GDEMO-HAT: Knowledge-guided Medical Text Classification Using Heterogeneous Graph-based Dependency Modeling with Hierarchical Attention

The medical texts contain rich information about the health status of patients, including personal demographic details, examination items, diagnostic narratives, and corresponding outcomes. There is often an underlying dependency between a patient and multiple medical texts. Such dependency has a substantial influence on the accuracy and reliability of the patient's comprehensive diagnostic outcomes, thus necessitating the modeling of dependency. Most of the existing studies on medical text classification focused only on textual content and often overlooked the inherent dependency between the patient and medical texts, leading to the loss of relevant contextual information. To address the above problems, we propose an approach for medical text classification using heterogeneous Graph-based DEpendency MOdeling with Hierarchical ATtention (GDEMO-HAT), which accounts for patient-notes dependency and retains the contextual information. We model the dependencies between the patients and their medical texts via heterogeneous graphs. To facilitate medical text mining, we incorporate domain knowledge into the construction of a Patient-Notes-Physician (PNP) graph. Based on the constructed PNP graph, we introduce the hierarchical attention mechanism for classification that assigns different levels of importance to nodes based on distinct semantic contexts and thereby effectively captures the intricate interdependencies among these nodes. The representations of dependencies are integrated at the semantic level, thereby enabling a more comprehensive understanding of textual data. To demonstrate the effectiveness of the proposed approach, we conduct a classification task of the diagnosis of cardiovascular and cerebrovascular diseases based on medical text. The classification accuracy rate achieves 94%, which is significantly higher than the performances of the other compared models, such as Bert.