Detection of Patients at Risk of Enterobacteriaceae Infection Using Graph Neural Networks: a Retrospective Study

While Enterobacteriaceae bacteria are commonly found in healthy human gut, their colonisation of other body parts can potentially evolve into serious infections and health threats. We aim to design a graph-based machine learning model to assess risks of inpatient colonisation by multi-drug resistant (MDR) Enterobacteriaceae. The colonisation prediction problem was defined as a binary classification task, where the goal is to predict whether a patient is colonised by MDR Enterobacteriaceae in an undesirable body part during their hospital stay. To capture topological features, interactions among patients and healthcare workers were modelled using a graph structure, where patients are described by nodes and their interactions by edges. Then, a graph neural network (GNN) model was trained to learn colonisation patterns from the patient network enriched with clinical and spatiotemporal features. The GNN model predicts colonisation risk with an AUROC of 0.93 (95% CI: 0.92-0.94), 7% above a logistic regression baseline (0.86 [0.85-0.87]). Comparing different graph topologies, the configuration that considers only in-ward edges (0.93 [0.92-0.94]) outperforms the configurations that include only out-ward edges (0.86 [0.85-0.87]) and both edges (0.90 [0.89-0.91]). For the top-3 most prevalent MDR Enterobacteriaceae, the AUROC varies from 0.92 (0.90-0.93) for Escherichia coli up to 0.95 (0.92-0.98) for Enterobacter cloacae, using the GNN - in-ward model. Topological features via graph modelling improves the performance of machine learning models for Enterobacteriaceae colonisation prediction. GNNs could be used to support infection prevention and control programmes to detect patients at risk of colonisation by MDR Enterobacteriaceae and other bacteria families.