QuBBD: Collaborative Research: Interactive Ensemble clustering for mixed data with application to mood disorders

大数据时代产生了前所未有的数据规模和复杂性。然而，由于传统的数据处理和分析方法经常失败或不合适，充分利用大数据资源来获取知识和发现是一个公开的挑战。本项目开发了一种创新的方法，利用大数据来改进情绪障碍的分类，以提高精神病患者的诊断和治疗效果。由于难以捉摸的本质，大数据问题对精神障碍来说本质上更严重。精神学界已经认识到，迫切需要一种更精确的、以证据为基础的方法来诊断和治疗疾病。事实上，最近由美国国家心理健康研究所（NIMH）资助的研究发现，精神病学干预对出现急性发作的患者有效的比例不到25%。鉴于精神障碍的普遍存在，这种低有效率尤其成问题。在美国，每5个成年人中就有1人在一生中的某个阶段经历过情绪障碍（例如抑郁症）。这个项目的动机是这样一个假设，即通过开发新的聚类方法来识别这些大型人群数据集的临床重要结构，可以获得更精确和个性化的精神健康疾病分类。然而，这种方法必须克服问题的复杂性和大规模现实世界电子健康数据的性质所带来的大量方法挑战。这些挑战包括复杂和未知的结构、高维、异质性、变量的复杂混合、缺失数据和稀疏性等。该奖项支持发起一项合作研究项目，由具有跨学科和互补技能的团队开展，以开发大数据方法，解决这类生物医学数据集成中固有的挑战。该团队的专业知识涵盖生物医学信息学、生物统计学、计算机和信息科学、电子和计算机工程、数学和精神病学等领域。在一个灵活和完全综合的框架内开发了一种新的方法，可扩展到其他生物医学数据和疾病。在此框架内，捕获数据中相关性的不同方面的聚类方法以严格的方式集成，不仅考虑了模型的不确定性，而且还产生了交互式可视化，对非专家来说具有很强的模型可解释性。具体来说，该方法将依赖于对泛化误差的自举估计器的新修改，以便在基于拓扑和机器学习方法推断的聚类集合上形成共识。该框架还支持基于用户输入（通过可视化）的共识解决方案的迭代细化，以合并领域专业知识。在异质人群中严格识别个体亚群将有助于对情绪障碍进行准确和有针对性的诊断，并为个性化的循证干预提供机会。应用程序集中于从双相情感障碍研究网络（BDRN）收集的数据中聚集患有情绪障碍（双相情感障碍和重度抑郁症）的个体。尽管如此，该方法仍可推广到在诊断和治疗方面面临类似挑战的其他疾病。事实上，该项目支持将这些方法推广到更复杂和较少管理的数据（如电子健康记录、社交媒体和其他来源）的长期愿景的第一步。该奖项由美国国立卫生研究院大数据到知识（BD2K）计划与美国国家科学基金会数学科学部合作支持。

The Big Data era has given rise to data of unprecedented size and complexity. However, fully leveraging Big Data resources for knowledge and discovery is an open challenge due to the fact that conventional methods of data processing and analysis often fail or are inappropriate. This project develops an innovative approach that utilizes Big Data to improve the classification of mood disorders for the purpose of improving diagnosis and outcomes for psychiatric patients. Big Data issues are inherently more severe for mental disorders because of their elusive nature. The psychiatric community has recognized the critical need for a more precise, evidence-based approach for the diagnosis and treatment of disease. In fact, recent studies funded by the National Institute of Mental Health (NIMH) have found that psychiatric interventions were effective in less than 25% of patients presenting with an acute episode. This low efficacy rate is especially problematic given the prevalence of mental disorders. Mood disorders alone (e.g., depression) will be experienced by 1 in 5 adults in the United States at some point in their lives. This project is motivated by the hypothesis that a more precise and personalized classification of mental health disease can be obtained through the development of novel clustering methods that identify clinically significant structures with these large population data sets. However, such an approach must overcome a large number of methodological challenges introduced by the complexity of the problem and the nature of large-scale real-world electronic health data. These challenges include, among others, complex and unknown structure, high dimensionality, heterogeneity, complex mixtures of variables, missing data, and sparsity. This award supports initiation of a collaborative research project, carried out by a team with interdisciplinary and complimentary skill sets, to develop methods for big data that address challenges inherent in the integration of biomedical data of this type. Collective expertise of the team spans the areas of biomedical informatics, biostatistics, computer and information science, electrical and computer engineering, mathematics, and psychiatry. A novel methodology is developed in a flexible and fully integrated framework that can be extended to other biomedical data and diseases. Within this framework, clustering methods that capture different aspects of relatedness in the data are integrated in a rigorous way that not only accounts for model uncertainty, but also results in an interactive visualization that is accessible with strong model interpretability for the non-expert. Specifically, the methodology will rely on novel modifications to bootstrap estimators of generalization error for the purpose of assembling a consensus over an ensemble of clusters inferred from topology-based and machine learning approaches. The framework also supports iterative refinement of the consensus solution based on user input (via the visualization) to incorporate domain expertise. The rigorous identification of sub-groups of individuals within heterogeneous populations will facilitate accurate and targeted diagnosis for mood disorders, and provide opportunity for personalized evidence-based interventions. Applications focus on clustering individuals with mood disorders (bipolar disorder and major depression) from data collected in the Bipolar Disorder Research Network (BDRN). Despite this focus, the methodology is generalizable to other diseases that face similar challenges for diagnosis and treatment. In fact, this project supports the first steps of a long-term vision of generalizing the methods to more complex and less curated data, such as electronic health records, social media, and other sources. This award is supported by the National Institutes of Health Big Data to Knowledge (BD2K) Initiative in partnership with the National Science Foundation Division of Mathematical Sciences.