LB-456090-4 NEAR-TERM PREDICTION OF LIFE-THREATENING VENTRICULAR ARRHYTHMIAS USING ARTIFICIAL INTELLIGENCE-ENABLED SINGLE LEAD AMBULATORY ECG
LB-456090-4 NEAR-TERM PREDICTION OF LIFE-THREATENING VENTRICULAR ARRHYTHMIAS USING ARTIFICIAL INTELLIGENCE-ENABLED SINGLE LEAD AMBULATORY ECG
复制标题
LB-456090-4 使用人工智能单导联动态心电图对危及生命的室性心律失常进行近期预测
DOI：
10.1016/j.hrthm.2023.04.036
复制
发表时间：
2023
期刊：
影响因子：
5.5
通讯作者：
É. Marijon
中科院分区：
文献类型：
--
作者：
L. Fiorina;Tanner Carbonati;K. Narayanan;Jia Li;C. Henry;J. Singh;É. Marijon
方法我们开发了一个基于深度学习的模型，使用前 24 小时的延长动态心电图记录来预测接下来两周内持续（≥ 30 秒）VT（集中判定）的风险。我们评估了该模型在至少 14 天持续时间的 Holter 记录中的性能，并且在前 24 小时内没有出现 VT。该模型在内部验证数据集上进行评估，并在独立数据集上进行外部验证，这两个数据集都不用于模型开发。使用室性早搏负荷、心率变异性参数 (SDNN)、患者年龄和性别进行多变量逻辑回归作为参考模型。结果我们使用在美国、英国、法国、捷克共和国、南非和印度收集的 78,294 份未经选择的动态心电图记录开发了该模型。在用于验证的 59,302 条记录中（患者平均年龄 61.3±17.3 岁，40% 为男性），222 条出现持续性 VT（平均速率 157±38 bpm，中位持续时间 62 秒 [IQR 42, 173]），其中绝大多数 (98%) 为单形性。在内部验证数据集上，该模型的 AUC 为 0.939，敏感性为 83.3%，特异性为 88.7%。在外部验证数据集上，AUC 为 0.911，敏感性和特异性分别为 78.9% 和 81.4%。该模型正确预测了 88% 快速 VT（≥ 180 bpm）动态心电图的 VT 发生情况。参考模型显示内部验证 AUC 为 0.833。
MethodsWe developed a deep learning-based model using the first 24 hours of extended Holter recordings to predict the risk of sustained (≥ 30 sec) VT (centrally adjudicated) in the following two weeks. We evaluated the performance of this model on Holter recordings of at least 14 days duration, with no VT in the first 24 hours. The model was evaluated on an internal validation dataset and externally validated on an independent dataset, both of which were not used for model development. Multivariable logistic-regression was performed as a reference model using Premature Ventricular Contraction burden, Heart Rate Variability parameter (SDNN), patient age and sex.ResultsWe developed the model using 78,294 unselected Holter recordings collected across the US, UK, France, Czech Republic, South Africa and India. Among 59,302 recordings used for validation (patients mean age 61.3±17.3 years, 40% male), 222 presented sustained VT (mean rate 157±38 bpm, median duration 62 seconds [IQR 42, 173]), with the vast majority (98%) being monomorphic. On the internal validation dataset, the model achieved an AUC of 0.939 with a sensitivity of 83.3% and a specificity of 88.7%. On the external validation dataset, the AUC was 0.911 with a sensitivity and specificity of 78.9% and 81.4%, respectively. The model correctly predicted VT occurrence in 88% of holters with rapid VT (≥ 180 bpm). The reference model revealed an internal validation AUC of 0.833.