AE2Vec: Medical concept embedding and time-series analysis for automated adverse event detection

7.项目总结/摘要 不良事件对医疗干预（药物、器械等）构成重大挑战，估计 2.3 1969年至2002年期间发生的百万例药物不良事件。不良事件导致住院时间延长 更高的医疗费用和更高的死亡率。显然需要进行不良事件监测，但 人工病历审查和自愿报告的标准既费时又不可持续。自愿 报告还遗漏了大多数不良事件病例。电子健康记录（EHR）的广泛采用 获取大多数美国患者的医学数据，并为可持续不良事件提供机会 通过自动化策略进行监控。然而，自动化不良事件监测存在两个障碍。 首先，不良事件在国际疾病分类（ICD）诊断中的代表性很差 代码.这阻碍了使用简单的基于规则的代码或标志/触发器方法的努力，而复杂和 高性能的文本挖掘方法由于难以适应其他医疗保健网站而受到阻碍 和大型数据网络进行更广泛的监控。第二，EHR中的时间信息固有的不利 事件的定时和排序对于捕获是具有挑战性的。现有方法面临的挑战包括： 将相关的医学概念作为独立的实体处理，数据的快速爆炸抑制了扩展到 大量的医学概念和人类的可解释性。我们的首要目标是扩大现有的 生物医学信息学工具，以更好地捕捉不良事件，并更全面地代表 完整的患者医疗轨迹，以识别不良事件发展的原型。我们将试点这些 用于接受免疫检查点抑制剂（ICI）治疗的癌症患者的方法。具体目标1： 结合医学概念嵌入和聚类方法来绘制疾病“地图”，分割成 “社区”标记为他们所描述的条件，包括不良事件。在第二阶段，我们将 测试一种在疾病地图上追踪病人轨迹的新方法，并假设我们可以识别出 使用时间序列聚类具有不同临床结果的原型患者轨迹。这项工作 解决了基于EHR的表型和不良事件监测方面的差距。它有可能告知 风险因素识别、不良事件发展预测和患者评估 结果，以及奠定了一个关键的踏脚石，为进一步发展的EHR为基础的表型， 生物医学信息学这个奖学金将使我能够发展我在生物医学信息学方面的技能 方法，将临床观点融入我的研究，磨练我的写作和演讲技巧，并扩大 我的专业网络在这个奖项结束时，我将朝着成为一个 独立的医生-信息学家，融合临床经验和信息学工具，以改善患者护理。

7. Project Summary/Abstract Adverse events pose a significant challenge to medical interventions (drugs, devices, others) with an estimated 2.3 million cases of adverse drug events between 1969-2002. Adverse events are responsible for longer hospital stay, higher healthcare costs, and higher mortality. There is a clear need for adverse event surveillance, but the standards of manual chart review and voluntary reporting are time-consuming and unsustainable. Voluntary reporting also misses most adverse event cases. The widespread adoption of electronic health records (EHRs) captures medical data for the majority of US patients and presents an opportunity for sustainable adverse event surveillance via automated strategies. However, there are two barriers to automating adverse event surveillance. First, adverse events are poorly represented by International Classification of Disease (ICD) diagnosis codes. This has inhibited efforts to use simple rules-based code or flag/trigger approaches, while complex and high-performing text-mining approaches are thwarted by the difficulty of adapting them to other healthcare sites and large data networks for wider surveillance. Second, temporal information in the EHR inherent to adverse event timing and sequencing is challenging to capture. The challenges to existing approaches include – treatment of related medical concepts as independent entities, the rapid explosion of data inhibiting scaling to large numbers of medical concepts, and human interpretability. Our overarching goal is to expand on existing biomedical informatics tools to better capture adverse events and more comprehensively represent the full patient medical trajectory to identify archetypes of adverse event development. We will pilot these methods for cancer patients undergoing immune checkpoint inhibitor (ICI) therapy. In Specific Aim 1, we will incorporate medical concept embedding and clustering methods to draw a “map” of disease, segmented into “neighborhoods” labeled for the conditions they describe, including adverse events. In Specific Aim 2, we will test a novel method for tracking patient trajectories on a map of disease and hypothesize that we can identify archetypal patient trajectories that have different clinical outcomes using time-series clustering. This work addresses gaps in EHR-based phenotyping and adverse event surveillance. It has the potential to inform risk factor identification, prediction of adverse event development, and prognostication of patient outcomes, as well as lay a crucial stepping-stone for further progression of EHR-based phenotyping in biomedical informatics. This fellowship award will enable me to develop my skills in biomedical informatics methods, integrate clinical perspective into my research, hone my writing and presentation skills, and expand my professional network. At the conclusion of this award, I will have made strides towards becoming an independent physician-informaticist, fusing clinical experience and informatics tools to improve patient care.