TSKG-ICU: temporal–semantic knowledge graph embeddings for ICU readmission prediction

Unplanned Intensive Care Unit (ICU) readmissions emerge from complex, time-dependent clinical processes that are poorly captured by existing electronic health records-based prediction models. Existing approaches capture either temporal patterns without understanding clinical meaning, or clinical relationships without preserving temporal context. This separation limits their ability to simultaneously model how clinical events relate to one another with how they evolve across an ICU stay. Bridging this gap requires models that integrate both temporal dynamics and semantic clinical knowledge within a unified framework. We introduce TSKG-ICU, a novel framework that jointly models temporal evolution and semantic relationships in ICU trajectories using temporal knowledge graph embeddings. Leveraging the MIMIC-III cohort, we constructed a temporal–semantic knowledge graph with over 11 million triples representing 60,000+ admissions and 200,000+ biomedical concepts. Clinical events are temporally encoded relative to ICU admission, preserving the full longitudinal structure of clinical records, and semantically annotated through ontology mapping. Patient trajectories are encoded as temporal knowledge graph embeddings and used as feature representations for supervised readmission prediction. TSKG-ICU achieves state-of-the-art performance (ROC-AUC: 0.786 and PR-AUC: 0.641). Ablation studies reveal that temporal modelling provides the primary predictive signal, whilst semantic enrichment offers complementary improvements, an ordering consistent across all three classifiers evaluated. Removing temporal information displaces roughly an order of magnitude more predictions than removing semantic structure, which instead refines a representation temporal encoding has already organised, confirming that unified temporal–semantic representations shape the clinical decisions a deployed model would support. Architectural evaluation demonstrates that asymmetric, complex-valued embeddings with explicit temporal decomposition outperform both simpler and more recent alternatives. TSKG-ICU establishes temporal–semantic knowledge graph embeddings as a new paradigm for ICU readmission prediction. By leveraging complete ICU timelines and grounding clinical events in a shared semantic space, this approach provides a clinically meaningful foundation for identifying high-risk patients and supporting critical care decision-making.