Artificial Intelligence for Predicting Mortality Due to Sepsis

Objective: Sepsis is a life-threatening organ dysfunction caused by a systemic host response to infection, and lead to multi organ failure and septic shock. The incidence of sepsis and sepsis-related death were reported approximately 48.9 million and 11.0 million in the worldwide. Specially, the mortality of the patients who were diagnosed severe sepsis and septic shock was 30 - 50%, and its can cause a tremendous loss for medical resource and socioeconomic cost. Although various clinical indicators were developed for the predicting mortality, these indicators need to apply plenty of the features, or confusing due to ambiguous criteria. To solve this topic, we focused on a deep learning (DL) model predicting mortality due to sepsis. Methods: We developed a prediction model employing the multilayer perceptron (MLP), and built the dataset with 577 survival patients and 208 mortality patients to train and evaluate. Results: To evaluate the proposed model performance, we applied 5 metrics such as the accuracy, sensitivity, specificity, F1-score and area under the receiver operating characteristic (AUROC) curve. The proposed model predicted mortality due to sepsis in 164 out of 208 mortality group (sensitivity: 73.81%) and survival in 438 out of 577 survival group (specificity: 74.14%). The F1-score was 60.19% and the AUROC was 0.835. Through the comparison with the clinical indicator and machine learning models, the proposed model showed outperforming. Conclusion: Although the MLP model is not a novel technique in the artificial intelligence field, we confirmed a possibility of the DL by this study that is first step for clinicians and patients.