Skip to main content
Journal of Central South University Medical Sciences logoLink to Journal of Central South University Medical Sciences
. 2023 Jul 28;48(7):995–1007. [Article in Chinese] doi: 10.11817/j.issn.1672-7347.2023.230018

基于机器学习的慢性瓣膜病合并心房颤动患者行Cox迷宫IV手术后心房颤动复发风险预测模型

Prediction model of atrial fibrillation recurrence after Cox-Maze IV procedure in patients with chronic valvular disease and atrial fibrillation based on machine learning algorithm

蒋 泽楠 1,2, 宋 珑 1, 梁 春水 2, 张 昊 1, 刘 立明 1,
Editor: 田 朴
PMCID: PMC10930048  PMID: 37724402

Abstract

目的

心房颤动(以下简称“房颤”)是一种常见的心律失常,Cox迷宫IV手术是外科治疗房颤的常用手术方法,目前Cox迷宫IV手术后患者房颤复发的风险因素尚不明确。近年来,机器学习算法在提高诊断准确率、预测患者预后和个性化治疗策略方面显示出巨大潜力。本研究旨在评估Cox迷宫IV手术治疗慢性瓣膜病合并心房颤动患者的疗效,使用机器学习算法识别心房颤动复发的潜在风险因素,构建Cox迷宫IV手术后房颤复发预测模型。

方法

回顾性纳入2012年1月至2019年12月中南大学湘雅二医院和陆军军医大学附属新桥医院符合条件的慢性瓣膜病合并房颤且行瓣膜手术合并Cox迷宫IV手术患者555例,年龄为(57.95±7.96)岁,根据患者术后房颤复发情况分为房颤复发组(n=117)和房颤未复发组(n=438)。采用Kaplan-Meier法分析窦性心律维持率,构建9个机器学习模型,包括随机森林、梯度提升决策树(gradient boosting decision tree,GBDT)、极限梯度提升(extreme gradient boosting,XGBoost)、引导聚集算法、logistic回归、类别提升(categorical boosting,CatBoost)、支持向量机、自适应增强和多层感知机。使用五折交叉验证和模型评估指标评估模型性能,评估指标包括准确度、精确度、召回率、F1分数和曲线下面积(area under the curve,AUC),筛选出2个表现最佳的模型进行进一步分析[包括特征重要性和沙普利加和解释(Shapley additive explanations,SHAP)]来识别房颤复发风险因素,以此构建房颤复发风险预测模型。

结果

患者术后5年窦性心律维持率为82.13%(95% CI 78.51%~85.93%)。9个机器学习模型中,XGBoost和CatBoost模型表现最好,AUC分别为0.768(95% CI 0.742~0.786)和0.762(95% CI 0.723~0.801),且在9个模型中有较高的准确率、精确率、召回率和F1值。特征重要性和SHAP分析显示房颤病史时长、术前左室射血分数、术后心律、术前左心房内径、术前中性粒细胞与淋巴细胞比值、术前心率和术前白细胞计数等是房颤复发的重要因素。

结论

Cox迷宫IV手术治疗房颤具有良好的窦性心律维持率,本研究通过机器学习算法成功识别多种Cox迷宫IV手术后房颤复发风险因素,成功构建2个房颤复发风险预测模型,可能有助于临床决策和优化房颤的个体化手术管理。

Keywords: Cox迷宫IV手术, 心房颤动, 机器学习, 风险因素, 预测模型


心房颤动(以下简称“房颤”)是临床常见的快速性心律失常,是成人心脏手术患者中最常见的心律失常,约占所有心律失常的三分之一,全球发病率接近1%[1-2]。房颤会增加缺血性脑卒中[3-4]、心肌梗塞[5]和肾功能不全的风险[6]

Cox迷宫手术是目前治疗房颤最有效的手术方法,是房颤外科治疗的金标准[7-8]。Cox迷宫IV手术使用双极射频消融线取代了Cox迷宫手术原本的“切割和缝合”技术[9-11]。既往研究[12-15]确定了影响Cox迷宫手术后房颤复发的一些风险因素,包括扩大的左心房内径(left atrial diameter,LAD)、较长的瓣膜病病史时长、冠状动脉疾病和较大的右心房内径等。但目前仍缺乏大样本、长期队列研究,Cox迷宫IV手术后房颤复发的风险因素有待进一步验证和评估。确定影响Cox迷宫IV手术疗效的风险因素,构建房颤复发风险预测模型,则有可能进一步提高Cox迷宫IV手术治疗房颤的有效性。以往的研究[14-15]主要采用传统的统计分析方法,如线性回归和logistic回归、线性判别分析等来研究变量之间的关系和分析临床数据,这些方法在处理大量数据和探索非线性关系的能力上存在局限性[16-17]。目前,机器学习越来越多地融入临床实践,应用范围从临床前数据处理到患者分层和治疗决策[18],主要包括疾病诊断、治疗风险评估、药物生产和医学数据分析等[18-21]。本研究通过回顾性队列研究评估Cox迷宫IV手术后慢性瓣膜病合并房颤患者的房颤复发情况,并通过机器学习模型算法探究Cox迷宫IV手术后房颤复发的风险因素,构建房颤复发预测模型。

