Machine learning algorithms for early sepsis detection in the emergency department: A retrospective study

Background: Early recognition and treatment of sepsis are crucial for improving patient outcomes. However, the diagnosis of sepsis remains challenging because of vague clinical presentations. Objectives: We aim to developed novel sepsis screening tools with machine learning models and compared their performance with traditional methods. Methods: We used machine learning algorithms to develop models for early risk prediction of sepsis based on retrospective single-center electronic health record data from adult patients who presented to the emergency department (ED) from June 2018 through May 2020. Available triage data including vital signs, baseline characteristics, and chief complaints served as predictors. In our study, 80% and 20% of the data were randomly split into training and testing sets, respectively. Derived from the training set, we built the models based on four machine learning algorithms: logistic regression, gradient boosting, random forest, and neural network. Our primary outcome was the model performance that predicted the final diagnosis of sepsis determined by the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and predictive performance compared with those of the reference models (quick sequential organ failure assessment [qSOFA], modified early warning score [MEWS], and systemic inflammatory response syndrome [SIRS]) using the testing dataset. Results: In total, 133,707 ED visits were analyzed. All machine learning models outperformed the reference models by achieving a higher AUROC (e.g., AUROC 0.931 [95% CI 0.921-0.944] in our best model (random forest algorithm) vs 0.635 [95% CI 0.613-0.660] in qSOFA, 0.688 [95% CI 0.662-0.715] in MEWS, and 0.814 [95% CI 0.794-0.833] in SIRS). Conclusion: The machine learning models demonstrated superior performance in prediction of sepsis diagnosis among emergency patients compared with that using the traditional screening tools. Further studies are needed to determine whether the models will enhance physicians' judgments and improve patient outcomes.