Multi-Length Meta-Path Sematic Fusion in Medical Heterogeneous Graph for Disease Dignosis

Heterogeneous graph neural networks have attracted significant attention in the field of disease diagnosis. Medical heterogeneous graphs encompass various types of nodes and edges, representing rich medical information and interconnections. However, there are limitations in applying inherited attention and multi-layer structures from graph neural networks to disease diagnosis tasks. Firstly, introducing attention to large medical heterogeneous graphs leads to significant computational complexity. Secondly, employing multi-layer structures when dealing with large medical heterogeneous graphs, with each layer performing semantic fusion, may cause semantic confusion and easily lead to issues such as vanishing or exploding gradients. To address these issues, a multi-length meta-path sematic fusion in medical heterogeneous graph for disease dignosis (MLM4DD) has been proposed. MLM4DD uses a lightweight average aggregator to precompute neighborhood aggregation, reducing computational complexity and improving information propagation efficiency. To better utilize semantic information and avoid issues like vanishing and exploding gradients, MLM4DD introduces a single-layer structure with multi-length meta-paths to expand the receptive field. It incorporates local attention and multi-scale attention fusion to capture features from different meta-paths, thus obtaining embedded representations of patient nodes. Extensive experiments on the MIMIC-IV dataset demonstrate that MLM4DD outperforms other baseline methods in terms of disease diagnostic performance, effectively improving the accuracy of disease diagnosis.