MATCHES: Making Telehealth Delivery of Cancer Care at Home Effective and Safe - Addressing missing data in the MATCHES study to improve ML/AI readiness

项目总结： Matches(使远程医疗在家中提供有效和安全的癌症护理)远程医疗研究 该中心旨在建立必要的证据基础，以建立远程保健癌症护理的最佳实践。 先前的工作表明，专注于肿瘤学的远程医疗可以取得良好的结果，但大规模试验 一直局限于特定的情况，如姑息治疗或生存。采用一直受到以下因素的限制 有限制的报销。火柴中心将通过执行预期执行来帮助弥补这一证据差距 试验和进行观察性分析。数据将从远程医疗平台的多层集成， 患者门户、移动跟踪设备和电子健康记录(EHR)。这将有助于开发一种新的 肿瘤学中的范例-精确的护理提供-最终目标是将个体患者与 在适当的时间提供基于诊所或远程医疗支持的居家护理的最有益组合- 所有这些都是基于动态可用数据的总和。这将通过应用数据科学来实现 方法--包括灵活的试验设计和机器学习--在远程医疗中的应用有限。 在匹配精选数据集中观察到丢失的数据，这也是 电子病历和患者报告的健康数据。由于存在缺失数据，匹配数据不会 为机器学习或人工智能应用程序做好准备，因为对丢失数据的不当处理可能 这既会导致偏差，也会导致统计能力的丧失。偏见尤其令人担忧的是，如果一小部分患者 可能有缺失的数据。例如，如果低收入患者更有可能跳过自我报告的结果 由于担心引发昂贵的工作，他们的经验将在数据和分析中被低估， 损害了结论的稳健性和普适性。这些问题在 统计文献和各种各样的工具已经被开发出来，以将丢失的数据归因于似是而非的值 从概率模型中获得，并执行识别某些数据点被归因于的分析。 但是，许多补偿方法不能扩展到匹配数据中的维度，而且可能也不能 对不同的丢失数据机制保持健壮。此外，没有关于如何检查失踪人员的指导 系统的数据模式，特别是在高维特征空间中，如匹配。因此，在这个 作为补充，我们提出并开发了基于机器学习的方法，这些方法将能够处理高- 维度特征矩阵，错位的复杂模式，以及更一般的错位机制。我们会 然后应用这些方法来检查复杂的缺失数据模式，并提供 准备好ML/AL应用程序，供Matches计划的研究人员使用并与其他人共享 远程医疗卓越研究中心(TRACE)。我们还将提供分析管道，将 帮助适当处理其他大型多医疗保健数据集中的缺失数据。

Project Summary: The MATCHES (Making Telehealth Delivery of Cancer Care at Home Effective and Safe) Telehealth Research Center aims to build the evidence base necessary to establish best practices for telehealth-enabled cancer care. Prior work demonstrates that oncology-focused telehealth can achieve favorable outcomes, but large-scale trials have been limited to specific contexts like palliative care or survivorship. Adoption has been constrained by restricted reimbursement. The MATCHES Center will help remediate this evidence gap by executing prospective trials and conducting observational analyses. Data will be integrated from multi-layers from telehealth platforms, patient portals, mobile tracking devices, and the electronic health record (EHR). This will help develop a new paradigm in oncology—precision care delivery—with the ultimate goal of matching individual patients with the most beneficial combination of clinic-based or telehealth-supported home-setting care at the appropriate time— all based on the totality of dynamically available data. This will be accomplished by applying data science methods—including nimble trial designs and machine learning—that have had limited application to telehealth. Missing data have been observed in the MATCHES curated data sets, which is also a common issue of both EHR and patient-reported health data. Due to the presence of missing data, the MATCHES data is not ready for machine learning or artificial intelligence applications as inappropriate handling of missing data can lead to both bias and loss of statistical power. Bias is particularly concerning if a subgroup of patients is more likely to have missing data. For example, if low-income patients are more likely to skip self-reported outcomes for fear of triggering costly work-up, their experience will be underrepresented in the data and analysis, compromising the robustness and generalizability of conclusions. These issues are well-recognized in the statistical literature and a wide array of tools have been developed to impute missing data with plausible values obtained from a probabilistic model and perform analyses recognizing that some data points are imputed. However, many imputation methods do not scale up to the dimensions in the MATCHES data, and they may not be robust to differentmissing data mechanisms. Additionally, there is no guidance on how to examine the missing data patterns systematically, especially in the high-dimensional feature space as in MATCHES. Hence in this supplement, we propose and develop machine-learning-based approaches that will be able to handle a high- dimensional feature matrix, complex patterns of missingness, and more general missing mechanisms. We will then apply these methods to examine the complex missing data patterns and provide imputed data sets that are ready for ML/AL applications both for the researchers of the MATCHES program and to be shared with others across the Telehealth Research Centers of Excellence (TRACE). We will also provide analysis pipelines that will help appropriately handle missing data in other large-scale multi-modality healthcare data sets.