A multi-modal approach for efficient, point-of-care screening of hypertrophic cardiomyopathy

项目摘要 肥厚型心肌病（HCM）是最常见的遗传性心肌病，影响高达0.5%的 一般人口。肥厚型心肌病的发病率和死亡率的风险增加，但仍然临床 被低估了传统上，HCM的诊断依赖于全面的评估， 超声心动图或磁共振成像，不能用于筛选的模态 一般人口。随着新的疾病修饰疗法的出现，需要有效的策略来 在专业中心外加强HCM筛查。在这个博士后奖学金提出的研究将 利用先进的计算方法和不断扩大的可穿戴和便携式技术的可用性 使机器学习算法适应HCM的有效、即时筛查。在目标1中，申请人 建议使用大型心电图（ECG）库来调整ECG信号以供可穿戴设备使用 并微调这些信号以检测HCM。去噪-去噪算法和跨模态预训练 与相应的超声心动图和心脏磁共振视频将确保模型 分别对噪声和学习HCM表型的关键表示具有鲁棒性。在目标2中，单视图，两个- 三维超声心动图视频将被提取、预处理和增强，以模拟 护理图像采集。通过自我监督，对比预培训方法，申请人将建立 基于反映质量的超声心动图视频筛选HCM的数据高效计算模型 和床旁使用的独特挑战。在目标3中，申请人提出了一项前瞻性病例对照研究， 对患有和不患有HCM的患者进行的研究，这些患者将接受床旁心电图检查， 超声心动图，以测试两阶段HCM筛选方案的可行性和真实世界性能 根据目标1和2。该提案得到了生物医学机器专家的有力指导 学习、计算机视觉和实施科学。耶鲁大学医学院提供的设施和 实现研究目标所需的计算资源，而耶鲁-纽黑文健康 电子健康记录和良好的表型超声心动图和心电图图书馆确保了获得多样化的 有代表性的人口。建议的指导研究期间将支持申请人的培训， 计算机视觉、高级分析和医学信息学。在此期间获得的经验、数据和技能 这段时间将进一步支持申请人在实施科学的成功职业生涯做好准备， 心血管人工智能技术。

PROJECT SUMMARY Hypertrophic cardiomyopathy (HCM) is the most common inherited cardiomyopathy, affecting up to 0.5% of the general population. HCM confers an increased risk of morbidity and mortality but remains clinically underrecognized. Traditionally, the diagnosis of HCM has relied on comprehensive assessment by echocardiography or magnetic resonance imaging, modalities which are not available for screening of the general population. As novel disease-modifying therapies emerge, there is a need for efficient strategies to improve HCM screening outside specialized centers. The research proposed in this post-doctoral fellowship will leverage advanced computational methods and the expanding availability of wearable and portable technologies to adapt machine learning algorithms for the efficient, point-of-care screening of HCM. In Aim 1, the applicant proposes to use a large electrocardiographic (ECG) library to adapt ECG signals for use with wearable devices and fine-tune those signals for the detection of HCM. Noising-denoising algorithms and cross-modal pre-training with corresponding echocardiographic and cardiac magnetic resonance videos will ensure that the models are robust to noise and learn key representations of the HCM phenotype, respectively. In Aim 2, single-view, two- dimensional echocardiographic videos will be extracted, pre-processed, and augmented to simulate point-of- care image acquisition. Through a self-supervised, contrastive pre-training approach, the applicant will build data-efficient computational models to screen for HCM based on echocardiographic videos reflecting the quality and unique challenges seen with point-of-care use. In Aim 3, the applicant proposes a prospective case-control study of patients with and without HCM, who will undergo point-of-care electrocardiography and echocardiography, to test the feasibility and real-world performance of a two-stage HCM screening protocol based on Aims 1 and 2. The proposal is supported by strong mentorship from experts in biomedical machine learning, computer vision, and implementation science. The Yale School of Medicine offers the facilities and computational resources necessary to accomplish the research goals, whereas the Yale-New Haven Health electronic health record and well-phenotyped echocardiographic and ECG libraries ensure access to a diverse and representative population. The proposed period of mentored research will support the applicant’s training in computer vision, advanced analytics, and medical informatics. The experience, data, and skillset acquired during this period will further support the applicant in preparing for a successful career in the implementation science of cardiovascular artificial intelligence technologies.