to improve the prediction accuracy. For DiagNet, we design a comparison experiment with four existing models Retain, Dipole, RetainEX and HiTANet, and the ablation study with the next-best variant on the MIMIC-IV and eICU datasets. Evaluations showcase that DiagNet outperforms other four models, especially on the MIMIC-IV dataset, achieving both a superior F1-Score and AUC score. Comparing with the next-best variant of the architecture, DiagNet performs better on the most crucial metric, AUC on both datasets. This research contributes to the field by providing an enhanced model architecture for healthcare risk prediction, offering the potential for improved patient care and outcomes.