EAGER: Data-Driven Learning and Decision Making in Healthcare

这一探索性研究早期拨款（EAGER）奖的目标是通过不断增长的数据可用性，在医疗保健系统中创建新一代学习和决策工具。将数据驱动的学习和决策作为医疗保健的一个组成部分，在提高医疗保健质量和降低医疗保健成本方面具有巨大的潜力。尽管如此，最先进的方法仍然不足以获得广泛接受，在医疗环境中缺乏循证决策的重大影响。该项目旨在研究的基本方法和科学挑战在于多学科的交叉点：决策过程、机器学习和统计学。该项目有助于推进这一多学科研究领域，并开发教学单位，在这一高影响力领域培养一批新的研究人员。另一方面，该项目有可能影响医疗保健以外的其他领域。特别是，强调动态学习和从大量数据中做出最佳决策，使其发现与许多领域相关，如金融、市场营销和电子商务。由于技术的进步和政府的激励措施，大量关于患者病情的数据正以电子方式提供。另一方面，机器学习和统计学的进步使“预测系统”的设计成为可能：算法可以通过筛选大量患者资料来学习可推广的模式，并对临床不良事件、治疗结果或医疗服务需求提供准确的未来预测。这些预测可以在捕获新数据时实时生成，并可以帮助指导医疗保健系统的决策。然而，当统计学习模型被用于指导医疗等决策时，临床医生根据他们的预测做出决策，这些预测可以改变随后收集的患者数据，并用于“重新训练”预测系统，从而在出现新证据时更新预测概率。目前在实践中的预测系统属于以下两类之一：(1)它们在首次安装后从未接受过重新培训；或者(2)随着新数据的到来，它们会定期重新训练。然而，当新数据到来，决策周围的环境发生变化时，第一种方法会导致糟糕的预测。另一方面，多武装强盗理论、强化学习及其在数字广告和推荐系统中的应用的最新进展表明，由于决策过程的内质性，第二种方法也可能导致低质量的预测。本项目开发的新的动态学习和决策工具将通过对数据元素、决策以及决策对新数据的影响之间的相互作用进行适当建模，从而强大地应对这些挑战。

The goal of this EArly-Grant for Exploratory Research (EAGER) award is to create a new generation of learning and decision making tools in healthcare systems enabled by growing availability of data. Making data-driven learning and decision making an integral part of healthcare has profound potential to both improve the quality of medical care and to reduce healthcare costs. Despite this, state-of-the art methods are still insufficient for achieving broad acceptance, and significant impact of evidence-based decision making in medical settings is lacking. The fundamental methodological and scientific challenges that this project aims to investigate lie at the intersection of multiple disciplines: decision processes, machine learning, and statistics. This project helps advance this multidisciplinary research area and also develops teaching units for training a new cohort of researchers in this high impact space. On the other hand, this project has the potential to impact other domains beyond healthcare. In particular, the emphasis on dynamic learning and optimal decisions from large amount of data make its findings relevant to a number of domains such as finance, marketing, and electronic commerce. Due to advances in technology and government incentives a large amount of data on patients' conditions is becoming available electronically. On the other hand advances in machine learning and statistics allow the design of "predictive systems": algorithms that can learn generalizable patterns by sifting through a large number of patient profiles and provide accurate future forecasts about clinical adverse events, treatment outcomes, or demand for healthcare services. These predictions can be produced in real-time as new data is captured and can help guide decisions in healthcare systems. However, when statistical learning models are used to guide decisions such as medical treatments, clinicians make decisions based on their predictions that can change subsequent patient data that are collected and used to "re-train" the predictive system, thereby updating the forecast probabilities at the presence of new evidence. Current predictive systems in practice fall in one of the following two categories: (1) they are never re-trained post first installation; or (2) they are periodically re-trained with the arrival of new data. However, the first approach leads to poor forecasts when the new data arrives and circumstances around the decisions change. On the other hand, recent advances in the theory of multi-armed bandits, reinforcement learning, and their applications to digital advertising and recommendation systems indicate that the second approach can also lead to low quality predictions due to the endogeneity of the decision making process. The new dynamic-learning and decision making tools developed in this project will be robust against these challenges by proper modeling of the interactions between the data elements, the decisions, and the consequences of the decisions on new data.