Mining Social Media Big Data for Toxicovigilance: Automating the Monitoring of Prescription Medication Abuse via Natural Language Processing and Machine Learning Methods

项目摘要 滥用处方药的问题已经达到了流行病的比例， 美国的根据国家药物研究所所长2014年的报告， 滥用（NIDA），估计有5200万人参与了非医疗使用 PM-其中很大一部分可归类为滥用。经常被滥用的PM 包括阿片类药物、中枢神经系统抑制剂和兴奋剂， 他们的虐待可能是严重的。在过去的15年里，PM误用和滥用的增加导致了 增加了急诊就诊率、吸毒成瘾率和过量死亡率。由于 发病率和死亡率迅速上升，目前正受到全国的关注。阿片类药物 这场危机的根源在于阿片类药物滥用，已被宣布为国家紧急状态， 美国总统尽管存在与总理虐待、监视相关的问题 处方药监测计划（PDMP）等计划不足， 许多缺点，从而限制了它们在真实的生活中的有用性。研究评估 不同类别的PM对滥用者群体的长期影响是罕见的， 行为。为了更好地描述问题的特征并实时监测， 需要查明信息，需要开发新的监测技术。到 为了解决这些问题，我们的项目旨在利用社交媒体数据进行表演， 毒物警戒社交媒体包含了大量关于PM滥用的知识， 以嘈杂的自然语言文本形式的滥用者。所提出的方法的核心是 机器学习系统，可以自动区分"虐待"和"非虐待" 表示从社交媒体收集的用户帖子。使用此分类系统，用户将 分为多个组别-（一）滥用者、（二）医疗使用者和（三）非使用者。的 开发的系统将收集纵向数据的用户暴露选定的PM通过定期 收集他们公开的帖子/讨论，并根据 年龄、性别和其他人口统计特征。这将使 对目标群体进行观察性研究，涉及数十万人， 队列成员。队列研究将重点分析从医疗过渡率 用于滥用不同的PM和过渡率从滥用PM非法类似物。 这个以数据为中心的框架将是开源的，它的实现将彻底改变 执行PM滥用监控的机制，并使未来的发展成为可能， 在最有效的时间段，针对特定群体的干预战略。

Project Summary The problem of prescription medication (PM) abuse has reached epidemic proportions in the United States. According to a 2014 report by the Director of the National Institute on Drug Abuse (NIDA), an estimated 52 million people, have been involved in the non-medical use of PMs— a significant portion of which can be classified as abuse. PMs that are commonly abused include opioids, central nervous system depressants and stimulants, and the consequences of their abuse may be severe. Increases in PM misuse and abuse over the last 15 years have resulted in increased emergency department visits, rates of addiction and overdose deaths. Due to the rapidly escalating morbidity and mortality, it is now receiving national attention. The opioid crisis, which has its root in opioid-based PM abuse, has been declared a national emergency by the president of the United States. Despite the problems associated with PM abuse, surveillance programs such as prescription drug monitoring programs (PDMPs) are inadequate and suffer from numerous shortcomings, thus limiting their usefulness in real life. Studies evaluating the long-term effects of distinct classes of PMs on cohorts of abusers are scarce and expensive to conduct. To better characterize the problem and to monitor it in real-time, new sources of information need to be identified and novel monitoring techniques need to be developed. To address these problems, our project aims to utilize social media data for performing toxicovigilance. Social media encapsulates an abundance of knowledge about PM abuse and the abusers in the form of noisy natural language text. At the heart of the proposed approach is a machine learning system that can automatically distinguish between `abuse' and `non-abuse' indicating user posts collected from social media. Using this classification system, users will be categorized into multiple groups—(i) abusers, (ii) medical users and (iii) non users. The developed system will collect longitudinal data for users exposed the selected PMs via periodic collection of their publicly available posts/discussions and automatically categorize them based on age, gender and additional demographic feature, when possible. This will enable the conducting of observational studies on targeted cohorts, involving hundreds of thousands of cohort members. The cohort studies will focus on analyzing the transition rates from medical use to abuse for distinct PMs and transition rates from abuse of PMs to illicit analogs. Implementation of this data-centric framework, which will be open source, will revolutionize the mechanism by which PM abuse monitoring is performed and enable the future development of intervention strategies targeted towards specific cohorts, at the most effective time periods.