1. 资料与方法

1.1. 资料与分组

回顾性纳入2012年1月至2019年12月在中南大学湘雅二医院和陆军军医大学新桥医院就诊的慢性瓣膜病合并房颤且行瓣膜手术合并Cox迷宫IV手术的555例患者。纳入标准:年龄35~70岁;术前心电图或24 h动态心电图确诊为房颤;术前射血分数≥40%;术前LAD≤75 mm;在2所医院进行瓣膜手术合并Cox迷宫IV手术,术后规律复诊。排除标准:同期行冠脉搭桥患者;急诊手术患者;住院/门诊电子病历及随访数据不完整或缺失值过多的患者。根据患者术后房颤复发情况,将患者分为房颤未复发组(n=438)和房颤复发组(n=117)。本研究已通过中南大学湘雅二医院和陆军军医大学附属新桥医院伦理委员会批准,伦理审批号为(2019)伦审[科](054)号及2019-研第056-01号,且获得所有患者的知情同意。

1.2. 房颤复发的定义和随访记录

房颤复发是指心电图(electrocardiogram,ECG)记录的房颤、心房扑动和房性心动过速持续发生超过30 s。非房颤复发的情况包括窦性心律、结性心律、房性早搏和室性早搏。诊断记录:体格检查及听诊、ECG、24 h动态ECG及相关病历。纳入患者的术后随访期从术后6个月开始,至少持续12个月。患者术后每12个月进行一次随访或登记复诊数据。随访记录症状和体征,并进行心脏超声、ECG等常规检查。对因个人原因无法到院复诊的患者,通过电话联系,建议在当地医院完成检查,并将结果报告给随诊记录员。如果患者出现心悸或疑似房颤复发,应随时进行ECG检查,并将结果报告给随访记录员。本研究的2个中心均按照统一的标准进行数据收集、数据质控,并统一进行数据输入和存储,保证数据的同质性。

1.3. 研究变量和数据预处理

根据住院数据、门诊数据和随访数据选择58个特征作为变量,包括人口统计学、病史、实验室检查结果和围手术期临床相关变量。数据从患者的门诊及住院电子病历、就诊记录和随访记录中收集,并进行数据清理和预处理。使用数字特征的中值和分类特征的模式来检查和估算缺失值,使用对数转换检查异常值和转换偏斜特征,以保证数据的质量和一致性。

1.4. 机器学习模型建立与模型评估

分别构建9种常见的机器学习模型,包括支持向量机(support vector machine,SVM)[22]、logistic回归[23]、极限梯度提升(extreme gradient boosting,XGBoost)[24]、随机森林(random forest,RF)[25]、类别提升(categorical boosting,CatBoost)[26]、自适应提升(adaptive boosting,AdaBoost)[27]、引导聚集算法(bootstrap aggregating,Bagging)[28]、梯度提升决策树(gradient boosting decision tree,GBDT)[29]以及多层感知机(multilayer perceptron,MLP)[30]。使用五折交叉验证(cross-validation,CV),将训练数据分成5份,即555例患者的临床数据集被分为5份,每份数据都在上述机器学习模型中进行训练和测试,每次选择其中一份作为验证集,剩余4份作为训练集,进行模型训练和评估。最终模型评估结果为5次训练结果的平均值,以更准确地评估模型的泛化性能。通过网格搜索,确定机器学习模型中最优的超参数组合。计算F1分数、准确率、召回率、精确率和曲线下面积(area under the curve,AUC)以及95% CI来评估这些模型的性能。通过上述指标可评估模型的整体性能并确定它们在临床决策中的适用性。F1分数是精确率和召回率的调和平均值,是平衡这2个指标之间的权衡的有用指标。准确率衡量正确预测的百分比;召回率衡量所有实际阳性案例中真阳性预测的百分比;精确率衡量所有正预测中真正正预测的百分比。按7꞉3分割训练集和测试集,绘制每个模型的接受者操作特征(receiver operating characteristic,ROC)曲线,直观显示各模型的AUC值差异[31]

