Interpretable Disease Prediction via Path Reasoning over medical knowledge graphs and admission history

Disease prediction based on patients' historical admission records is an essential task in the medical field, but current predictive models often lack interpretability, which is a critical aspect in clinical practice. In this paper, we propose a Knowledge Guided Interpretable Disease Prediction method (KGxDP) via Path Reasoning over Medical Knowledge Graphs and Admission History. In KGxDP, the representation of a patient is formulated via a personalized medical knowledge graph, which is then combined with the patient's admission sequence embedding to form an inclusive subgraph. This admission sequence embedding is modeled by a Transformer based on the patient's admission history, capturing the time-based variations of each diagnosis. Furthermore, the subgraph is updated via graph reasoning by using a node-type and edge-type specified Graph Attention Network (GAT) and subsequently combined with admission sequence embedding for disease prediction. This process also facilitates interpretability by extracting critical paths within the subgraphs. Empirical evaluations on public MIMIC-III, MIMIC-IV and eICU datasets demonstrate that KGxDP outperforms state-of-arts models in predicting patients' future diseases while also providing convincing explanations. The extracted paths are used as prompts for ChatGPT to generates user friendly, understandable Natural Language Explanations (NLE) for the prediction results, which also shows that the extracted paths by KGxDP have strong interpretability. This augmentation in predictive accuracy and explanation reliability holds significant potential to positively impact clinical decision-making.

Automated summary

This paper proposes a Knowledge Guided Interpretable Disease Prediction method that combines medical knowledge graphs and admission history to predict diseases. By extracting critical paths and using admission sequence embedding, the model achieves better predictive performance and interpretability.