A Machine Learning Sepsis Prediction Algorithm for Intended Intensive Care Unit Use (NAVOY Sepsis): Proof-of-Concept Study

• DOI: 10.2196/28000
• Publication Date: 2021-08-01T20:52:05Z
• Authors: Inger Persson; Andreas Östling; Martin Arlbrandt; Joakim Söderberg; David Becedas
• Keywords: Infectious Medicine; Original Paper; EHR; algorithm; proof of concept; detection; R; Infektionsmedicin; electronic health record; prediction; intensive care unit; 3. Good health; sepsis; 03 medical and health sciences; machine learning; software as a medical device.; 0302 clinical medicine; ICU; Medicine; early detection

Background Despite decades of research, sepsis remains a leading cause mortality and morbidity in intensive care units worldwide. The key to effective management patient outcome is early detection, for which no prospectively validated machine learning prediction algorithm currently available clinical use Europe. Objective We aimed develop high-performance based on routinely collected unit data, designed be implemented European units. Methods was developed using convolutional neural networks, Massachusetts Institute Technology Lab Computational Physiology MIMIC-III data from patients aged 18 years or older. model uses 20 variables produce hourly predictions onset sepsis, defined by international Sepsis-3 criteria. Predictive performance externally hold-out test data. Results algorithm—NAVOY Sepsis—uses 4 hours input can identify with high risk developing (area under the receiver operating characteristics curve 0.90; area precision-recall 0.62) up 3 before onset. Conclusions NAVOY Sepsis superior that existing warning scoring systems comparable those other algorithms predict has excellent predictive properties are