Cardiovascular medicine has blazed the path as a data-driven and evidence-rich specialty. Throughout its rich history, cardiovascular care has leveraged progressively complex data for clinical management, beginning with acoustic data on a stethoscope, electrical information on 12-lead ECG and implanted devices, cardiac and vascular measurements with invasive pressure transducers, visualization of the vasculature on angiography, and continually evolving landscape of biomarkers and noninvasive multimodality imaging. These data streams are reflected in everyday cardiovascular practice and have historically had 2 crucial elements: (1) acquisition of data in a health care setting and (2) interpretation by expert clinicians. Over the past decade, computational developments have allowed the design of complex algorithms that can leverage these diverse data streams to enhance cardiovascular care. The growing interpretive and generative power of artificial intelligence (AI) has been brought to the forefront with innovation in large language models and foundation models more generally. Despite these advancements, development has been limited by reliance on snapshots around clinical care rather than continuously captured data. Moreover, the diagnostic or prognostic labels used to advance such developments are derived from routine clinical care, thereby encoding limitations in our processes into our algorithms.

We envision AI-driven cardiovascular care innovation in 3 key domains: (1) sensing—leveraging a broader and growing stream of data integrated into clinical care, (2) interpreting—algorithmic capability in handling progressively complex data with valid and unbiased inference, and (3) implementing—integrating solutions into clinical care workflows for high-quality, equitable care. In a historical context, the majority of the last century has focused on developing innovation in sensing. This innovation is best reflected in a growing list of diagnostic modalities and the digitization of health care records. The past decade represented a major boost in interpreting the data, represented by AI-driven diagnostic interpretation and inference (Figure).

Figure. Past and future development in cardiovascular artificial intelligence (AI).

How do we envision the future will look that integrates innovation across domains?

In our current paradigm, we capture high-quality data infrequently. Therefore, diagnostic tests provide exquisite detail about some element of health or illness but are timed with health encounters. Such snapshots do not capture the full range of the human condition and are limited to sparse assessments during selective periods. To address this gap, various sensors have been developed. For example, historically, portable and noninvasive devices, such as the Holter monitor, enabled the collection of electrocardiographic data at home. However, they were largely limited to short periods of assessments. Such a restriction reflected the inconvenience of the devices and also the deluge of data that had to be manually reviewed. The ubiquity of wearable devices¹ has eased data acquisition but worsened the data deluge.

Despite this overwhelming quantity, individual-generated health data represent a complementary paradigm where high-frequency assessments of less specific testing can provide comprehensive coverage of the range of the human condition. This is increasingly possible as wearable devices have sensors capturing motion, sleep, and multiple physiological measures, including heart rate, oxygen saturation, and even electrocardiographic recordings. These data streams are, by design, large and complex, primed for interpretation by deep learning and AI.² How successfully we leverage these advances in model development and implementation represents a new frontier in methodologic and applied data science.

Advances in sensing ushered in an era of big data. However, over the past decade, clinicians have spent increasingly more time with patient data while unable to leverage a majority of generated data. For the data to be useful, algorithms that can organize, process, and transform them into actionable knowledge are essential. The current approach to algorithm development has largely relied on highly curated, human-annotated data streams. More broadly, AI advances in computer vision and natural language processing have been driven by methodologic innovation but also by the curation of large human-annotated data sets across these key domains. For example, ImageNet, a publicly available data set of 14M hand-annotated images with more than 20 000 categories, has fueled progress in deep learning for the past decade.³ While images in ImageNet can be downloaded and labeled quickly and inexpensively, amassing clinical data and high-quality annotations requires significant time and expertise.

