面向电子病历动态建模的可解释深度学习方法研究

深度学习与海量电子病例（EHR）相结合，为医疗健康领域带来翻天覆地的变化。然而，在该领域，深度学习的不可解释性却遭遇巨大挑战。虽然可解释深度学习为应对该挑战提供了崭新而有效的方法。但是，如何发现适于EHR动态建模的可解释深度学习方法，是目前亟待解决的关键科学问题，极具挑战性。本项目拟运用表征学习、深度学习、注意力机制、领域知识图谱、多语义元路径、链接分析及医学科学理论与方法，开展面向EHR动态建模的可解释深度学习方法研究，导出相应的融合异构特征的EHR医学概念表征方法、可解释的通用EHR动态建模方法、融合多领域知识的EHR建模方法及EHR动态模型与知识图谱互惠增强方法，籍此揭示用于EHR动态建模的“精确+解释”深度学习方法的规律。本项目属于计算机科学与医学的交叉前沿方向，不仅为深度学习乃至机器学习的可解释性方法研究提供创新思路，而且拓展可解释性深度学习的应用领域，对医学现代化意义重大。

The combination of deep learning and massive Electronic Health Records(EHR) has brought tremendous changes to the medical and health domain. However, the un-interpretability of deep learning in the medical field has encountered significant challenges. Fortunately, interpretable deep learning method provides a new and effective way to handle this challenge. However, how to find an interpretable deep learning method suitable for EHR dynamic modeling is currently a critical scientific problem, which is exceptionally challenging. This project intends to use representation learning, deep learning, attention mechanism, the domain knowledge graph, multi-semantic meta-paths, link analysis, and medical science theory and technique, to research on interpretable deep learning for EHR dynamic modeling. Specifically, the study contains the EHR medical concept characterization method, interpretable EHR dynamic modeling method, EHR modeling method combining various domain knowledge, and the reciprocal enhancement method for EHR dynamic model and knowledge graph. The ultimate goal of the project is revealing the law of the "accurate + interpretable" deep learning ways for EHR dynamic modeling. The project belongs to the cross frontier direction of computer science and medicine. On the one hand, it provides innovative ideas for the research of interpretable machine learning methods. On the other hand, it expands the application field of interpretable deep learning, which is of considerable significance to medical modernization.