| Sumario: | Enhancing patient care quality depends on preventing health-related adverse events (HAEs), including in-hospital mortality, right from the start of hospitalization. Today, machine learning tools offer innovative solutions for predicting such events. Nevertheless, and despite their predictive potential, graph neural networks (GNNs) remain largely unexplored in the literature for analyzing complex, interconnected, real-life clinical hospital data from a clinical data warehouse (CDW). In this study, we present a robust evaluation of two GNN models which are benchmarked against strong baseline methods for predicting in-hospital death. Our findings demonstrate that these models perform as well as the best benchmarks established in previous research. Additionally, we highlight promising opportunities enabled by the graph structure of the data, including graph model optimization and enhanced prediction explainability.
|
|---|