Machine Learning for Atrial Fibrillation Ablation

摘要 在美国，房颤是最常见的心律失常，影响着600多万人，是一种主要的 公共卫生问题。房颤对医疗保健系统来说是昂贵的并导致显著的健康后果(例如， 中风、心力衰竭、痴呆症、生活质量下降)。随着时间的推移，房颤患者的频率会增加 以及房颤发作的持续时间。散发性房颤发作的随机发生和抗凝治疗的必要性 预防中风使房颤难以管理。许多房颤患者依次寻求心房颤动消融术(AFA) 提高生活质量，减少房颤发作。AFA，即烧灼左心房的区域，是最严重的 对持续性/阵发性房颤的有效治疗。AFA的成功率各不相同，但许多患者并不是没有房颤 在AFA之后。在领先的AFA中心，首次AFA后一年和两年的无房颤率分别为40%和37%， 分别进行了分析。鉴于AFA的成功率不高，选择这种手术的患者应该会得到更多 请注意。社会人口学和临床表型数据已被用来预测AFA反应，但 总的来说，他们的预测能力很差。电子健康记录(EHR)系统的广泛采用 为预测AFA结果提供了一个模式转变的成熟机会。更好地了解患者 预测AFA结果的特定因素将为患者选择该手术提供信息。为此，我们建议 为了使用机器学习技术来开发对主要AFA过程的结果的预测模型， 解决以下具体目标和研究问题： 1.目标1：使用机器学习来预测不良的AFA结果。 ·在初始手术前，现有的风险评分对AFA并发症的预测效果如何？ ·根据EHR数据训练的机器学习模型能否更好地预测AFA并发症？ 2.目标2：数据驱动的AFA结果亚组识别。 ·聚类分析能否根据结果轨迹确定有用的子组？ ·是否有其他非监督最大似然算法(如序列模式挖掘)可供分析 病人的结局轨迹？ 3.目标3：开发开源软件工具包。 该项目将为未来改进现有的机器学习方法以及 新方法的发展，以改善房颤复发的预测。

SUMMARY Affecting over 6 million people in the U.S., atrial fibrillation (AF), the most common cardiac arrhythmia, is a major public health concern. AF is costly to the health care system and leads to significant health consequences (e.g., stroke, heart failure, dementia, decreased quality of life). With time, AF patients experience increased frequency and duration of AF episodes. Random occurrence of sporadic AF episodes and the need for anticoagulation to prevent stroke make AF difficult to manage. Many AF patients seek out atrial fibrillation ablation (AFA) in order to improve quality of life and decrease AF episodes. AFA, cauterization of areas of the left atrium, is the most effective treatment for persistent / paroxysmal AF. AFA success rates vary, but many patients will not be AF-free following AFA. At leading AFA centers, AF-free rates at one and two years after initial AFA were 40% and 37%, respectively. Given the modest success rates of AFA, patient selection for this procedure should receive more attention. Sociodemographic and clinical phenotype data have been used to predict AFA response, but collectively they have poor predictive ability. The widespread adoption of electronic health record (EHR) systems presents a ripe opportunity for a paradigm shift for predicting AFA outcomes. A better understanding of patient specific factors predicting AFA outcome will inform patient selection for this procedure. To this end we propose to use machine learning techniques to develop predictive models for outcomes of primary AFA procedures, addressing the following specific aims and research questions: 1. Aim 1: Predict adverse AFA outcomes using machine learning. • How well do existing risk scores predict AFA complications prior to initial procedure? • Can a machine learning model trained on EHR data provide better prediction of AFA complications? 2. Aim 2: Data-driven AFA outcome subgroup identification. • Can cluster analysis identify useful subgroups based on outcome trajectory? • Are other unsupervised ML algorithms such as sequential pattern mining alternatives for analyzing patient outcome trajectories? 3. Aim 3: Develop an open-source software toolkit. This project will lay the foundation for future refinement of existing machine learning methods as well as development of new methods to improve prediction of AF recurrence following AFA.