基于人工智能的超声心动图舒张功能评估：一项系统综述。
完整摘要 · 中文译文
对左心室舒张功能障碍和充盈压进行准确的无创评估，是心力衰竭、尤其是射血分数保留的心力衰竭（HFpEF）诊断和管理的核心。尽管经胸超声心动图被广泛使用，但基于指南的算法常常无法得出结论、可重复性不一，并且难以在常规实践中一致应用。人工智能（AI）已成为一种潜在策略，可用于标准化测量、整合多参数超声心动图数据并减少诊断不确定性。我们进行了一项诊断准确性研究的系统综述，评估将AI或机器学习方法应用于超声心动图以自动评估舒张功能或估测左心室充盈压的研究。数据库检索自建库至2026年7月7日。参考标准包括侵入性血流动力学或基于指南的舒张功能分类。使用QUADAS-3评估偏倚风险和适用性。共纳入14项研究，分别报告了开发样本以及验证或测试队列。验证范围从小的留出侵入性队列到大型独立外部数据集。针对侵入性血流动力学参考标准，报告的AUC范围为0.761至0.883，反映了血流动力学目标、阈值和验证设计的异质性。评估与基于指南或专家得出的分类的一致性以及单个舒张参数自动化复制的研究，通常报告了较高的判别 尽管这些代表了不同的临床任务，且在验证的独立性和规模上差异显著，但多项研究展示了外部或前瞻性验证，而针对有创血流动力学的稳健评估仍然有限。鉴于方法学和临床异质性较大，研究结果采用描述性综合，未计算汇总估计值。尽管回顾性研究中显示出有前景的诊断辨别能力，基于人工智能的超声心动图舒张功能评估仍受限于参考标准不一、验证方法多样，以及缺乏前瞻性干预性临床影响研究和针对有创血流动力学的有限前瞻性多中心验证，从而在现阶段阻碍了常规临床实施。
完整摘要 · 英文原文
Accurate non-invasive assessment of left ventricular diastolic dysfunction and filling pressures is central to the diagnosis and management of heart failure, particularly heart failure with preserved ejection fraction (HFpEF). Despite widespread use of transthoracic echocardiography, guideline-based algorithms are frequently indeterminate, variably reproducible, and difficult to apply consistently in routine practice. Artificial intelligence (AI) has emerged as a potential strategy to standardize measurement, integrate multiparametric echocardiographic data, and reduce diagnostic uncertainty. We performed a systematic review of diagnostic accuracy studies evaluating AI or machine-learning approaches applied to echocardiography for automated assessment of diastolic function or estimation of left ventricular filling pressures. Databases were searched from inception to July 7, 2026. Reference standards included invasive hemodynamics or guideline-based diastolic classification. Risk of bias and applicability were assessed using QUADAS-3. Fourteen studies were included, with development samples and validation or test cohorts reported separately. Validation ranged from small held-out invasive cohorts to large independent external datasets. Against invasive hemodynamic reference standards, reported AUCs ranged from 0.761 to 0.883, reflecting heterogeneous hemodynamic targets, thresholds, and validation designs. Studies evaluating agreement with guideline- or expert-derived classifications and automated replication of individual diastolic parameters generally reported high discrimination, although these represented distinct clinical tasks and varied substantially in the independence and scale of validation. Several studies demonstrated external or prospective validation, but robust evaluation against invasive hemodynamics remained limited. Given substantial methodological and clinical heterogeneity, findings were synthesized descriptively and no summary estimate was calculated. Despite promising diagnostic discrimination in retrospective studies, AI-based echocardiographic diastolic assessment remains limited by heterogeneous reference standards, variable validation approaches, and an absence of prospective interventional clinical-impact studies and limited prospective multicenter validation against invasive hemodynamics, precluding routine implementation at this stage.