Given these challenges, in health care, we have focused on (1) a small number of publicly available labeled data sets and (2) opportunistically labeled data sets arising from routine clinical care. Despite the relatively small size, publicly available data have had a disproportionate impact on the research community.⁴ First, it is easier to work with previously collected data, and second, publicly available data offer an opportunity for benchmarking and reproducibility critical to advancing AI. Even small data sets such as the MIT-BIH Arrhythmia database Physionet have fueled thousands of AI articles. While methodologically impactful, such analyses are somewhat limited in clinical scope. With the advent of the electronic health record, labeled data streams available through routine clinical care have prompted additional analyses. For example, electrocardiograms have been used to identify those with reduced left ventricular systolic function.⁵ This, too, however, is only possible for conditions that are frequently manifested and diagnosed. For other conditions, where data streams are not labeled as part of routine care (eg, unstructured notes), AI has been limited by its need for large, labeled data sets. This, in turn, has dissociated development from need, especially because a bulk of development in health care is underresourced.

Addressing these underresourced problems requires a new approach. Instead of relying entirely on carefully annotated data, we may turn to a new generation of models, called foundation models, which learn data representations useful across various tasks.⁶ Once developed, such models can achieve strong performance while training with limited amounts of explicitly labeled data.⁶ Such models are often trained in an unsupervised or self-supervised approach. This means that no external input is provided to the data, but creative training signals are introduced, such as learning to reconstruct the data (eg, fill in the blank) to ensure that the model learns the data structure. Once developed, such models can then undergo further training (ie, fine-tuning) with limited amounts of data that are explicitly labeled. The foundation provides a reasonable starting point for the training algorithm, meaning that it is more likely to converge to a good solution, even with a limited amount of data. The model development dissociated from specific tasks also allows development to be performed on large data sets, improving population coverage, while also eliminating the necessity of sharing patient data. The sharing of models as starting points across institutions and not just end points promises to accelerate progress.

Health care transformation is often limited by the last-mile challenges. This has been reflected in numerous quality measurement initiatives, where diagnostic tests or therapeutic interventions are underused. This has also been the case with the deployment of AI in clinical care. The scientific developments have historically ceased at the time of publication of the model and the results. However, there has been a much-needed transition in understanding AI integration into clinical workflows as a necessary component of scientific development. This requires navigating issues in data stream integration, platform development for delivering and optimizing care, and building clinician trust for deployed algorithms. Once integrated, there is also substantial oversight needed in monitoring the continued use of AI because data and clinical workflows can change.⁷

The impact of patient risk stratification models, such as the Epic deterioration index, which have been deployed widely, is still largely unknown. Understanding the impact on patient outcomes requires careful prospective evaluation.⁸ The integration and implementation of AI-based health solutions require navigating unexpected challenges. Moreover, integration challenges may result in perfect and accurate models failing to have a positive impact in a real-world clinical setting against an imperfect model with thoughtful integration into clinical workflows that may transform care. These challenges would be amplified for the broader foundation models, which require testing for each important task. As more researchers translate AI solutions from in silico to in situ,⁹ we will better understand how to adapt clinical workflows and measure utility and impact, rather than accuracy and calibration.

Furthermore, while the technology that leverages novel, patient-generated data evolves, there will be a specific need to develop the capacity for incorporating such data within the electronic health record. This would require creating specific data standards and storage strategies. Alternatively, if learning and deploying these technologies can be done directly at the level of participants, with the integration of predictive outputs from these algorithms integrated with clinical care, these would have to be specifically evaluated and designed. In addition to infrastructural and logistical considerations, how a clinician should respond to the algorithm care, the liability of algorithmic failures, and required oversight by users of the technology in clinical care will need to be explicitly defined for each technology.¹⁰ If the innovation in AI ceases to extend beyond algorithm development to care innovation, the value of AI to future clinical care will be limited.

In summary, we are in the midst of an AI summer. We can acquire and process large data streams. Our innovation milieu has largely focused on hardware and software solutions to address problems of sensing and interpreting complex data, but the actual change in care would require addressing many key challenges around aligning development with clinical needs and ensuring a focus on integration while improving trust in technology through its rigorous development. While we move toward a data-driven and AI-fueled future of cardiovascular care, the decisions that align development with implementation will get us closer to the health care utopia of optimal cardiovascular care quality and patient outcomes.

