DynEdges-TGN: Dynamic-Edges-Based Temporal Graph Network for Early Sepsis Prediction

Supriya Bajpai, Manohar Kaul

Abstract: In this paper, we introduce a novel method to improve the early prediction of sepsis for patients in intensive care units (ICUs), directly contributing to the essential aim of life saving. Physiological parameters of the body, intricately interact in a dynamic and complex web of relationships that evolve with sepsis progression, potentially offer crucial diagnostic insights in sepsis prediction. Based on this, we propose Dynamic-Edges-based Temporal Graph Network (DynEdges-TGN), a novel graph-based network designed for early sepsis prediction using irregularly sampled multi-variate time series clinical data obtained from electronic health records. DynEdges-TGN takes into account the dynamic nature of the relationship between physiological parameters during disease progression and learns time-varying inter-feature dependency graphs for each sample. By aggregating the temporal information present in these dependency graphs, DynEdges-TGN effectively captures the temporal structure present within a data sample. The model learns the graph structure with evolving correlations among features over time and, generates feature embeddings through message passing that exploit the relational structure among features. Our experimental results demonstrate that DynEdges-TGN outperforms the existing state-of-the-art method with an improvement of \(8.3\%\) in mean AUROC score and \(40.1\%\) in mean AUPRC score, showing promising potential for early sepsis detection. The ability of the proposed DynEdges-TGN to leverage the evolving interplay between features over time emphasizes its viability for improved early sepsis detection.