SCH: Tackling Progressive Disease - Learning from Longitudinal Observational Clinical Data in the Presence of Noise and Confounding

迫切需要提高我们对阿尔茨海默病（AD）患者进行分层和治疗的能力。为此，像阿尔茨海默病神经成像倡议（ADNI）这样的努力旨在收集广泛的生物标志物数据，并招募了数百名参与者进行纵向跟踪。虽然这些研究是重要和必要的，但有证据表明，阿尔茨海默病的发展早在症状出现前22年就开始了。因此，这种前瞻性研究需要一段时间才能产生足够的数据来阐明这种疾病的长期进展。该项目超越了像ADNI这样的管理数据集，并开发了新的技术，可以利用常规收集的电子健康记录（EHR）数据，对轻度认知障碍和AD诊断前后的患者轨迹进行新的分析。根据观察数据估计患者风险的工具有可能推广到阿尔茨海默病以外的其他进展缓慢的疾病。我们期望所提出的工作为临床系统奠定基础，通过识别最有可能从早期干预中受益的患者，并通过测量可改变的风险因素的影响来推荐降低风险的行动，从而直接影响社会。这项工作在开发评估AD患者风险的工具的背景下，推进了用于患者风险分层和个体治疗效果评估的机器学习（ML）领域。在风险分层方面，研究人员探索了在存在标签噪声的情况下进行学习的新方法，以及利用事件顺序约束信息（例如，死亡不能先于AD）进行多事件生存分析的新方法。此外，新的机器学习技术被开发出来，以提高我们使用观察数据估计因果关系的能力（例如，高血压如何影响一个人患AD的风险），重点是解决与混杂相关的偏差。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

There is a pressing need to improve our ability to stratify and treat patients with or at risk of developing Alzheimer’s disease (AD). To this end, efforts like the Alzheimer’s Disease Neuroimaging Initiative (ADNI) aim to collect data on a wide range of biomarkers and have enrolled hundreds of participants who are followed longitudinally. While such studies are important and necessary, there is evidence to suggest that the development of AD starts as early as 22 years prior to symptom onset. Thus, it will be some time before such prospective studies produce enough data to shed light on the long-term progression of the disease. This project moves beyond curated datasets like ADNI and develops new techniques that can leverage routinely collected electronic health record (EHR) data for novel analyses of patient trajectories prior to and following a diagnosis with mild cognitive impairment and AD. Tools for estimating patient risk from observational data have the potential to generalize beyond AD to other conditions that progress slowly over the course of years. We expect the proposed work to lay the groundwork for clinical systems that directly impact society by identifying patients most likely to benefit from early intervention and recommend actions to reduce risk through measuring the effect of modifiable risk factors.This work advances the fields of machine learning (ML) for patient risk stratification and individual treatment effect estimation in the context of developing tools for estimating patient risk for AD. In terms of risk stratification, new approaches for learning in the presence of label noise and for multi-event survival analysis that leverage information about the constraints on the ordering of events (e.g., death cannot precede AD) are explored. In addition, novel ML techniques are developed to advance our ability to estimate causal effects using observational data (e.g., how does hypertension affect one’s risk of developing AD), with a focus on addressing bias related to confounding.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.