Advancing early detection of sepsis with physiological variable interactions and temporal contrastive learning in critical care

Sepsis presents a critical challenge in Intensive Care Units (ICUs) due to its rapid onset and complex etiology, necessitating accurate and timely diagnosis to reduce mortality. However, existing methods often fail to capture the intricate interactions among physiological variables and lack mechanisms to enhance the discovery of frequency-domain patterns, which are crucial for detecting subtle and clinically significant signs of sepsis. To address these limitations, we propose a novel sepsis prediction framework that integrates a Variable Interaction Graph Neural Network (VIGNN) with a Temporal Contrastive Loss (TCL). First, we design VIGNN to effectively model the intricate relationships among physiological variables. Second, we introduce a frequency-masking augmentation strategy that selectively focuses on important frequency components, generating augmented views to emphasize critical frequency-domain features. Finally, we develop TCL to align the representations of frequency-enhanced and original views of the same sample while distinguishing them from other samples at multiple temporal scales. This mechanism forces our model to uncover meaningful frequency-domain patterns that complement time-domain features, enabling a richer and more robust representation. Experimental results on the Beth Israel Deaconess Medical Center dataset and Emory University Hospital dataset demonstrate that our framework achieves AUROC scores of 81.17% and 84.48%, respectively. These results represent improvements of 2.49% and 2.45% over state-of-the-art methods, enabling clinicians to deliver more timely and targeted interventions. The code is publicly available at https://github.com/Hgnnhd/VIGNN-TCL-master.