Health Data Science EPSRC CDT

房性期前收缩（PAC）是常见的心律失常，虽然通常被认为是良性的，但可能表明潜在的心血管疾病（CVD），并可能导致严重的后果。目前，对其严重程度和临床意义的了解有限。运动测试可用于暴露不太频繁的心律失常，如PAC，这可能会在标准的10秒静息ECG测试中被遗漏。利用广泛的英国生物银行（UKB）数据集[5]，我的目标是开发和评估一种机器学习模型，用于在运动期间检测PAC，并调查运动诱导的PAC，主要心血管疾病和其他健康状况之间的关联。目前，在UKB运动队列（N= 95，071）这样大的数据集中，还没有量化单个PAC搏动的研究。解决这一差距可以促进生物医学研究，调查PAC负担与心血管和整体健康的关联。这方面的知识是至关重要的，以改善早期诊断，风险分层，并在临床实践中的预防性干预。其主要目的是加强检测和了解PAC及其对心血管和整体健康的影响。我假设运动引起的PAC可以使用机器学习模型准确检测，并且这些PAC与主要的心血管疾病和其他健康状况显著相关。我提出三个目标：1。开发和验证机器学习模型，以使用UKB数据集在运动期间检测PAC。该模型将使用运动ECG队列的预注释子集（112名参与者，79，113个心跳标签）进行训练和验证。使用卷积神经网络的初步结果显示，PAC类的精度为0.81，召回率为0.87，这将作为本研究的一部分进行微调。然后，它将在外部运动数据集上进行测试，以确保其鲁棒性和通用性，并作为开源工具提供给更广泛的研究社区。运动诱发的PAC与主要心血管疾病之间的相关性研究。使用经验证的PAC检测模型，将检查英国生物样本库队列中运动期间PAC发生率与主要CVD患病率的关系。该研究将扩展Duijvenboden等人对室性早搏和心血管疾病的关联分析，这反映了拟议研究的可能效力。以无假设的方式探索运动诱导的PAC与各种其他健康状况之间的关系。将使用英国生物银行数据集进行探索性分析，以确定运动期间PAC与各种健康状况之间的潜在关联。这项工作将遵循Watts等人的研究中阐述的无假设格式。该研究项目将产生一个专门设计用于在运动期间检测个体PAC的开源机器学习模型，这是一种新方法，可以解决当前诊断能力的显着差距。如果时间允许，该模型也可以适用于静息ECG数据。此外，对PAC与各种疾病的关联的无假设探索是一种创新策略，可能会发现以前未知的健康影响。衍生的表型将与英国生物库社区共享，以促进对未充分探索的心律失常的开放研究。通过加强运动期间PAC的检测和理解，这项工作将有助于改善心血管和其他疾病的早期诊断和预防策略。最终，这将导致更好的患者结局，并减少与未诊断或管理不善的心律失常和相关疾病相关的医疗负担。

Premature atrial contractions (PACs) are common cardiac arrhythmias which, although usually considered benign, can indicate underlying cardiovascular disease (CVD) and potentially lead to severe outcomes. Currently, there is limited understanding of their severity and clinical implications. Exercise testing can be used to expose less frequent arrhythmias, such as PACs, which might be missed during a standard 10-second resting ECG test. Leveraging the extensive UK Biobank (UKB) dataset [5], I aim to develop and evaluate a machine learning model for detecting PACs during exercise, and to investigate the associations between exercise-induced PACs, major CVDs, and other health conditions. Currently, there are no studies quantifying individual PAC beats in such a large dataset as the UKB exercise cohort (N=95,071). Addressing this gap can facilitate biomedical studies investigating the associations of PAC burdens with cardiovascular and overall health. This knowledge is crucial for improving early diagnosis, risk stratification, and preventive interventions in clinical practice.This project falls within the EPSRC Healthcare Technologies theme. The main aim is to enhance the detection and understanding of PACs and their implications for cardiovascular and overall health. I hypothesise that exercise-induced PACs can be accurately detected using machine learning models and that these PACs are significantly associated with major CVDs and other health conditions. I propose three objectives:1. Development and validation of a machine learning model to detect PACs during exercise using the UKB dataset. The model will be trained and validated using a pre-annotated subset of the exercise ECG cohort (112 participants, 79,113 per-heartbeat labels). Preliminary results using a convolutional neural network show promising performance on the PAC class with precision of 0.81 and recall of 0.87 which will be fine-tuned as part of this study. It will then be tested on an external exercise dataset to ensure its robustness and generalisability, and made available to the wider research community as an open-source tool.2. Investigation of the association between exercise-induced PACs and major CVDs. Using the validated PAC detection model, PAC incidence during exercise in relation to the prevalence of major CVDs within the UK Biobank cohort will be examined. The study will expand on the associative analysis of premature ventricular contractions and CVDs by Duijvenboden et al. which reflects the likely power of the proposed study.3. Exploration of the relationship between exercise-induced PACs and various other health conditions in a hypothesis-free manner. An exploratory analysis using the UK Biobank dataset will be carried out to identify potential associations between PACs during exercise and a wide range of health conditions. This work will follow the hypothesis-free format which is illustrated in the study by Watts et al. This research project will result in an open-source machine learning model specifically designed to detect individual PACs during exercise, a novel approach that addresses a significant gap in current diagnostic capabilities. If time permits, the model could also be adapted for use on resting ECG data. Additionally, the hypothesis-free exploration of PACs' associations with various diseases is an innovative strategy that may uncover previously unknown health implications. The derived phenotypes will be shared with the UK Biobank community in order to promote open research into underexplored arrhythmias. By enhancing the detection and understanding of PACs during exercise, this work will contribute to improved early diagnosis and preventive strategies for cardiovascular and other diseases. Ultimately, this will lead to better patient outcomes and reduced healthcare burdens associated with undiagnosed or poorly managed arrhythmias and related conditions.