Sepsis Prediction and Vital Signs Ranking in Intensive Care Unit Patients

• DOI: 10.48550/arxiv.1812.06686
• Publication Date: 2018-01-01
• Authors: Avijit Mitra; Khalid Ashraf
• Keywords: Vital signs; Predictive modelling

We study multiple rule-based and machine learning (ML) models for sepsis detection. report the first neural network detection prediction results on three categories of sepsis. have used retrospective Medical Information Mart Intensive Care (MIMIC)-III dataset, restricted to intensive care unit (ICU) patients. Features were created from only common vital sign measurements. show significant improvement AUC score using based ensemble model compared single ML models. For sepsis, severe septic shock, our achieves an 0.97, 0.96 0.91, respectively. Four hours before positive hours, it predicts same with 0.90, 0.91 0.90 Further, we ranked features found that six signs consistently provides higher all tested. Our novel highest in detecting predicting shock MIMIC-III ICU patients, is amenable deployment hospital settings.