The use of wearable electrocardiogram sensors to improve the prediction of cardiovascular disease

英国国家医疗服务体系（NHS）已将心血管疾病（CVD）确定为未来10年可以挽救生命的最大单一领域。预防心血管疾病的一项关键策略是使用风险预测模型，针对高危人群采取针对性的预防性干预措施。然而，许多人发现得太晚，超过50%的重大心脏事件发生在未被归类为高风险的患者身上。这意味着失去了预防的机会，尽管潜在的疾病发展了很多年。可穿戴传感器，如心电图贴片，有可能在患者的日常生活中连续、无创、无痛地测量心血管疾病的危险因素。显然有必要更好地评估心电图在使用该技术的健康个体纵向队列中的CVD风险预测。该项目旨在通过开发可重复的机器学习方法来弥合这一差距。使用英国生物银行数据集，收集了96000名参与者的临床心电图，这些参与者的健康状况被纵向随访。除此之外，还有可穿戴心电图数据的独特访问权限，这些数据收集于约30,000名参与者中（n=约10,000已收集）。该项目将开发深度学习方法，为参与队列生成更新的风险评分。目的和目标：1)建立临床心电测量的深度学习模型，预测未来CVD：研究将从学习与CVD预测相关的心电特征的过程开始。在2009-2010年期间测量数据的79,209名参与者和2012-20137年期间测量数据的20,218名参与者中，使用4导联ECG设备以500Hz记录数据。这些参与者已经发生了5000多起心血管疾病事件。候选方法包括在心电图诊断研究中发展卷积神经网络。2)检查临床心电图测量相对于当前临床标准的效用。这将包括与英格兰和威尔士日常临床实践中使用的QRISK3模型进行比较。除了在深度学习环境中使用Qrisk3特征外，该目标还将探索将学习到的ECG特征与现有Qrisk3模型相结合的潜力。根据统计分析，这将允许在发展到可穿戴传感器时确定最佳方法。3)研究穿戴式心电测量的加入是否可以改善cvd的预测。该目标提出迁移学习模型的建立，将早期隐藏层从临床训练模型中迁移过来。这有助于从可穿戴心电图数据中预测心血管疾病的发生。这将利用来自20,000名英国生物银行参与者的数据，这些参与者在2018年至2023年期间参加了成像评估诊所，并进行了2周的可穿戴心电图监测。由于该领域的大部分成功来自使用多个ECG导联的研究，该项目旨在将这一成功复制到可穿戴设备上。这将通过迁移学习来捕获目标1和2.4中使用的多个临床心电导联的关键特征来实现。探索潜在心血管疾病的检测机会，以及临床医生如何从时间序列心电数据中解释深度学习特征。深度学习模型，特别是时间序列数据，通常难以解释，这限制了它们最终临床应用的潜力。这一目的将探讨在心电图数据中检测潜在异常的机会。目标是将这些输入目标1中概述的风险预测器。这种情况的检测将有助于临床医生确定风险的潜在原因，并可与一系列可解释的方法相结合。这种方法将有助于确定哪些心电图特征导致相关的决定。该项目属于EPSRC医疗保健技术和人工智能（AI）主题，旨在应用最先进的深度学习方法来推进

The NHS has identified cardiovascular disease (CVD) as the single biggest area where lives can be saved over the next 10 years1. A key strategy in the prevention of CVD is the use of risk prediction models to target preventative interventions for people at higher risk2. However, many people are identified too late and over 50% of major cardiac events are in patients who were not classified as high-risk3. This means lost opportunities for prevention despite the underlying disease developing over many years. Wearable sensors, such as electrocardiogram (ECG) patches, have the potential to measure CVD risk factors continuously, noninvasively, and painlessly in patient's everyday lives4. There is a clear need to better evaluate ECGs for CVD risk prediction in a longitudinal cohort of healthy individuals using this technology.This project aims to bridge this gap via the development of reproducible machine learning methods. Using the UK Biobank dataset, which has collected in-clinic ECGs in 96,000 participants whose health outcomes are longitudinally followed-up8. This is in addition to unique access of wearable ECG data, collected in ~30,000 participants (n=~10,000 already collected). This project will develop deep learning methods to produce updated risk scores for the participating cohort.Aims and objectives:1) Establish a deep learning model for in-clinic ECG measurements to predict future CVD:The investigation will start with the process of learning ECG features relevant to the prediction of CVD. Data is recorded at 500Hz with a 4-lead ECG device in 79,209 participants who had data measured between 2009-2010, and 20,218 participants between 2012-20137. Over 5,000 incident CVD events have already occurred in these participants. Candidate methods include the development of a convolutional neural network in ECG diagnostic studies.2) Examine the utility of in-clinic ECG measurements over current in-clinic standards.This will include a comparison to the QRISK3 model which is used in day-to-day clinical practice in England and Wales10. This aim will explore the potential of combining learned ECG features with the existing Qrisk3 model, in addition to the use of Qrisk3 features in a deep learning setting. Following statistical analysis, this will allow for the optimal approach to be identified, when progressing to wearable sensors.3) Investigate if the addition of wearable ECG measurements can improve the prediction of CVDThis aim proposes the building of transfer learning models, with the early hidden layers transferred from the in-clinic trained model. This is to help predict incident CVD from wearable ECG data. This will utilise data from 20,000 UK Biobank participants, with wearable ECG monitoring for 2 weeks, who attended an imaging assessment clinic between 2018-2023. With much of the success in the field coming from studies using multiple ECG leads, this project aims to replicate this success to wearable devices. This will be achieved using transferring learning to capture the key features from the multiple in-clinic ECG leads used in aims 1 and 2.4) Explore detection opportunities for underlying cardiovascular conditions and how clinicians can interpret deep learned features from time-series ECG dataDeep learning models, particularly for time-series data, are often difficult to interpret, which limits their potential for eventual clinical use. This aim will investigate the opportunity to detect underlying abnormalities in the ECG data. With the goal of feeding these into the risk predictor outlined in aim 1. Detection of such conditions will aid clinicians in identifying the underlying cause of the risk and can be combined with a range of interpretable methods. Such methods will help to to identify which ECG features led to the relevant decision.This project falls within the EPSRC Healthcare technologies and Artificial intelligence (AI) themes, with the aim of applying state-of-the-art deep learning methods to advan