Dr Khera receives support from the National Heart, Lung, and Blood Institute (NHLBI) of the National Institutes of Health (under award K23HL153775) and the Doris Duke Charitable Foundation (under award 2022060). Dr Wiens receives support from the National Institutes of Health (NHLBI, National Library of Medicine, and National Institute on Aging), the National Science Foundation, the Gordon and Betty Moore Foundation, Cisco Research, the Agency for Healthcare Research and Quality, Juvenile Diabetes Research Foundation, the Kahn Foundation, the Alfred P. Sloan Foundation, and the Michigan Department of Health and Human Services.

Disclosures In addition to the National Institutes of Health and the Doris Duke Charitable Foundation, Dr Khera also receives research support, through Yale, from Bristol-Myers Squibb, NovoNordisk, and BridgeBio. He is a coinventor of US Provisional Patent Applications 63/177,117, 63/428,569, 63/346,610, 63/484,426, and 63/508,315. He also co-founded Evidence2Health and Ensight-AI, health platforms to improve evidence-based cardiovascular care. In addition to the external agencies listed above, Dr Wiens receives research support from the University of Michigan. She also serves on the board of Machine Learning for Healthcare.

The opinions expressed in this article are not necessarily those of the editors or of the American Heart Association.

The American Heart Association celebrates its 100^th anniversary in 2024. This article is part of a series across the entire AHA Journal portfolio written by international thought leaders on the past, present, and future of cardiovascular and cerebrovascular research and care. To explore the full Centennial Collection, visit https://www.ahajournals.org/centennial

For Sources of Funding and Disclosures, see page 205.

心血管医学作为数据驱动和证据丰富的专业开辟了道路。纵观其丰富的历史，心血管护理已经利用越来越复杂的数据进行临床管理，从听诊器上的声学数据、12 导联心电图和植入设备上的电气信息、使用侵入式压力传感器进行心脏和血管测量、血管造影上脉管系统的可视化开始，以及不断发展的生物标志物和无创多模态成像领域。这些数据流反映在日常心血管实践中，并且历史上有两个关键要素：（1）在医疗保健环境中获取数据；（2）由专业临床医生解释。在过去的十年中，计算的发展使得复杂算法的设计成为可能，这些算法可以利用这些不同的数据流来增强心血管护理。随着大型语言模型和更广泛的基础模型的创新，人工智能 (AI) 日益增长的解释能力和生成能力已被推到最前沿。尽管取得了这些进步，但由于依赖临床护理快照而不是持续捕获的数据，发展受到限制。此外，用于推进此类发展的诊断或预后标签源自常规临床护理，从而将我们流程中的局限性编码到我们的算法中。

我们设想人工智能驱动的心血管护理创新将在 3 个关键领域实现：(1) 感知——利用更广泛且不断增长的数据流，并将其整合到临床护理中；(2) 解释——通过有效且公正的推理来处理日益复杂的数据的算法能力；以及(3) 实施——将解决方案整合到临床护理工作流程中，以实现高质量、公平的护理。在历史背景下，上个世纪的大部分时间都集中在传感领域的创新发展上。这种创新最好地体现在不断增长的诊断方式和医疗保健记录的数字化上。过去十年代表了数据解释的重大进步，以人工智能驱动的诊断解释和推理为代表（图）。

数字。 心血管人工智能（AI）的过去和未来发展。

我们如何展望跨领域整合创新的未来？

在我们当前的范例中，我们很少捕获高质量的数据。因此，诊断测试提供了有关健康或疾病某些因素的详细信息，但与健康状况同步。此类快照无法全面​​反映人类状况，并且仅限于选择性时期的稀疏评估。为了解决这一差距，人们开发了各种传感器。例如，历史上，便携式和非侵入性设备（例如动态心电图监测仪）可以在家中收集心电图数据。然而，它们很大程度上仅限于短期评估。这种限制反映了设备的不便以及必须手动审查的大量数据。可穿戴设备¹的普及简化了数据采集，但也加剧了数据泛滥。

