Medical Heterogeneous Graph Transformer for Disease Diagnosis

construction of medical heterogeneous map for disease diagnosis using electronic medical records is a research hotspot of medical artificial intelligence. However, existing disease diagnosis networks based on message passing mechanisms have certain limitations. For instance, these models exhibit limited expressiveness and suffer from issues such as over-compression and oversmoothing, which subsequently affect the accuracy of disease diagnosis. To address these issues, a disease diagnosis framework named Trans4DD is proposed, based on the medical heterogeneous graph Transformer. In Trans4DD's medical heterogeneous graph encoder, edge type embeddings and residual connections are introduced. Edge type embeddings effectively capture the node structure and heterogeneous information in the graph. Residual connections aid in avoiding oversmoothing and gradient vanishing problems. A node-level graph Transformer is adopted to overcome the limitations of the message passing mechanism. By employing a multi-hop node context sampling strategy, a broader range of global attention mechanisms is introduced to obtain more accurate patient representations. Experimental results on the MIMIC-IV dataset demonstrate that Trans4DD outperforms other baseline methods in terms of disease diagnosis performance, effectively enhancing the accuracy of disease diagnosis.