Explainable Graph Neural Networks in Intensive Care Unit Mortality Prediction: Edge-Level and Motif-Level Analysis
ACTA INFOLOGICA, vol.9, no.2, pp.770-789, 2025 (ESCI, TRDizin)
- Publication Type: Article / Article
- Volume: 9 Issue: 2
- Publication Date: 2025
- Doi Number: 10.26650/acin.1835775
- Journal Name: ACTA INFOLOGICA
- Journal Indexes: Emerging Sources Citation Index (ESCI), TR DİZİN (ULAKBİM)
- Page Numbers: pp.770-789
- Istanbul University Affiliated: No
Abstract
Accurate forecasting of ICU patient outcomes is essential for clinical decision support. However, most high-performing machine learning models function as black boxes, limiting their interpretability and clinical adoption. This study introduces a graph-based explainable risk-prediction framework, in which patient-patient relations are modeled through a diagnosis-based similarity network. An undirected graph was derived from the eICU-CRD demo subset (PhysioNet v2.0) by linking individuals sharing three-digit ICD-9 categories, and a GCN was trained for in-hospital mortality prediction. Despite the dataset's modest size and imbalance, meaningful discrimination was achieved (AUROC = 0.708; AUPRC = 0.308). A two-layer explainabilityanalysiswas applied to clarify the model's decision process. Each prediction was driven bya combination of patient-specific clinical attributes and signals from a small number of influential neighbors, according to GNNExplainer. SubgraphX, a Shapley-value-based motif discovery method, identified compact and clinically coherent subgraphs with strong causal influence on the prediction. Consistency between edge-and motif-level explanations indicated that the model relies on stable relational patterns with clinical relevance. These findings suggest that integrating GNNs with structured explainability methods can transform a single risk score into a transparent, data-driven decision-support mechanism that provides interpretable and hypothesis-generating insights to clinicians.