New Framework Boosts Clinical Risk Prediction with Knowledge Graphs

Predicting clinical risk from Electronic Health Records (EHRs) is challenging due to data sparsity and the difficulty of integrating diverse external knowledge and unstructured notes. Existing methods often fail to capture the nuances of disease severity, treatment responses, and patient progression. To overcome these limitations, researchers developed TRACER, a novel trajectory-aware and clinically grounded prediction framework. TRACER operates through several key components: First, it constructs a medical knowledge graph that is enriched with severity information derived from extensive medical literature. Second, it retrieves clinically relevant, severity-weighted paths of a patient's progression from this knowledge graph. Third, it extracts crucial clinical events from unstructured clinical notes, which are often overlooked. Finally, it augments the patient's context by incorporating similar peer cases. Experiments on the MIMIC-III and MIMIC-IV datasets demonstrated substantial improvements over current state-of-the-art methods, showing up to a 28.5% increase in Macro F1 score for mortality prediction and a 19.7% increase for readmission prediction.