Development and Validation of an Explainable Deep Learning Model to Predict Adverse Event During Hospital Admission in Patients with Sepsis

Abstract

Sepsis is among the most common conditions requiring emergency hospitalization. The early and accurate identification of sepsis patients with a high risk of in-hospital adverse events can aid physicians in making optimal clinical decisions. This study aimed to develop an explainable neural network model to predict adverse events during hospital admission in patients with suspected sepsis. Patient data were collected from a single medical center in Taiwan for the period of 2018–2020. The adverse events considered during hospital admission were cardiac arrest, respiratory failure requiring mechanical ventilation, and transfer to intensive care unit during admission. This study included 9398 patients in the analysis, with 6794 and 2603 patients in the development and validation sets, respectively. The proposed model could predict adverse events with an area under the receiver operating curve of 0.88 and 0.85 in the development and validation sets, respectively. Of the 2603 patients in the test set, 523 (20.1%) were classified as having adverse events during hospital admission. Of these patients, 104 eventually experienced adverse events. Thus, the model can predict adverse events with good performance and therefore, can be regarded as a gatekeeper before patients with sepsis are admitted to the general ward.