Multi-level dynamic fusion and temporal role-aware network for diagnosis prediction

Disease association modeling and dynamic evolution analysis in electronic health records (EHR) is a core challenge in diagnosis prediction. Existing approaches typically treat diagnostic codes as independent entities, ignore the clinical significance of disease combinations or using only clinical monitoring indicators, which make it difficult to capture potential trends in disease progression. To this end, this paper proposes the Dynamic Fusion and Role-aware Network(DFR-Net). Specifically, we construct a feature-disease similarity matrix to fuses diagnostic codes with clinical monitoring indicators to capture potential associations between clinical monitoring indicators and disease progression, thereby providing a decision basis for disease progression. The DFR-Net further defines the dynamic roles of persistent, evolving, and episodic diseases, and uses matrixed gating units to model the stage-shifting patterns of the disease states, in order to analyze the key mechanisms in the evolution of the disease, and adaptively fuses diagnostic dependencies with clinical feature relevance to capture disease progression dynamics. This leads to improved performance on the diagnosis prediction task. Experiments on the MIMIC-III dataset demonstrate that the proposed model outperforms cutting-edge models for diagnosis prediction.