Sepsis, a life-threatening condition caused by the body's dysregulated response to infection, remains a leading cause of mortality in ICUs worldwide. Despite advancements in medical technology, accurately predicting sepsis outcomes remains a significant challenge. Traditional scoring systems, such as APACHE-II, often fall short in providing timely and precise risk assessments.
In this study, researchers developed a two-stage Transformer-based model that processes hourly and daily time-series data from ICU patients. The model, trained on data from over 13,000 sepsis patients, demonstrated robust predictive performance, with an AUC of 0.92 by the fifth day of ICU admission. This improvement reflects the model's ability to assimilate longitudinal physiological patterns, offering clinicians a powerful tool for early intervention.
The study also utilized SHAP-derived temporal heatmaps to visualize mortality-associated feature dynamics over time. These heatmaps revealed key biomarkers, such as lactate levels, tidal volume, and chloride levels, which are strongly correlated with patient outcomes. This visualization bridges the gap between model predictions and clinically interpretable biomarkers, providing valuable insights for clinicians.
The model's external validation across diverse cohorts, including Chinese sepsis data and the MIMIC-IV database, confirmed its generalizability. With an accuracy of 81.8% on Chinese data and 76.56% on MIMIC-IV, the model showcases its adaptability across different populations and healthcare settings.