Enhancing sepsis prediction with transformer-based feature extraction, PCA, and TabNet: a transparent and interpretable framework

Sepsis, a critical and potentially fatal condition arising from the body’s uncontrolled immune response to infection, remains a major challenge in global healthcare. Despite ongoing advancements in medical diagnostics, early sepsis detection continues to be problematic due to its rapid progression, complex clinical variability, and the limitations inherent in traditional diagnostic tools such as SOFA and qSOFA. These conventional methods often fail to identify subtle physiological changes during the early stages of sepsis. This study introduces an innovative framework to enhance sepsis prediction in intensive care settings by combining Transformer-based feature extraction, Principal Component Analysis (PCA), and the TabNet architecture. The framework integrates dimensionality reduction techniques, self-attention mechanisms for analyzing feature interactions, and an interpretable classification model for accurate predictions. To ensure transparency in clinical decision-making, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) are employed. The proposed model demonstrates robust performance with an accuracy of 78% and a ROC-AUC score of 0.855, providing a reliable and interpretable solution for real-time diagnostic needs in critical care environments.