目的介绍一种基于住院患者心电图及临床特征开发的机器学习模型,用于诊断反射性晕厥。
方法入选2018年6月20日至2022年5月11日于天津医科大学第二医院心脏科住院治疗的晕厥患者,经过临床评估和调查研究获得相关基线资料。确定了晕厥患者的15个特征,并进行特征排序。采用不同的机器学习方法构建反射性晕厥的诊断模型,如Logistic回归分析、感知机、支持向量机、决策树、随机森林和K最近邻算法等方法。
结果最终入选410例患者,首次晕厥事件的年龄(64.5±14.6)岁,其中男236例(236/410,57.6%),65例患者确诊为反射性晕厥。纳入特征重要性排序结果位于前4位的特征构建模型,随机森林模型诊断反射性晕厥的性能最佳,曲线下面积为0.644,精确率(Precision)、召回率(Recall)和F1得分(F1 score)分别为0.794、0.849和0.791。
结论人工智能算法能够识别反射性晕厥,可作为一种经济有效的筛查工具。
ObjectiveTo develop a machine learning model based on electrocardiogram and clinical characteristics of hospitalized patients for identifying reflex syncope.
MethodsSyncope patients hospitalized between June 20, 2018 and May 11, 2022 were included in Department of Cardiology, the Second Hospital of Tianjin Medical University.Standardized clinical variables were collected by evaluation.Fifteen features of syncope patients were identified, and features ranking was developed.Different machine learning methods were used to construct a prediction model for reflex syncope, such as Logistic regression, perceptron, support vector machines, decision tree, random forest and K-nearest neighbor.
ResultsA total of 410 patients were enrolled, and 65 patients were diagnosed as reflex syncope.The average age of first onset was (64.5±14.6) years.There were 236 males (236/410, 57.6%). The best prediction performance was obtained by using random forest with the top 4 features derived from feature importance ranking.The algorithm showed an area under the curve of 0.644 for diagnosis of cardiac syncope with a Precision, Recall and F1 score of 0.794, 0.849 and 0.791, respectively.
ConclusionArtificial intelligence algorithms can identify reflex syncope as a cost-effective and efficient screening tool.
李歆慕,章德云,高欣怡,等. 人工智能心电图及患者特征诊断反射性晕厥[J]. 中华心律失常学杂志,2022,26(05):418-423.
DOI:10.3760/cma.j.cn113859-20220704-00131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。
重要性评分基于变量对模型性能的贡献。NT-proBNP为N末端脑钠肽前体,age为年龄,cTnI为心肌肌钙蛋白I,CK-MB为肌酸激酶同工酶,QT为QT间期,PR为PR间期,RR为RR间期,QTc为校正的QT间期,axis为QRS波电轴,HR为心率,QRS为QRS时限,post-exercise为运动后,standing为站立,GI stimulation为胃肠道刺激,others为其他诱因
特征 | 模型 | Precision | Recall | F1 score | ROC- AUC |
---|---|---|---|---|---|
Top1 | 决策树 | 0.707 | 0.841 | 0.768 | 0.544 |
K最近邻算法 | 0.760 | 0.754 | 0.756 | 0.528 | |
Logistic回归 | 0.694 | 0.707 | 0.841 | 0.768 | |
感知机 | 0.693 | 0.705 | 0.829 | 0.762 | |
随机森林 | 0.707 | 0.841 | 0.768 | 0.587 | |
支持向量机 | 0.707 | 0.841 | 0.768 | 0.412 | |
Top2 | 决策树 | 0.707 | 0.841 | 0.768 | 0.547 |
K最近邻算法 | 0.789 | 0.820 | 0.799 | 0.590 | |
Logistic回归 | 0.655 | 0.707 | 0.841 | 0.768 | |
感知机 | 0.668 | 0.705 | 0.832 | 0.763 | |
随机森林 | 0.752 | 0.841 | 0.777 | 0.655 | |
支持向量机 | 0.707 | 0.841 | 0.768 | 0.376 | |
Top3 | 决策树 | 0.707 | 0.841 | 0.768 | 0.638 |
K最近邻算法 | 0.789 | 0.820 | 0.799 | 0.591 | |
Logistic回归 | 0.651 | 0.706 | 0.838 | 0.766 | |
感知机 | 0.676 | 0.765 | 0.712 | 0.696 | |
随机森林 | 0.774 | 0.846 | 0.781 | 0.639 | |
支持向量机 | 0.636 | 0.894 | 0.743 | 0.379 | |
Top4 | 决策树 | 0.794 | 0.846 | 0.789 | 0.560 |
