An Explainable AI based Clinical Assistance Model for Identifying Patients with the Onset of Sepsis

The high mortality rate of sepsis, especially in Intensive Care Unit (ICU) makes it third-highest mortality disease globally. The treatment of sepsis is also time consuming and depends on multi-parametric tests, hence early identification of patients with sepsis becomes crucial. The recent rise in the development of Artificial Intelligence (AI) based models, especially in early prediction of sepsis, have improved the patient outcome. However, drawbacks like low sensitivity, use of excess features that leads to overfitting, and lack of interpretability limit their ability to be used in a clinical setting. So, in this research we have developed a smart, explainable and a highly accurate AI based model (called XAutoNet) that provides quick and early prediction of sepsis with a minimal number of features as input. An application based novel convolutional neural network (CNN) based autoencoder is also implemented that improves the performance of XAutoNet by dimensional reduction. Finally, to unbox the "Black Box" nature of these models, Gradient based Class Activation Map (GradCAM) and SHapley Additive exPlanations (SHAP) are implemented to provide interpretability of autoencoder and XAutoNet in the form of visualization graphs to assist clinicians in diagnosis and treatment.