Early Sepsis Prediction in ICU Trauma Patients with Using An Improved Cascade Deep Forest Model

Abstract

In the intensive care unit (ICU), Trauma is one of the leading causes of death worldwide, about 40\% of death occurred during hospitalization and 22\% are caused by sepsis. Sepsis is a complication disease that is easily caused by the ICU patient, especially in trauma patients. Sepsis is a systemic inflammatory response that is very dangerous to patient’s lives. So early prediction of sepsis for trauma patients can help doctors intervene and diagnose in advance to improve treatment and patients outcomes. This work developed an early sepsis prediction diagnostic model by using machine learning method-a improved cascade deep forest model. We extracted the patient’ s Electronic medical record (EMR) data with the first 24 hours of ICU admission to classify patients as sepsis by using the new definition of Sepsis-3. From MIMIC-III dataset, 3125 patients were included(sepsis patients =1187, non sepsis =1938). Compare to different machine learning methods, the improved cascade deep forest model is superior or comparable predictive performance. Our model achieved AUROC of 0. S0, Sensitivity is 0.79 and specificity is 0.64. Besides, related risk factors of sepsis and important predictors variables were scored.