目的开发1项基于12导联心电图的机器学习模型预测冷冻球囊消融术后心房颤动(房颤)复发。
方法本研究为前瞻性、单中心队列研究。纳入2020年5月至2023年9月就诊于天津医科大学第二医院并接受冷冻球囊消融术的房颤患者。收集患者临床资料以及术前和术后24 h内的标准12导联心电图。利用XGBoost方法构建3种基于术前和术后心电图数据差异的人工智能心电图模型。这些模型考虑了心电图参数、心电图深度特征、早期复发和基线房颤类型等55个特征的不同组合,并对其在模型预测中的重要性进行排序。
结果共入选患者201例,年龄(68.0±9.3)岁,其中男91例(45.3%,91/201)。随访222(124,368) d,26例(12.9%,26/201)患者复发。最佳机器学习模型是以术前、术后心电图深度特征差值作为输入的模型,受试者工作特征曲线的曲线下面积( AUC)为0.872,F1得分为0.600,敏感度(召回率)为60.0%,特异度为94.4%,准确度为90.2%。
结论人工智能辅助分析心电图能够预测冷冻球囊消融术后房颤复发。
ObjectiveTo develop a machine learning model based on 12-lead electrocardiogram (ECG) to predict recurrence of atrial fibrillation (AF) after cryoballoon ablation.
MethodsIt was a prospective, single-center cohort study. Patients with AF who were admitted to The Second Hospital of Tianjin Medical University from May 2020 to September 2023 and underwent cryoballoon ablation were enrolled. Clinical data and standard 12-lead ECG within 24 hours before and after ablation were collected. Using the XGBoost method, three types of artificial intelligence (AI) ECG models were constructed based on the differences in ECG data before and after ablation. These models took into account considered different combinations of 55 features, including ECG parameters, ECG deep features, early recurrence and baseline AF types, and ranked their importance in model predictions.
ResultsA total of 201 patients were included, with an average age of (68.0±9.3) years and 91 (45.3%, 91/201) males. After 222 (124, 368) days of follow-up, there were 26 (12.9%, 26/201) patients with recurrence. The best prediction performance was obtained from the model using ECG deep features as input, with area under curve ( AUC) of 0.872, F1 score of 0.600, sensitivity (recall) of 60.0%, specificity of 94.4% and accuracy of 90.2%.
ConclusionAI algorithms can predict recurrence of AF after cryoballoon ablation.
宋文华,耿世佳,唐功政,等. 基于人工智能心电图差异预测冷冻消融术后心房颤动复发[J]. 中华心律失常学杂志,2024,28(02):139-146.
DOI:10.3760/cma.j.cn113859-20240225-00018版权归中华医学会所有。
未经授权,不得转载、摘编本刊文章,不得使用本刊的版式设计。
除非特别声明,本刊刊出的所有文章不代表中华医学会和本刊编委会的观点。
注:房颤为心房颤动
项目 | 复发组 | 非复发组 | P值 |
---|---|---|---|
例数 | 26 | 175 | |
年龄(岁,
|
67.7±11.3 | 68.1±9.1 | 0.819 |
男[例(%)] | 7(26.9) | 84(48.0) | 0.044 a |
BMI(kg/m
2,
|
24.0±3.5 | 25.3±3.7 | 0.120 |
阵发性心房颤动[例(%)] | 25(96.2) | 140(80.0) | 0.045 a |
收缩压(mmHg,
|
135.9±19.5 | 131.5±17.0 | 0.234 |
舒张压(mmHg,
|
81.1±11.1 | 80.2±11.4 | 0.706 |
合并疾病[例(%)] | |||
高血压 | 16(61.5) | 105(60.0) | 0.881 |
冠心病 | 11(42.3) | 63(36.0) | 0.534 |
糖尿病 | 5(19.2) | 32(18.3) | 0.908 |
用药史[例(%)] | |||
胺碘酮 | 8(30.8) | 72(41.1) | 0.313 |
索他洛尔 | 5(19.2) | 44(25.1) | 0.512 |
美托洛尔 | 5(19.2) | 31(17.7) | 0.033 a |
利伐沙班 | 14(53.8) | 113(64.6) | 0.290 |
艾多沙班 | 6(23.1) | 36(20.6) | 0.769 |
辅助检查 | |||
左心房内径(mm,
|
41.2±5.9 | 42.5±5.8 | 0.301 |
右心房上下径(mm,
|
47.3±7.3 | 49.4±6.1 | 0.124 |
右心房左右径(mm,
|
36.0±4.9 | 38.2±5.4 | 0.050 |
左心室射血分数(%,
|
59.2±7.7 | 60.5±6.3 | 0.334 |
左心耳最大充盈速度[cm/s, M( Q 1, Q 3)] | 51.0(36.3,65.1) | 49.0(39.7,64.0) | 0.782 |
左心耳最大排空速度[cm/s, M( Q 1, Q 3)] | 51.0(35.7,72.5) | 54.0(37.0,70.0) | 0.750 |
Hs-cTnI [ng/L, M( Q 1, Q 3)] | 0.010(0.004,0.017) | 0.015(0.010,0.024) | 0.636 |
NT-pro BNP [ng/L, M( Q 1, Q 3)] | 464.1(190.3,1 252.5) | 402.3(142.5,1 076.2) | 0.393 |
CK[U/L, M( Q 1, Q 3)] | 67.0(52.0,109.2) | 67.8(53.1,96.7) | 0.462 |
CK-MB [U/L, M( Q 1, Q 3)] | 10.9(9.3,12.5) | 11.5(9.3,14.7) | 0.185 |
早期复发[例(%)] | 8(30.8) | 22(12.6) | 0.015 a |
注:BMI为体重指数,hs-cTnI为高敏肌钙蛋白I,NT-pro BNP为N末端脑钠肽前体,CK为肌酸激酶,CK-MB为肌酸激酶同工酶;1 mmHg=0.133 kPa, a为差异具有统计学意义
心电图测量值 | 术前 | 术后 | P值 |
---|---|---|---|
心率(次/min) | 73.86±12.48 | 76.03±22.09 | 0.223 |
RR间期(ms) | 828.39±134.20 | 838.03±197.44 | 0.211 |
PR间期(ms) | 154.25±12.47 | 155.41±13.52 | 0.959 |
QRS时限(ms) | 96.71±16.66 | 96.59±16.30 | 0.949 |
QT间期(ms) | 391.51±39.95 | 381.36±35.72 | 0.049 a |
QTc间期(ms) | 431.97±38.02 | 422.72±41.26 | 0.013 a |
注:QTc间期为校正的QT间期, a为差异具有统计学意义
注:房颤为心房颤动,QTc间期为校正的QT间期
注: AUC为曲线下面积
宋文华、耿世佳:设计和实施研究、数据采集、论文撰写、论文修订、统计分析;唐功政、王悦、章德云:数据处理、模型训练、特征提取;王乔、吕童莲、刘莹、上官文锋、缪帅:数据收集;李广平:设计和实施研究、研究指导;洪申达、刘彤:设计和实施研究、论文修订、研究指导、经费支持

你好,我可以帮助您更好的了解本文,请向我提问您关注的问题。