Hilgendorf Lukas, Petursson Petur, Andersson Erik, Rawshani Aidin, Bhatt Deepak L, Råmunddal Truls, Gupta Vibha, Skoglund Kristofer, Omerovic Elmir, Sjöland Helen, Taha Amar, Kim David, Lundgren Peter, Rawshani Araz
Institute of Medicine, Department of Molecular and Clinical Medicine, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden; Wallenberg Centre for Molecular and Translational Research (WCMTM), Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden.
Institute of Medicine, Department of Molecular and Clinical Medicine, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden; Department of Cardiology, Sahlgrenska University Hospital, Gothenburg, Sweden.
JACC Adv. 2025 Jul 16;4(8):102011. doi: 10.1016/j.jacadv.2025.102011.
Rapid detection of acute myocardial infarction (AMI) reduces morbidity and mortality. Deep learning may enhance automated electrocardiogram (ECG) interpretation.
The purpose of the study was to develop and validate a deep learning model for AMI detection using ECG data, demographics, and symptoms.
This retrospective cohort study used ECG data from 2 centers in Västra Götaland County, Sweden (January 2015-June 2023), for model training and validation, with a third center for external testing. Patients with chest pain or dyspnea who received a prehospital or in-hospital ECG were included. A residual convolutional neural network was trained on ECG features, age, sex, and symptoms to predict AMI, defined by International Classification of Diseases codes at discharge. Performance was assessed using area under the receiver operating characteristic, sensitivity, and specificity.
The study included 104,507 individuals (208,366 ECGs), with 8.17% in the training set and 8.59% in the external set diagnosed with AMI. The model achieved AUROCs of 0.8221 ± 0.0101 (internal validation ± 95% CI) and 0.8314 ± 0.0085 (external validation). Performance was consistent across sex but slightly lower for ambulance-arriving patients (area under the receiver operating characteristic: 0.8081 ± 0.0095). Saliency maps highlighted focus on ST segments and T waves.
The deep learning model demonstrated strong AMI detection across diverse patient groups. A randomized trial is needed to compare its performance with emergency physicians.
急性心肌梗死（AMI）的快速检测可降低发病率和死亡率。深度学习可能会增强自动心电图（ECG）解读能力。
本研究旨在开发并验证一种利用心电图数据、人口统计学信息和症状来检测AMI的深度学习模型。
这项回顾性队列研究使用了瑞典韦斯特罗斯-哥德兰省两个中心（2015年1月至2023年6月）的心电图数据进行模型训练和验证，并在第三个中心进行外部测试。纳入了因胸痛或呼吸困难接受院前或院内心电图检查的患者。基于心电图特征、年龄、性别和症状训练了一个残差卷积神经网络，以预测AMI，AMI由出院时的国际疾病分类代码定义。使用受试者操作特征曲线下面积、敏感性和特异性评估模型性能。
该研究纳入了104,507名个体（208,366份心电图），训练集中有8.17%被诊断为AMI，外部数据集中有8.59%被诊断为AMI。该模型在内部验证中的受试者操作特征曲线下面积为0.8221±0.0101（±95%可信区间），在外部验证中的值为0.8314±0.0085。不同性别的模型性能一致，但救护车送来的患者性能略低（受试者操作特征曲线下面积：0.8081±0.0095）。显著性图突出显示了对ST段和T波的关注。
深度学习模型在不同患者群体中均表现出强大的AMI检测能力。需要进行一项随机试验，将其性能与急诊医生的性能进行比较。