1.5. 特征重要性和模型可解释性分析

根据模型评估的结果筛选出2种表现最佳的机器学习模型。对所选模型进行特征重要性分析,以确定特征对预测的影响,每个特征的重要性根据模型的内部机制来计算。本研究分析2个模型中的前20个特征,分析其特征重要性并根据它们的重要性进行排名,使用特征重要性图进行可视化。此外,采用沙普利加和解释(Shapley additive explanations,SHAP)来分析和解释机器学习模型结果。SHAP是一种用来解释机器学习模型预测结果的工具,可以更直观地展示每个特征对模型结果的贡献方向以及各特征的总贡献,还可通过SHAP的单个样本实例图来了解每个样本中各特征对模型预测的贡献,从而了解各特征对房颤复发的正负贡献及SHAP值的大小及范围,通过观察SHAP值的分布情况,可以发现哪些特征对预测结果有显著影响。通过以上方法构建Cox迷宫IV术后房颤复发风险预测模型。

1.6. 统计学处理

使用SPSS 27.0、Python 3.8、Scikit-learn(Sklearn) library 0.23.2和R4.0.2进行统计分析。符合正态分布的计量资料以均数±标准差( x¯ ±s)表示,采用t检验进行组间比较;非正态分布的计量资料以中位数(第1四分位数,第3四分位数)[M(P 25, P 75)]表示,采用非参数检验进行组间比较;计数资料采用例数和百分比(%)表示。分类数据之间的比较采用χ2检验。检验水准α=0.05,P<0.05为差异有统计学意义。

2. 结 果

2.1. 基线特征表

2组患者的住院基线特征如表1所示。所有患者中位随访时间为5年,年龄(57.95±7.96)岁。2组患者房颤病史时长、LAD、术前左室射血分数(left ventricular ejection fraction,LVEF)、术后心律等特征的差异均有统计学意义(均P<0.05,表1)。

表1.

患者的基线资料

Table 1 Baseline characteristics of the patients

组别 n 女性/[例(%)] 年龄/岁 房颤病史时长/年 身高/cm 体重/kg 收缩压/mmHg
房颤未复发组 438 297(67.8) 57.72±8.02 3.3(2.0, 4.3) 158.45±7.98 58.55±10.34 113.70±16.07
房颤复发组 117 73(62.4) 58.80±7.67 4.8(3.7, 6.1) 160.47±8.14 59.49±10.13 114.57±16.49
合计 555 370(66.7) 57.95±7.97 3.8(2.8, 5.0) 158.87±8.06 58.74±10.30 113.88±16.16
t/χ 2 0.921 -1.177 -6.247 -2.502 -0.876 -0.462
P 0.337 0.195 <0.001 0.016 0.381 0.607
组别 舒张压/mmHg BMI/(kg·m-2) 术前心率/(次·min-1) 术前LAD/mm 术前RAD/mm 术前LVD/mm
房颤未复发组 71.61±11.05 23.25±3.21 91.93±20.70 51.22±7.71 38.54±6.13 49.93±8.19
房颤复发组 72.00±11.00 22.99±2.76 87.77±18.35 55.19±10.40 40.46±7.81 49.93±7.85
合计 71.69±11.04 23.19±3.13 91.06±20.30 52.05±8.50 38.94±6.56 49.93±8.12
t/χ 2 -0.422 0.825 1.989 -3.776 -2.357 0.047
P 0.737 0.443 0.049 <0.001 0.016 0.997
组别 术前RVD/mm 术前LVEF/% 术前高血压/[例(%)]

术前糖尿病/

[例(%)]

术前冠心病/

[例(%)]

术前脑梗史/

[例(%)]

房颤未复发组 36.72±5.85 62.08±8.33 40(9.1) 11(2.5) 30(6.8) 27(6.2)
房颤复发组 37.56±6.60 59.03±6.70 10(8.6) 6(5.1) 12(10.3) 7(6.0)
合计 36.90±6.02 61.44±8.11 50(9.0) 17(3.0) 42(7.6) 34(6.1)
t/χ 2 -1.303 3.637 0.039 2.129 1.532 0.005
P 0.181 <0.001 0.870 0.138 0.203 0.866
组别

术前肺动脉

高压/[例(%)]

吸烟或饮酒/

[例(%)]