尽管数量巨大，但个人生成的健康数据代表了一种补充范式，其中对不太具体的测试进行高频评估可以全面覆盖人类状况的范围。随着可穿戴设备具有捕捉运动、睡眠和多种生理测量（包括心率、氧饱和度，甚至心电图记录）的传感器，这种可能性越来越大。从设计上来说，这些数据流庞大且复杂，准备好由深度学习和人工智能进行解释。²我们如何成功地利用模型开发和实施方面的这些进步代表了方法论和应用数据科学的新前沿。

传感技术的进步开创了大数据时代。然而，在过去的十年中，临床医生在患者数据上花费的时间越来越多，但却无法利用大部分生成的数据。为了让数据发挥作用，能够组织、处理数据并将其转化为可操作知识的算法至关重要。当前的算法开发方法在很大程度上依赖于高度策划的、人工注释的数据流。更广泛地说，人工智能在计算机视觉和自然语言处理方面的进步不仅受到方法创新的推动，而且还受到这些关键领域中大型人工注释数据集的管理的推动。例如，ImageNet 是一个公开数据集，包含 1400 万张手工注释图像，包含超过 20,000 个类别，在过去十年中推动了深度学习的进步。³虽然 ImageNet 中的图像可以快速且廉价地下载和标记，但积累临床数据和高质量注释需要大量时间和专业知识。

鉴于这些挑战，在医疗保健领域，我们重点关注（1）少量公开可用的标记数据集和（2）常规临床护理中产生的机会性标记数据集。尽管规模相对较小，但公开数据对研究界产生了不成比例的影响。⁴首先，使用以前收集的数据更容易，其次，公开数据提供了对推进人工智能至关重要的基准测试和可重复性的机会。即使是像 MIT-BIH 心律失常数据库 Physionet 这样的小数据集也催生了数千篇人工智能文章。虽然在方法上有影响力，但此类分析在临床范围上有所限制。随着电子健康记录的出现，通过常规临床护理获得的标记数据流促进了额外的分析。例如，心电图已被用来识别那些左心室收缩功能降低的人。⁵然而，这也仅适用于经常表现和诊断的情况。对于其他情况，数据流没有被标记为日常护理的一部分（例如，非结构化笔记），人工智能因其对大型标记数据集的需求而受到限制。这反过来又使发展与需求脱节，特别是因为卫生保健领域的大部分发展都是资源不足的。

解决这些资源不足的问题需要采取新的方法。我们可以转向新一代模型，称为基础模型，而不是完全依赖仔细注释的数据，它学习跨各种任务有用的数据表示。⁶一旦开发出来，此类模型可以在使用有限数量的明确标记数据进行训练时实现强大的性能。⁶此类模型通常采用无监督或自监督的方法进行训练。这意味着不向数据提供外部输入，而是引入创造性的训练信号，例如学习重构数据（例如填空）以确保模型学习数据结构。一旦开发出来，此类模型就可以使用有限数量的明确标记的数据进行进一步的训练（即微调）。该基础为训练算法提供了合理的起点，这意味着即使数据量有限，它也更有可能收敛到良好的解决方案。与特定任务无关的模型开发还允许在大型数据集上进行开发，提高人口覆盖率，同时还消除了共享患者数据的必要性。将模型共享作为跨机构的起点而不仅仅是终点，有望加速进展。

