EAGER: Patient Similarity Learning with Massive Clinical Data and Its Applications in Cohort Identification

电子健康记录（EHRs）在美国医疗保健系统中的迅速采用，加上将电子健康记录与研究生物储存库联系起来的能力，为开展大规模精准医学研究提供了独特的机会。使此类研究成为可能的一个关键步骤是通过定义纳入和排除标准来确定队列，该标准根据可用的临床数据通过算法选择患者组。对于大多数现有的研究，生成这些患者队列的标准是手动定义的，这使得整个过程缓慢，劳动密集型且不可扩展。该项目开发患者相似学习算法，以实现自动队列识别，这将加速精准医学的研究。患者周围的大量临床数据具有高度的异质性和稀疏性。虽然有一些患者相似学习算法，但它们通常只处理单一类型的患者数据（例如，仅使用患者EHR中的诊断信息），不能有效地处理上述挑战。该项目开发了先进的患者相似度学习算法：1)通过从多个基础相似度度量的改进过程来学习复合患者相似度，每个基础相似度从特定的患者数据来源或特定形式的患者表示进行评估；2)整合来自多个相关辅助领域的信息，如药物、疾病和基因组信息。这些信息有效地规范了患者相似性学习过程，使其对数据稀疏性不那么敏感。

The rapid adoption of Electronic Health Records (EHRs) across the U.S. healthcare systems coupled with the capability of linking EHRs to research biorepositories provides a unique opportunity for conducting large-scale Precision Medicine research. A critical step to make such research possible is identification of cohorts by defining inclusion and exclusion criteria that algorithmically select sets of patients based on available clinical data. For most of the existing research, the criteria for generating those patient cohorts are defined manually, which makes the entire process slow, labor intensive and not scalable. This project develops patient similarity learning algorithms to enable automatic cohort identification, which will accelerate the research of precision medicine.The massive clinical data around patients are highly heterogeneous and sparse. Although there are some patient similarity learning algorithms, they typically work with a single type of patient data (e.g., just using diagnosis information in patient EHR) and cannot handle those challenges mentioned above effectively. This project develops advanced patient similarity learning algorithms by 1) learning composite patient similarities through a refinement process from multiple base similarity measures, with each base similarity being evaluated from a specific source of patient data or a specific form of patient representation; and 2) integrating information from multiple related auxiliary domains, such as drug, disease, and genomic information. Those information effectively regularizes the patient similarity learning process and makes it less sensitive to data sparsity.