HAS-BLED评分/[例(%)]
0 1 2 4
房颤未复发组 121(27.6) 18(4.1) 300(68.5) 122(27.9) 15(3.4) 1(0.2)
房颤复发组 28(23.9) 6(5.1) 90(76.9) 24(20.5) 2(1.7) 1(0.9)
合计 149(26.8) 24(4.3) 390(70.3) 146(26.3) 17(14.5) 2(1.7)
t/χ 2 0.642 -0.167 1.384
P 0.459 0.614 0.149
组别 CHA2DS2-VASc评分/[例(%)]
0 1 2 3 4 5
房颤未复发组 109(24.9) 262(59.8) 35(8.0) 30(6.8) 1(0.2) 1(0.2)
房颤复发组 33(28.2) 72(61.5) 5(4.3) 5(4.3) 1(0.9) 1(0.9)
合计 142(25.6) 334(60.2) 40(7.2) 35(6.3) 2(0.4) 2(0.4)
t/χ 2 1.022
P 0.395
组别 EuroScore II评分 术前WBC/(×109·L-1) 术前中性粒细胞比例/% 术前红细胞/(×1012·L-1)

术前血红

蛋白/(g·L-1)

术前中性粒

细胞/(×109·L-1)

术前淋巴细胞/(×109·L-1)
房颤未复发组 1.64±0.978 6.37±2.02 60.34±10.19 4.78±1.32 129.95±15.62 4.38±1.73 1.90±0.83
房颤复发组 1.61±0.77 5.95±1.57 60.06±10.02 4.67±1.39 128.52±16.50 4.66±1.61 1.76±0.75
合计 1.63±0.94 6.28±1.94 60.28±10.15 4.75±1.33 129.65±15.82 4.44±1.71 1.87±0.82
t/χ 2 0.301 2.187 0.218 0.783 0.764 -1.748 1.685
P 0.787 0.017 0.792 0.434 0.386 0.124 0.093
组别 术前NLR

术前PLT/

(×109·L-1)

术前INR 术前PT/s 术前AST/(U·L-1) 术前ALT/(U·L-1)
房颤未复发组 2.30(1.58, 3.25) 190.63±72.90 1.13±0.38 12.98±4.54 28.40(20.66, 49.99) 26.40(17.20, 40.80)
房颤复发组 2.55(1.84, 3.90) 195.21±79.68 1.14±0.28 12.99±3.13 32.63(21.60, 53.52) 26.40(18.700, 45.20)
合计 2.34(1.65, 3.42) 191.59±74.39 1.13±0.36 12.99±4.28 29.90(20.90, 50.70) 26.40(17.30, 42.20)
t/χ 2 -2.816 -0.534 -0.160 -0.057 -1.485 -1.059
P 0.005 0.557 0.878 0.988 0.138 0.290
组别 术前肌酐/(µmol·L-1) 术前尿素氮/(mmol·L-1)

术前TBIL/

(μmol·L-1)

术前NYHA分级/[例(%)]
I&II III IV
房颤未复发组 72.10(59.50, 90.30) 6.23±2.58 14.90(10.60, 21.80) 25(5.7) 355(81.1) 58(13.2)
房颤复发组 77.60(63.80, 92.00) 6.29±2.76 14.20(10.20, 19.10) 7(6.0) 92(78.6) 18(15.4)
合计 72.80(60.20, 90.80) 6.24±2.62 14.70(10.60, 21.00) 32(5.8) 447(80.5) 76(13.7)
t/χ 2 -1.318 -0.205 0.865 0.389
P 0.187 0.81 0.387 0.797
组别 主手术类型/[例(%)]

合并MVP手术/

[例(%)]

合并TVP手术/[例(%)]
AVR DVR MVR 单纯TVP 单纯MVP 单纯TVR
房颤未复发组 6(1.4) 185(42.2) 230(52.5) 1(0.2) 15(3.4) 1(0.2) 18(4.1) 379(86.3)
房颤复发组 0(0) 44(37.6) 69(59.0) 2(1.7) 2(1.7) 0(0) 2(1.7) 100(86.2)
合计 6(1.1) 229(41.3) 299(53.9) 3(0.5) 17(3.1) 1(0.2) 20(3.6) 479(86.3)
t/χ 2 -0.991 1.531 -0.004
P 0.293 0.222 0.972
组别 合并TVR手术/[例(%)] 体外循环时间/min 主动脉阻断时间/min 左心房血栓/[例(%)] 术后非窦性心律/[例(%)]

