Early Sepsis Prediction Using TabNet: A Deep Learning Approach with Sequential Attention

Abstract

Sepsis continues to be a major cause of death in ICUs, with 48.9 million cases and about 11 million deaths being registered each year. However, while early detection is crucial, methods like SIRS criteria and SOFA score always seem to fail to identify sepsis before severe organ dysfunction sets in. This research suggests a deep learning TabNet model for early sepsis prediction, which uses a dataset of ICU data of 1.55 million rows collected for PhysioNet/CinC Challenge 2019. The proposed model reaches the accuracy of 84.46% and AUC-ROC of 0.923, outperforming such well-known models as LSTM (AUC 0.902), XGBoost (AUC 0.895), Random Forest (AUC 0.883) and Logistic Regression (AUC 0.845). The comparison was made under the same experimental conditions (dataset, split, SMOTE data preprocessing). TabNet, as opposed to other models, utilizes a sequential attention mechanism to select important clinical features at each decision point and, thus, provides model-level explainability. Lactate, white blood cell count, heart rate, and mean arterial pressure are found as key predictors. The proposed model resulted in a mean AUC of 0.908 and standard deviation of 0.009.