医疗保健转型往往受到最后一英里挑战的限制。这已反映在许多质量测量举措中，其中诊断测试或治疗干预措施未得到充分利用。人工智能在临床护理中的部署也是如此。从历史上看，科学发展在模型和结果发布时就停止了。然而，在理解人工智能融入临床工作流程作为科学发展的必要组成部分方面，我们急需进行转变。这需要解决数据流集成、提供和优化护理的平台开发以及建立临床医生对部署算法的信任等方面的问题。一旦整合，还需要对人工智能的持续使用进行大量监督，因为数据和临床工作流程可能会发生变化。⁷

患者风险分层模型（例如已广泛部署的 Epic 恶化指数）的影响仍然很大程度上未知。了解对患者结果的影响需要仔细的前瞻性评估。⁸基于人工智能的健康解决方案的集成和实施需要应对意想不到的挑战。此外，集成挑战可能会导致完美而准确的模型无法在现实世界的临床环境中对不完美的模型产生积极影响，而无法将不完美的模型深思熟虑地集成到可能改变护理的临床工作流程中。对于更广泛的基础模型来说，这些挑战将被放大，这需要对每项重要任务进行测试。随着越来越多的研究人员将人工智能解决方案从计算机转化为现场，⁹我们将更好地了解如何调整临床工作流程并衡量效用和影响，而不是准确性和校准。

此外，虽然利用患者生成的新颖数据的技术不断发展，但仍特别需要开发将此类数据合并到电子健康记录中的能力。这需要创建特定的数据标准和存储策略。或者，如果可以直接在参与者层面学习和部署这些技术，并将这些算法的预测输出与临床护理相结合，则必须对这些技术进行专门评估和设计。除了基础设施和后勤考虑之外，还需要针对每种技术明确定义临床医生应如何应对算法护理、算法故障的责任以及临床护理中技术用户所需的监督。¹⁰如果人工智能的创新不再从算法开发延伸到护理创新，人工智能对未来临床护理的价值将受到限制。

总而言之，我们正处于人工智能的夏天。我们可以获取和处理大数据流。我们的创新环境主要集中在硬件和软件解决方案上，以解决感知和解释复杂数据的问题，但护理领域的实际变化需要解决许多关键挑战，围绕使开发与临床需求保持一致，并确保注重集成，同时提高对技术的信任经过其严谨的研发。当我们迈向数据驱动和人工智能驱动的心血管护理未来时，将发展与实施相结合的决策将使我们更接近最佳心血管护理质量和患者结果的医疗保健乌托邦。

Khera 博士得到了美国国立卫生研究院国家心肺血液研究所 (NHLBI)（奖励 K23HL153775）和 ​​Doris Duke 慈善基金会（奖励 2022060）的支持。 Wiens 博士获得美国国立卫生研究院（NHLBI、国家医学图书馆和国家老龄化研究所）、国家科学基金会、戈登和贝蒂摩尔基金会、思科研究中心、医疗保健研究和质量局、青少年糖尿病中心的支持研究基金会、卡恩基金会、阿尔弗雷德·P·斯隆基金会和密歇根州卫生与公众服务部。

除了美国国立卫生研究院和多丽丝·杜克慈善基金会之外，Khera 博士还通过耶鲁大学获得了百时美施贵宝、诺和诺德和 BridgeBio 的研究支持。他是美国临时专利申请 63/177,117、63/428,569、63/346,610、63/484,426 和 63/508,315 的共同发明人。他还共同创立了 Evidence2Health 和 Ensight-AI 健康平台，以改善基于证据的心血管护理。除了上面列出的外部机构之外，维恩斯博士还获得密歇根大学的研究支持。她还担任医疗保健机器学习董事会成员。

本文表达的观点不一定代表编辑或美国心脏协会的观点。

美国心脏协会将于 2024 年庆祝成立 100 周年。^本文是国际思想领袖撰写的整个 AHA 期刊系列文章的一部分，内容涉及心脑血管研究和护理的过去、现在和未来。要探索完整的百年纪念收藏，请访问 https://www.ahajournals.org/centennial

有关资金来源和披露信息，请参阅第 205 页。