Artificial Intelligence for Sepsis Prediction in ICU

Overview

The development of sepsis prediction model in line with Chinese population, and extended to clinical, assist clinicians for early identification, early intervention, has a good application prospect. This study is a prospective observational study, mainly to evaluate the accuracy of the previously established sepsis prediction model. The occurrence of sepsis was determined by doctors' daily clinical judgment, and the results of the sepsis prediction model were matched and corrected to improve the clinical accuracy and applicability of the sepsis prediction model.

The sepsis prediction model adopted in this study has been completed in the preliminary preparation, which was constructed on 7,000 patients since the establishment of comprehensive ICU, and the sepsis 3.0 diagnostic standard was adopted.The sepsis prediction model was built using Python platform and XGBoost algorithm, which was used to predict the incidence of sepsis in ICU patients within 24 hours. The overall accuracy was 82%, and the area under the Auroc curve was 0.854. Patients who met the inclusion and exclusion criteria were given a daily prediction of sepsis model, and a quantitative checklist was formed based on the test results.There are two kinds of forecast outcomes: low risk and high risk.Quantitative checklists are available to attending physicians to improve diagnostic efficiency.The results were kept confidential to the clinician. All patients were diagnosed with sepsis by two senior attending physicians at a fixed time. The diagnosis consisted of two types: yes and no.If two attending physicians have different opinions, the third attending physician will be included for correction diagnosis, and the presence of sepsis will be determined in a 2:1 manner.The attending physicians are independent of each other. When the diagnosis results of the attending physician are input into the system, the prediction results of yesterday's sepsis prediction model are compared and calculated to determine the accuracy of the prediction model

Conditions