Deep Learning of ECG for the Prediction of Postoperative Atrial Fibrillation
Deep Learning of ECG for the Prediction of Postoperative Atrial Fibrillation
复制标题
心电图深度学习预测术后房颤
DOI：
10.1161/circep.122.011579
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Tsutsui Hiroyuki
中科院分区：
文献类型：
--
作者：
Tohyama Takeshi;Ide Tomomi;Ikeda Masataka;Nagata Takuya;Tagawa Koshiro;Hirose Masayuki;Funakoshi Kouta;Sakamoto Kazuo;Kishimoto Junji;Todaka Koji;Nakashima Naoki;Tsutsui Hiroyuki
术后心房颤动（POAF）是指在术后阶段诱发的心房颤动（AF）。POAF被认为是一种良性的短暂性房颤。然而，最近的研究表明，POAF与延长住院时间、发展为自发性房颤的风险以及晚期卒中的高风险相关。2因此，建立实用的POAF高危患者筛查势在必行。在此，我们开发了一种深度学习方法来预测术前心电图的POAF。本研究经九州大学医院机构审查委员会批准（批准号：2020-335），并按照《赫尔辛基宣言》进行。2015年至2020年期间在九州大学医院接受手术的患者参与了这项研究，并有机会选择退出。受试者为年龄≥18岁且术前30天记录非房颤心电图的住院患者。排除在分析之外的术前心电图如下：心房颤动、有噪声的心电图、电极移位记录失败、重复心电图、右侧心电图、心率> 120或< 40次/分钟、室性心动过速或频繁的室性早搏、不可评估或其他被认为不稳定且临床上不适合术前心电图检查的心电图。术前非房颤心电图包括：窦性心律、起搏节律和偶尔的早搏。回顾性观察术后7天12导联心电图记录，确定POAF的发生。医院内死亡被视为一种竞争风险。我们开发了一个可以处理事件时间数据的深度学习模型（图[a]）。12导联心电图波形数据（10 s）作为深度学习模型的输入。数据增强是通过随机修剪输入数据来实现的。模型的输出为一个2× 9张量，对应于事件类型（POAF或院内死亡）的数量和时间点（0-7天、8天或更晚）。为了提高准确性，我们在启用模型随机性的情况下重复了10倍的推理，并对结果进行了集成。该模型开发的第一步是执行网格搜索以识别最佳超参数。之后，我们使用137551张心电图对模型进行年龄和性别预测的预训练，这些心电图包括未术前患者，但排除了测试数据中的受试者。最后，我们使用训练数据对主模型进行微调。请参阅GitHub获取训练过程的代码（https://github）。com/Takeshi-Tohyama/DLmodel_POAF)。利用随时间变化的接收机工作特性曲线和C统计量对模型进行了评估，并利用滤波加权法的逆概率估计模型，以减少放电造成的偏差。我们分析了27 564例无房颤患者的43 980张术前心电图。这些心电图按个人基础、训练、调整和内部验证分为3个数据集，比例为7:1:2。对于内部验证，POAF的发生率为3.6%。图（B和C）显示了所开发模型预测POAF的性能，7天时的时间相关C统计量为0.83 （95% CI, 0.80-0.85）。作为参考，使用年龄和性别变量的简单网络实现了0.64。关于7的临床指标
Postoperative atrial fibrillation (POAF) is defined as the atrial fibrillation (AF) induced during the postoperative phase. POAF has been recognized as a benign and transient AF. 1 However, recent studies revealed that POAF is associated with the risk of prolonged hospitalization, development of spontaneous AF, and higher risk of stroke in the late phase. 2 Therefore, it is imperative to establish practical screening for high-risk patients with POAF. Herein, we developed a deep learning approach to predict POAF using the preoperative ECG. This study was approved by the Institutional Review Board of Kyushu University Hospital (approval number: 2020-335) and conducted in accordance with the Declaration of Helsinki. Patients who underwent a surgical operation at the Kyushu University Hospital between 2015 and 2020 participated in this study and were given the opportunity to opt out. The subjects were hospitalized patients aged≥ 18 years with non-AF ECGs recorded up to 30 days before the operation. The preoperative ECGs excluded from the analysis were as follows: AF, noisy ECGs, recording failure with dislodged electrodes, duplicated ECGs, right-sided ECGs, heart rate> 120 or< 40 beats per minute, ventricular tachycardia or frequent premature ventricular contractions, and unevaluable or others deemed unstable and clinically inappropriate for preoperative ECG tests. The preoperative non-AF ECGs included in this analysis are as follows: sinus rhythm, pacing rhythm, and occasional premature contractions. The occurrence of POAF was identified in the recordings of 12-lead ECGs within 7 days after the operation retrospectively. In-hospital death was dealt with as a competing risk. We developed a deep learning model that can handle the time-to-event data (Figure [A]). The 12-lead ECG waveform data (10 s) were used as input for the deep learning model. Data augmentation was performed by randomly trimming the input data. The output of the model was a 2× 9 tensor, corresponding to the number of event types (POAF or in-hospital death) and time points (days 0–7, day 8, or later). To improve accuracy, we repeated the inference 10× with enabled randomness of the model and ensembled the results. The first step in the development of this model was to perform a grid search to identify the best hyperparameters. Thereafter, we pretrained the model for age and sex prediction with 137 551 ECGs that included nonpreoperative patients but excluded subjects in the test data. Finally, we fine-tuned the main model using training data. See GitHub for the code of the training processes (https://github. com/Takeshi-Tohyama/DLmodel_POAF). The model was evaluated with time-dependent receiver operating characteristic curves and C statistics, which were estimated with the inverse probability of the censoring weighting approach to reduce bias due to discharge. We analyzed 43 980 preoperative ECGs from 27 564 patients without AF. These ECGs were divided into 3 data sets on an individual basis, training, tuning, and internal validation, in the ratio of 7: 1: 2. For the internal validation, the incidence of POAF was 3.6%. The Figure (B and C) exhibited the performance of the developed model for predicting POAF, and the time-dependent C statistic at 7 days was 0.83 (95% CI, 0.80–0.85). For reference, a simple network using variables of age and sex achieved 0.64. Regarding clinical metrics at 7