Streamlined machine learning model for early sepsis risk prediction in burn patients

Abstract Sepsis is the leading cause of mortality in burn patients, yet early identification remains difficult due to persistent hyperinflammatory responses and altered baseline physiology. We developed a streamlined machine learning model for early sepsis risk prediction in burn patients using data from 6629 patients across 11 centers participating in the German Burn Registry. The model was trained using only six admission-level variables (age, burned body surface area, deep partial-thickness burns, full-thickness burns, inhalation injury, and hypertension), selected through multiple feature selection methods and evaluated using cross-validated machine learning pipelines. The final Random Forest model achieved an AUROC of 0.91, sensitivity of 0.81, specificity of 0.85, and a negative predictive value of 0.98, enabling reliable early risk stratification immediately upon ICU admission. By relying solely on admission-level variables, this model offers a reliable and interpretable solution for early sepsis risk detection in burn patients, supporting timely interventions and potentially improving critical care outcomes.