Modeling Clinical Thinking Based on Knowledge Hypergraph Attention Network and Prompt Learning for Disease Prediction

Abstract

Disease prediction requires a deep understanding of both clinical thinking and medical knowledge. We propose a framework for the deep and dynamic interaction of empirical and basic knowledge to simulate the interaction between intuitive and analytical reasoning of dual-process theory. Specifically, we construct an interactive logical hypergraph by integrating empirical knowledge with electronic medical record text as a special node into a sub-logical hypergraph of extracting k-hop binary and n-ary logical knowledge from basic knowledge. All the nodes and relations are represented and updated through the logical hypergraph attention network, which is capable of measuring the uncertainty and importance of symbolic-based basic knowledge. Then, we extract important basic knowledge and integrate it into intuitive reasoning dominated by empirical knowledge through deep interactive textual representations based on knowledge-guided prompt learning. Experiments demonstrate modeling clinical thinking can effectively improve disease prediction accuracy and interpretability in line with expert diagnostic thinking.

Keywords: clinical thinking, dual-process theory, directed hypergraph attention network, logical knowledge, Representation Learning, prompt learning