Machine Learning for Identifying Adverse Drug Events

描述(由申请人提供)：由于对患者安全的直接影响，FDA、AHRQ和医学研究所已将新兴药物的上市后药物警戒标记为国家研究的高度优先事项。FDA、NIH基金会和Pharma已经建立了观察医疗结果合作伙伴关系，以开发和比较识别不良药物事件(ADE)的方法，FDA已经宣布了其哨兵倡议。这项拟议的工作将开发和研究用于ADE识别和预测的机器学习。后者是更简单的ADE预测任务，它假设已经确定了ADE--例如COX2抑制剂(Cox2ib)和心肌梗死(MI)之间的关联--目标是构建一个模型，能够准确地预测哪些患者最容易发生ADE，例如，如果他们服用COX2抑制剂，就会出现MI。我们的初步结果表明，使用机器学习，我们已经可以做出75%的灵敏度和75%的特异度的预测。由于我们没有观测到的类别变量，所以ADE辨识的任务比ADE预测更困难。鉴于一种新药已经上市，这项任务试图确定是否有任何以前没有预料到的不良事件是由该药物引起的。因为我们事先不知道这个事件是什么--它可能甚至不对应于现有的诊断代码--所以这个任务不能很好地适应标准的监督学习范例。我们的方法是使用反向机器学习来建立一个上市后监测工具，以便从电子医疗记录(EMR)或索赔数据中预测和/或检测药物的不良反应。我们从理论上和初步的经验结果都表明，这种方法可以发现一个或多个亚群的患者，他们的特征是以前没有预料到的不良事件-服用药物的患者比没有服用药物的患者遭受更高的发生率。这些事件不必对应于先前定义的ADE。为了建立和评估基于机器学习的ADE识别和预测系统，这项提议将解决以下具体目标：(1)将有监督的机器学习应用于ADE预测任务-预测如果给予药物，哪些患者最有可能遭受已知的ADE；(2)应用反向机器学习来识别新的ADE；(3)提供基于机器学习的ADE识别和预测的完整软件系统。该系统将在马什菲尔德诊所的EMR上和通过观察性医疗结果伙伴关系(OMOP)获得的真实和合成数据集上进行测试，其中一些初步结果将在本提案中提出。

DESCRIPTION (provided by applicant): Because of the direct effect on patient safety, the FDA, AHRQ and Institute of Medicine have flagged post- marketing pharmacovigilance of emerging medications as a high national research priority. The FDA, Foundation for the NIH and PhARMA have formed the Observational Medical Outcomes partnership to develop and compare methods for identification of adverse drug events (ADEs), and the FDA has announced its Sentinel Initiative. The proposed work will develop and study machine learning for ADE identification and prediction. The latter, easier task of ADE prediction assumes that an ADE has already been identified -- such as the association between Cox2 inhibitors (Cox2ib) and myocardial infarction (MI) - and the goal is to construct a model that can accurately predict which patients are most susceptible to having the ADE, e.g., having an MI if they take a Cox2 inhibitor. Our preliminary results show that using machine learning we can already make predictions at 75% sensitivity with 75% specificity. The task of ADE identification is more difficult than ADE prediction, because we do not have an observed class variable. Given that a new drug has been placed on the market, this task seeks to determine whether any previously-unanticipated adverse event is caused by the drug. Because we do not know in advance what this event is - it may not even correspond to an existing diagnosis code - this task does not neatly fit into the standard supervised learning paradigm. Our approach is to use reverse machine learning to build a post- marketing surveillance tool in order to predict and/or detect adverse reactions to drugs from electronic medical records (EMRs) or claims data. We show both theoretically and with preliminary empirical results that this approach can discover one or more subgroups of patients who are characterized by previously-unanticipated adverse events - events that patients on the drug suffer at a higher rate than patients not on the drug. These events do not have to correspond to previously-defined ADEs. In order to build and evaluate a machine learning-based system for ADE identification and prediction, this proposal will address the following specific aims: (1) apply supervised machine learning to the task of ADE prediction - predicting which patients are most likely to suffer a known ADE if given the drug; (2) apply reverse machine learning to identify novel ADEs; (3) provide a complete software system for machine learning-based identification and prediction of ADEs. This system will be tested on both the Marshfield Clinic's EMR, some preliminary results of which are presented in this proposal, and on real and synthetic datasets available through the Observational Medical Outcomes Partnership (OMOP).