Developing novel time-varying phenotype representations for clinical systems

表征学习推动了机器学习领域的一些重大进展；例如，卷积网络（cnn）学习的成像数据集的低维、定性有意义的表示。临床数据缺乏这种自然表征，这是限制相关机器学习进展的一个因素。这在决策领域尤其成问题，因为对于一个特定的病人来说，什么是最好的存在很大的不确定性。大多数干预措施都有潜在的有害副作用，即使是完善的重症监护病房也与随机对照结果的改善有关。因此，开发表征或健康表型是患者状态的稳健表征是至关重要的。我们的目标是发展健康表型，并率先使用它们来理解健康意味着什么，识别相似的患者，改善医疗保健服务，并量化可能的干预措施的影响。为了开发健康表型，我们需要整合数据，包括自由文本注释、编码程序和诊断、血压等生命体征的时间序列，以及ECG等高频信号。对于任何给定的病人，不同的数据可能对特定的任务很重要。虽然之前的工作针对特定任务（例如，纵向结果或诊断代码的预测），但我们计划将重点放在集成不同数据类型的表示上，包括编码账单数据（稀疏二进制张量）、生命体征和实验室（密集不规则采样时间序列数据）、临床叙述（可变长度文本向量）和基因组学数据（密集二进制矩阵）。学习的表型必须是可靠和稳健的，因为在临床实践中应用机器学习可能会产生许多潜在的意想不到的后果。将这些数据整合到表型潜在状态估计中需要跨时间尺度和粒度进行映射。这项拟议的研究将对我们理解健康的含义以及人类表型的差异产生重大影响。随着在实际环境中使用机器学习系统的意图，表示学习在其他领域的重要性也在增加。我们对异质性数据中有意义的表型发现的关注在许多其他应用领域也很强大，合作者期望在机器学习方法中证明学习表征的鲁棒性。

Representation learning has prompted some of the greatest advances seen in machine learning; for example, the lower dimensional, qualitatively meaningful representations of imaging datasets learned by convolutional networks (CNNs). Clinical data lacks such natural representations, which is a limiting factor to the progress of relevant machine learning. This is especially problematic in the realm of decision making, where there is much uncertainty about what is best for a particular patient. Most interventions have potentially harmful side-effects, and even well-established intensive care unit are associated with improved outcomes in randomized controlled. It is therefore critical to develop representations or health phenotypes that are robust representations of patient state. Our goal is to develop health phenotypes, and pioneer their use to understand what it means to be healthy, identify similar patients, improve healthcare delivery, and quantify the impact of possible interventions. In order to develop health phenotypes, we need to integrate data that span free text notes, coded procedures and diagnoses, timeseries of vitals like blood pressure, and high-frequency signals like ECG. For any given patient, different data may be important for a specific task. While prior work has targeted specific tasks (e.g., the prediction of longitudinal outcomes or diagnostic codes, we plan to focus instead on representations that integrate distinct data types, including coded billing data (sparse binary tensors), vitals and labs (dense irregularly sampled time series data), clinical narratives (variable length text vectors), and genomics data (dense binary matrices). Learned phenotypes must be reliable and robust as there are many potential unintended consequences that may result from the application machine learning in clinical practice. Incorporating these data into phenotypic latent state estimates requires mapping across time-scales and granularity. The proposed research will have a strong impact on our understanding of what it means to be healthy, and variance within the human phenotype. The importance of representation learning in other fields has increased along with the intent to use the machine learning systems in a practical setting. Our focus on meaningful phenotypes discovery in heterogeneous data is also powerful in many other application areas where justifications about the robustness of a learned representation in machine learning methods are expected by collaborators.