InKrat: Interpretable diagnosis prediction models based on cross-modal knowledge graph semantic retrieval fusion

Although deep learning models have made breakthrough achievements in many fields, they still face some challenges in diagnostic prediction tasks in healthcare. Existing methods either use graph structures or sequence structures one-sidedly or disjointedly, failing to obtain high-quality representations of EMR data. Some knowledge-enhanced methods rely on strategies based on name or identifier matching, lacking flexibility while introducing semantically mismatched noise. On the other hand, attention-based models for interpretable analysis can only provide the importance of different factors rather than intuitive and easily understandable natural language descriptions. To address the above issues, we propose InKrat, a new KG-enhanced method. Specifically, we designed a novel temporal graph structure that models the structure and temporal information in EMR by integrating anchor nodes as a bridge. We also developed a cross-modal semantic retrieval method, utilizing a large language model (LLM) to compute the semantic similarity between the KG and medical notes, filtering the knowledge accordingly. Finally, based on the knowledge prompts, the LLM generates interpretable descriptions of the prediction results. We have extensively validated the effectiveness of InKrat through experiments on two commonly used real-world datasets. The results demonstrate that our proposed method achieves state-of-the-art performance. Our code can be found at https://github.com/lzh-nwpu/InKrat.