Mining Social Network Postings for Mentions of Potential Adverse Drug Reactions

描述（由申请人提供）： 药物在市场上广泛使用之前，要经过广泛的动物试验和人体临床试验。上市前测试产生了关于药物作为其被批准的条件的治疗的有效性的合理高质量的信息，但给出了药物的安全性的非常不完整的画面。上市后监测目前主要依赖于医疗保健专业人员（最近，患者本人）通过FDA的安全信息和不良事件报告计划MedWatch向FDA自愿报告。自我报告的患者信息捕获了一个有价值的观点，已发现其质量与卫生专业人员提供的信息相似，目前仅通过正式的MedWatch表格捕获。该应用程序的总体目标是部署所需的基础设施，以探索非正式社交网络帖子的价值，作为药物上市后不久潜在药物不良反应的“信号”来源，特别注意这些信息可能具有的价值，以比目前可能的更早地检测不良事件，并检测传统手段不易捕获的影响。尽管处理口语文本存在重大挑战，但我们在这方面的原型研究在识别这些帖子中提到的不良反应方面表现出了良好的表现，公众提到的效果与我们研究的药物记录的效果之间存在显著相关性。具体目标包括：（1）。建立基础设施，以便处理健康相关社交网络网站上关于药物的在线用户评论。特别是，我们试图识别和提取这些非正式帖子中提到的不利影响，并将其映射到标准术语。我们将建立在我们初步的词汇方法来寻找提及，并提出了一个机器学习的变化（通常称为主动学习），其中机器学习框架具有控制将选择哪些实例用于训练数据的能力，以及其他用于规范化的创新语义方法（将提及映射到已建立的正式术语）和情感分析（发现提及是否报告了积极或消极的影响）; 2）评价提取和鉴定系统的灵敏度和特异性，以及通过对一组具有众所周知的不良反应的药物的特定案例研究和通过监测关于自2007年以来发布的一组精选药物。我们现有的手动注释金标准将通过药理学家（Karen Smith）领导的专门注释工作进行扩展。3)将从患者评论中提取的知识与从FDA监督的既定药物安全性监测计划中获得的知识进行比较。我们认识到，通过部署的基础设施获得的数据将无法用于定义独立的ADR。然而，如果这种方法得到验证，它可以提供有用的信号，以补充已经建立的流程和数据源。

DESCRIPTION (provided by applicant): Drugs undergo extensive testing in animals and clinical trials in humans before they are marketed for widespread use in the population. Pre-market testing produces reasonably high quality information about the efficacy of the drug as a treatment for the condition for which it was approved, but gives a very incomplete picture of the drug's safety. Post-marketing surveillance currently relies mainly on voluntary reporting to the FDA by health care professionals (and recently, patients themselves) through MedWatch, the FDA's safety information and adverse event reporting program. Self-reported patient information captures a valuable perspective that has been found to be of similar quality to that provided by health professionals, and currently it is only captured via the formal MedWatch form. The overarching goal of this application is to deploy the infrastructure needed to explore the value of informal social network postings as a source of "signals" of potential adverse drug reactions soon after the drugs hit the market, paying particular attention at the value such information might have to detect adverse events earlier than currently possible, and to detect effects not easily captured by traditional means. Despite the significant challenge of processing colloquial text, our prototype study in this direction showed promising performance in identifying adverse reactions mentioned in these postings, with significant correlations between the effects mentioned by the public and those documented for the drugs we studied. Specific aims to be addressed include: 1). To establish the infrastructure that enables processing of online user comments about the drug on health-related social network websites. Particularly, we seek to recognize and extract mentions of adverse effects in those informal postings, and to map them to standard terminology. We will build on our preliminary lexical approach for finding the mentions, and propose a variation of machine learning (commonly referred to as active learning) where the machine learning framework has the ability to control what instances will be selected for use in the training data, among other innovative semantic approaches to normalization (mapping of the mentions to established, formal terms) and sentiment analysis (to discover whether a mention is reporting a positive or a negative effect); 2) To evaluate the sensitivity and specificity of the extraction and identification systems, as well as the predictive value of the extracted knowledge through specific case studies of a set of drugs with well known adverse reactions and by monitoring postings about a select group of drugs released since 2007. Our existing manually annotated gold standard will be expanded through a dedicated annotation effort led by a pharmacologist (Karen Smith). 3) To compare the knowledge extracted from patient comments to what is derived from the established drug safety monitoring scheme overseen by the FDA. We recognize that the data obtained through the deployed infrastructure would not be able to be used to define an ADR standing on its own. However, if this method is validated, it could provide useful signals to complement the already established processes and data sources.