III: Small: Collaborative Research: Social Media Based Analysis of Adverse Drug Events: User Modeling, Signal Reliability, and Signal Validation

药物不良反应(ADR)与严重的发病率和死亡率有关，是住院的一个重要原因，占所有住院人数的5%。美国每年报告的严重不良反应约有200万例；每年有10万例死亡与药物不良事件有关；严重不良反应在死亡原因中排名第4至第6位。这个问题源于这样一个事实，即一种给定药物的ADR概况在官方批准时很少是完整的。通常有限的审批前评估往往导致这样一种可能性，即当药物最终被批准用于普通人群(种族、性别、年龄、生活方式具有显著多样性)时，经常会观察到一些以前未确定的ADR。这一问题对于精神药物来说是严重的，因为大多数患有精神疾病的人往往有其他健康问题，个人同时服用多种药物(精神药物和非精神药物)，它们之间的相互作用往往未知。初步结果显示，使用社交媒体数据进行ADR信号检测是有希望的。然而，这些方法仍然面临着两个关键挑战，即信号可靠性和生物验证。因此，本项目建议对信号可靠性的关键决定因素进行详细研究：社交媒体来源的可信度、生成源内容的用户模型、从这些来源生成信号以及生成的信号的验证。这项工作将涉及负责药品审批、药品监督和疾病监测的政府机构、制药公司、医院和公众。拟议工作的影响将超出药物监测，因为拟议的方法可以适用于其他医疗问题，以及其他情景，如金融市场和国家安全。计划的教育活动包括扩展到高中生，以及本科生和研究生的参与。研究成果将通过专业期刊上的技术出版物和会议演示文稿进行传播。该项目有三个具体目标：(1)利用可信度分析、用户建模和通过深度学习进行信号融合，丰富社交媒体对不良药物事件的信号可靠性；(2)通过分子水平分析验证信号；(3)原型开发和评估。来自各种社交媒体渠道和其他用户生成内容来源的数据无处不在、准确和多样，因此必须认真考虑其可信度、新颖性、独特性和显着性。为了丰富信号的可靠性，该团队将提出使用可信度分析进行ADR信号检测的新方法，以及基于深度学习技术的用户建模和信号融合方法。对于信号验证，将使用对假想的ADR的生物学支持，基本上将来自社交媒体相互作用的高级别观察与分子水平网络和路径上的潜在关联联系起来。这些结果将改变目前在很大程度上依赖自愿报告的上市后药物监测的被动方法，确保基于社交媒体的方法的可靠性，从而使公众成为主动药物监测系统的组成部分。用于用户建模和信号生成的信号融合和深度学习的想法可以扩展到药物监督以外的其他用途。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Adverse drug reactions (ADRs) have been associated with significant morbidity and mortality, and have been a significant cause of hospital admissions, accounting for as much as 5% of all admissions. About 2,000,000 serious ADRs are reported yearly in the US; 100,000 annual deaths are related to adverse drug events; serious ADRs rank 4th to 6th as causes of death. The problem stems from the fact that the ADR profile of a given drug is rarely complete at the time of official approval. The typically limited preapproval evaluation often results in the possibility that when the drug is finally approved for use in the general population (with significant diversity in race, gender, age, lifestyle), some previously unidentified ADRs are often observed. This problem is acute for psychotropic medications, given the fact that most people with psychiatric diseases tend to have other health issues, with the individual taking multiple drugs at the same time (both psychotropic and non-psychotropic), with often unknown interactions between them. Initial results have shown the promise of using social-media data for ADR signal detection. However, these methods are still faced with two critical challenges, namely, signal reliability and biological validation. Thus, this project proposes a detailed study on key determinants of signal reliability: credibility of social media sources, model of the users that generate source content, signal generation from such sources, and validation of the generated signals. This work will be relevant to government agencies charged with drug approval, drug monitoring, and disease monitoring, drug companies, hospitals, and the general public. The impact of the proposed work will go beyond drug surveillance, since the approaches proposed can be adapted for other healthcare problems, and for other scenarios, such as financial markets, and national security. Planned educational activities include outreach to high-school students, and involvement of undergraduate and graduate students. Research results will be disseminated via technical publications in professional journals and conference presentations. The project has three specific aims: (1) Enrich signal reliability in social media analysis of adverse drug events, using credibility analysis, user modeling and signal fusion via deep learning; (2) Signal validation via molecular level analysis; (3) Prototype development and evaluation. The ubiquity, veracity and diversity of data from various social media channels and other sources of user-generated content necessitate a serious consideration of their credibility, recency, uniqueness and salience. To enrich signal reliability, the team will propose novel methods for ADR signal detection using credibility analysis, and for user modeling and signal fusion based on deep leaning techniques. For signal validation, biological support for hypothesized ADRs, essentially connecting high-level observations from social media interactions to potential associations at molecular level networks and pathways, will be used. The results will change the current largely passive approach to post-marketing drug surveillance that relies heavily on voluntary reports, by ensuring reliability in social-media based approaches, thus making the public an integral part of a proactive drug surveillance system. The idea of signal fusion and deep learning for user modeling and signal generation can be extended for other uses beyond drug surveillance.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.