Sepsis: current challenges and new solutions based on modern technologies. A clinical management approach

Abstract

Introduction: Despite high associated mortality and high treatment costs, sepsis remains difficult to diagnose. A recent supplement to sepsis management are systems based on machine learning (ML). Purpose: Proof of concept and presentation of a MLbased clinical application for the early prediction of sepsis. Material and methods: The data comes from the publicly accessible database Early Prediction of Sepsis from Clinical Data - the PhysioNet Computing in Cardiology Challenge 2019 and include 40366 intensive care clinical cases, of which 7.26% are patients with sepsis, and 92.74% - with other diagnoses. Exploratory data analysis and data processing are performed in RStudio (R programming language), and machine learning is based on the H2O platform (www.h2o.ai). Results: Based on the processing of the large data set, an intelligent system is built, which allows the prediction of sepsis 4 hours before the onset and which can be delivered as an application for clinical use. The performance metrics are: accuracy - 0.91, specificity - 0.93 and sensitivity - 0.84.