Explainable Graph-Neural Architectures for ICU Mortality Prediction Using Logic Explained Networks

The increasing digitalisation of intensive care has led to vast repositories of patient data that remain underexploited for transparent clinical decision support. In this work, we present an explainable deep-learning framework that combines Logic Explained Networks (LENs) with Graph Neural Networks (GNNs) for predicting 28-day mortality in mechanically ventilated ICU patients. Each patient is represented as a node in a similarity graph constructed from physiological and clinical variables, enabling the GNN to capture relational dependencies between comparable cases. The LEN component augments the neural model with logic-based reasoning layers, producing human-interpretable first-order rules that approximate network predictions. Experiments on the MIMIC-II clinical dataset demonstrate that integrating relational structure improves mortality prediction accuracy (≈ 0.9) compared to traditional baselines, while LEN-derived rules achieve high fidelity (≈ 0.7) and align with known clinical risk factors such as age, organ dysfunction, and comorbidities. These results suggest that symbolic interpretability and deep relational learning can coexist in a unified framework, offering both strong predictive performance and transparent reasoning for critical care decision support.