K最近邻算法 | 0.813 | 0.846 | 0.814 | 0.589 | |
Logistic回归 | 0.741 | 0.843 | 0.774 | 0.687 | |
感知机 | 0.704 | 0.820 | 0.758 | 0.666 | |
随机森林 | 0.794 | 0.849 | 0.791 | 0.644 | |
支持向量机 | 0.707 | 0.841 | 0.768 | 0.424 | |
Top5 | 决策树 | 0.723 | 0.831 | 0.768 | 0.613 |
K最近邻算法 | 0.771 | 0.777 | 0.772 | 0.562 | |
Logistic回归 | 0.683 | 0.777 | 0.841 | 0.786 | |
感知机 | 0.654 | 0.824 | 0.723 | 0.661 | |
随机森林 | 0.774 | 0.846 | 0.781 | 0.660 | |
支持向量机 | 0.636 | 0.894 | 0.743 | 0.443 | |
All | 决策树 | 0.812 | 0.846 | 0.794 | 0.701 |
K最近邻算法 | 0.754 | 0.791 | 0.769 | 0.502 | |
Logistic回归 | 0.649 | 0.735 | 0.826 | 0.769 | |
感知机 | 0.636 | 0.829 | 0.719 | 0.641 | |
随机森林 | 0.707 | 0.841 | 0.768 | 0.752 | |
支持向量机 | 0.638 | 0.894 | 0.745 | 0.430 |
注:Top1为特征排序结果位于首位的特征[N末端脑钠肽前体(NT-proBNP)],Top2为特征排序结果位于前2位的特征(NT-proBNP、年龄),Top3为特征排序结果位于前3位的特征[NT-proBNP、年龄、心肌肌钙蛋白I(cTnI)],Top4为特征排序结果位于前4位的特征[NT-proBNP、年龄、cTnI、肌酸激酶同工酶(CK-MB)],Top5为特征排序结果位于前5位的特征(NT-proBNP、年龄、cTnI、CK-MB、QT间期),All为纳入特征排序的所有特征,Precision为精确率,Recall为召回率,F1 score为F1得分,ROC- AUC为受试者工作特征-曲线下面积
项目 | 反射性晕厥组 | 非反射性晕厥组 | P值 | |
---|---|---|---|---|
例数 | 65 | 345 | ||
首发年龄(岁,
|
57.6±16.5 | 65.8±13.8 | <0.001 a | |
男[例(%)] | 42(64.6) | 194(56.2) | 0.210 | |
收缩压(mmHg,
|
133.4±23.1 | 131.8±23.9 | 0.624 | |
舒张压(mmHg,
|
79.3±14.0 | 78.2±34.6 | 0.812 | |
合并症[例(%)] | ||||
高血压 | 33(50.8) | 192(55.7) | 0.468 | |
冠心病 | 22(33.8) | 161(46.7) | 0.056 | |
糖尿病 | 9(13.8) | 81(23.5) | 0.085 | |
无合并症 | 12(18.5) | 26(7.5) | 0.005 a | |
其他合并症 | 28(43.1) | 152(44.1) | 0.884 | |
晕厥发作诱因[例(%)] | ||||
长时间站立 | 7(10.8) | 5(1.4) | <0.001 a | |
体位改变 | 3(4.6) | 12(3.5) | 0.716 | |
情绪性(疼痛、恐惧) | 6(9.2) | 7(2.0) | 0.002 a | |
胃肠道刺激 | 12(18.5) | 19(5.5) | <0.001 a | |
运动后 | 2(3.1) | 14(4.1) | 1.000 | |
环境闷热 | 1(1.5) | 3(0.9) | 0.500 | |
排尿 | 3(4.6) | 4(1.2) | 0.083 | |
咳嗽 | 4(6.2) | 5(1.4) | 0.039 a | |
其他诱因 | 4(6.2) | 9(2.6) | 0.135 | |
心电图参数[M( Q 1, Q 3)] | ||||
心率(次/min) | 68.0(62.0,77.5) | 71.0(61.0,85.0) | 0.856 | |
RR间期(ms) | 856.0(765.0,954.5) | 842.0(702.0,970.5) | 0.797 | |
PR间期(ms) | 167.0(149.5,175.0) | 166.0(149.5,190.0) | 0.217 | |
QRS时限(ms) | 98.0(93.0,104.0) | 100.0(92.0,111.0) | 0.243 | |
QT间期(ms) | 391.0(369.0,426.5) | 400.0(375.0,444.5) | 0.139 | |
QTc间期(ms) | 427.0(417.0,446.5) | 439.0(420.5,465.5) | 0.046 a | |
QRS轴(°) | 28.0(6.0,49.5) | 27.0(0,53.5) | 0.930 | |
实验室检查[M( Q 1, Q 3)] | ||||
cTnI(ng/L) | 0.011(0.002,0.020) | 0.015(0.010,0.060) | <0.001 a | |
NT-proBNP(ng/L) | 111.0(38.8,255.8) | 344.2(84.0,1 265.0) | <0.001 a | |
CK(U/L) | 75.0(56.8,96.6) | 73.8(51.2,133.3) | 0.485 | |
CK-MB(U/L) | 9.7(7.0,15.4) | 11.6(8.4,18.0) | 0.011 a |
注:QTc间期为校正的QT间期,cTnI为心肌肌钙蛋白I,NT-proBNP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
受试者工作特征(ROC)曲线纳入的特征数如标题所示,不同模型的ROC曲线以不同的彩色线表示。五折交叉验证的平均(Mean) AUC见ROC曲线右下角。DT为决策树,KNN为K最近邻,LR为Logistic回归,PPN为感知机,RF为随机森林,SVM为支持向量机, AUC为曲线下面积,NT-proBNP为N末端脑钠肽前体,cTnI为心肌肌钙蛋白I,CK-MB为肌酸激酶同工酶
李歆慕、章德云:设计和实施研究、数据采集、论文撰写、论文修订、统计分析;高欣怡:数据采集、论文撰写;李秀莲、梁燕:数据收集;刘文玲: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。