Unleashing the power of explainable AI: sepsis sentinel's clinical assistant for early sepsis identification

Sepsis is a severe and potentially life-threatening condition that occurs when the body's immune response becomes excessively intense in reaction to an infection. If not promptly treated, it can result in organ failure and even death. So, early identification of patients at risk for sepsis is crucial to improve the patient’s outcome in critical care. The main objective of this work is to create a highly accurate model named XAutoNet that utilizes optimal number of clinical features to predict sepsis 6 h before its onset, also providing diagnostic map behind its prediction that will help health workers in better treatment. The importance of this work is heightened in resource-scarce settings, where not all tests are available, or the turnaround time is excessive. A novel convolutional neural network based autoencoder architecture is also implemented to augment the performance of XAutoNet by reducing the input dimensions into an optimal number of dimensions. For explaining the participation of features in feature extraction, Gradient-based Class Activation Map is used to visualize the gradients in individual layers of the encoder block via heatmaps. For the explainability of XAutoNet, a visualization tool named SHapley Additive exPlanations (is used to interpret the features’ contribution in the model’s global and local prediction. The proposed XAutoNet model has an accuracy of 93%, Precision of 90%, F1 score of 92%, and Recall of 94%. The performance of the convolutional neural network -based autoencoder was also compared with its other variants, including Principal Component Analysis, which showed its high feature extraction power. The XAutoNet has also outperformed other comparable method by a significant margin. The performance of XAutoNet is instrumental in predicting sepsis in advance by understanding the non-linearity and complexity of the data of Intensive Care Unit patients with the help of the proposed autoencoder.