Oikonomou Evangelos K, Batinica Bruno, Dhingra Lovedeep S, Aminorroaya Arya, Coppi Andreas, Khera Rohan
Section of Cardiovascular Medicine, Department of Internal Medicine, Yale School of Medicine, New Haven, CT, USA.
Cardiovascular Data Science (CarDS) Lab, Yale School of Medicine, New Haven, CT, USA.
medRxiv. 2025 Aug 28:2025.08.25.25334266. doi: 10.1101/2025.08.25.25334266.
Artificial intelligence (AI) applied to routine electrocardiograms (ECGs) offers promise for screening of structural heart disease (SHD), yet broad clinical integration remains limited by high false positive rates and the lack of tailored deployment strategies.
We developed TARGET-AI, a multimodal AI-enabled pipeline that integrates longitudinal electronic health record (EHR) data with ECG images to identify optimal intersections of healthcare encounters and patient phenotypes for targeted AI-ECG screening. The approach is built on (1) a foundation model pretrained on 118 million coded EHR events from 159,322 individuals to generate temporal patient embeddings and identify high-risk screening candidates, followed by (2) a contrastive vision-language model trained on 754,533 ECG-echocardiogram pairs to detect SHD with tunable performance characteristics. We evaluated this joint strategy in a temporally distinct cohort of 5,198 individuals referred for their first transthoracic echocardiogram (TTE) within 90 days of an ECG and externally in 33,518 participants from the UK Biobank undergoing ECG and cardiac magnetic resonance imaging.
Our pre-trained AI-ECG image foundation model discriminated 27 SHD subtypes, from left ventricular systolic dysfunction (AUROC of 0.90) to severe aortic stenosis (AUROC of 0.85) and elevated right ventricular systolic pressure (AUROC of 0.82). Compared with untargeted AI-ECG screening, EHR-informed TARGET-AI-guided screening significantly reduced false positive predictions across SHD labels (median reduction: 87.8%; interquartile range [IQR], 82.4%-98.2%) and improved F1 score (median increase: 0.25; IQR, 0.19-0.41). In the UK Biobank, targeted screening reduced false positives by 61.7% (IQR, 50.4%-89.1%) while preserving sensitivity.
TARGET-AI enables the context-aware deployment of AI-ECG screening by leveraging key longitudinal EHR phenotypes and multimodal ECG-echocardiogram representations, thereby defining an interoperable, data-driven strategy for the more precise deployment of AI screening tools across health systems.
应用于常规心电图（ECG）的人工智能（AI）为结构性心脏病（SHD）的筛查带来了希望，但广泛的临床应用仍受到高假阳性率和缺乏针对性部署策略的限制。
我们开发了TARGET-AI，这是一种多模态人工智能驱动的流程，它将纵向电子健康记录（EHR）数据与ECG图像相结合，以确定医疗接触和患者表型的最佳交叉点，用于有针对性的AI-ECG筛查。该方法基于以下两点构建：（1）一个基础模型，在来自159,322名个体的1.18亿条编码EHR事件上进行预训练，以生成时间患者嵌入并识别高风险筛查候选者，随后（2）一个对比视觉语言模型，在754,533对ECG-超声心动图对上进行训练，以检测具有可调性能特征的SHD。我们在一个时间上不同的队列中评估了这一联合策略，该队列中有5198名个体在ECG后90天内接受了首次经胸超声心动图（TTE）检查，并在外部对来自英国生物银行的33,518名接受ECG和心脏磁共振成像的参与者进行了评估。
我们预先训练的AI-ECG图像基础模型能够区分27种SHD亚型，从左心室收缩功能障碍（AUROC为0.90）到严重主动脉瓣狭窄（AUROC为0.85）和右心室收缩压升高（AUROC为0.82）。与无针对性的AI-ECG筛查相比，基于EHR的TARGET-AI引导筛查显著降低了SHD标签上的假阳性预测（中位数降低：87.8%；四分位间距[IQR]，82.4%-98.2%），并提高了F1分数（中位数增加：0.25；IQR，0.19-0.41）。在英国生物银行中，针对性筛查减少了61.7%的假阳性（IQR，50.4%-89.1%），同时保持了敏感性。
TARGET-AI通过利用关键的纵向EHR表型和多模态ECG-超声心动图表示，实现了AI-ECG筛查的情境感知部署，从而定义了一种可互操作、数据驱动的策略，以便在整个卫生系统中更精确地部署AI筛查工具。