术后心率/

(次·min-1)

房颤未复发组 4(0.9) 98(80, 120) 64(48, 77) 39(8.9) 107(24.4) 84.89±14.73
房颤复发组 0(0) 101(81, 120) 63(48, 80) 13(11.1) 61(52.1) 84.37±15.66
合计 4(0.7) 99(80, 120) 63(48, 78) 52(9.4) 168(30.3) 84.78±14.93
t/χ 2 0.152 -0.646 -0.480 -0.547 -4.607 0.325
P 0.304 0.475 0.605 0.445 <0.001 0.738
组别 术后LAD/mm 术后RAD/mm 术后LVD/mm 术后RVD/mm 术后LVEF/%

ICU住院

时间/h

IABP或

ECMO/[例(%)]

房颤未复发组 42.14±6.35 35.46±4.31 45.47±5.60 34.50±3.64 64.82±7.18 39.81±12.60 1(0.2)
房颤复发组 45.34±8.14 37.30±5.88 46.36±5.88 35.76±5.12 64.33±6.88 38.80±12.44 0(0)
合计 42.81±6.89 35.84±4.74 45.65±5.67 34.77±4.03 64.72±7.12 39.60±12.57 1(0.2)
t/χ 2 -3.972 -3.685 -1.504 -3.010 0.713 0.911 0.038
P <0.001 <0.001 0.134 0.003 0.509 0.443 0.607
组别

心脏复律/

[例(%)]

永久起搏器/[例(%)]

血液透析/

[例(%)]

术后AKI/

[例(%)]

术后脑梗塞/

[例(%)]

术后住院

时间/d

随访时长/年
房颤未复发组 8(1.8) 1(0.2) 1(0.2) 1(0.2) 1(0.2) 11(8, 15) 4(3, 6)
房颤复发组 3(2.6) 0(0) 0(0) 0(0) 0(0) 13(9, 18) 5(3, 6)
合计 11(2.0) 1(0.2) 1(0.2) 1(0.2) 1(0.2) 11(8, 16) 5(3, 6)
t/χ 2 -0.123 0.038 0.038 0.038 0.038 -2.492 -1.206
P 0.600 0.607 0.607 0.607 0.607 0.007 0.222

1 mmHg=0.133 kPa。正态分布的数据采用均数±标准差表示,非正态分布的数据采用M(P 25, P 75)表示。BMI:体重指数;LAD:左心房内径;RAD:右心房内径;LVD:左心室内径;RVD:右心室内径;LVEF:左室射血分数;PHTN:肺动脉高压;NLR:中性粒细胞与淋巴细胞比值;WBC:白细胞计数;PLT:血小板;INR:国际标准化比值;PT:凝血酶原时间;AST:天冬氨酸转氨酶;ALT:谷丙转氨酶;TBIL:总胆红素;NYHA:纽约心脏协会分类;AVR:主动脉瓣置换术;MVP:二尖瓣成形术;TVP:三尖瓣成形术;TVR:三尖瓣置换术;ICU:重症监护室;IABP:主动脉内球囊反搏;ECMO:体外膜肺氧合;AKI:急性肾损伤。

2.2. 窦性心律维持率曲线

通过Kaplan-Meier 方法分析绘制随访队列的窦性心律维持率曲线(图1)。患者术后5年窦性心律维持率为82.13%(95% CI 78.51%~85.93%)。

图1.

图1

窦性心律维持率曲线

Figure 1 Sinus rhythm maintenance rate curve

2.3. 术后房颤复发预测模型的构建

五折交叉验证评估机器学习模型的结果如表2所示。绘制模型使用7꞉3比例划分训练集和测试集的ROC曲线(图2)。ROC曲线和模型评估结果显示: 9个模型中XGboost、CatBoost模型的AUC值最高,分别为0.768(95% CI 0.742~0.786)和0.762(95% CI 0.723~0.801),均优于传统logistic回归模型。

表2.

模型性能评估表

Table 2 Performance summary of machine learning models

模型 准确率(95% CI) 精确率(95% CI) 召回率(95% CI) F1(95% CI) AUC(95% CI)
XGBoost 0.802(0.760~0.844) 0.799(0.782~0.816) 0.633(0.551~0.772) 0.706(0.664~0.748) 0.768(0.742~0.786)
CatBoost 0.807(0.729~0.861) 0.764(0.716~0.812) 0.631(0.532~0.778) 0.697(0.647~0.747) 0.762(0.723~0.801)
RF 0.788(0.769~0.802) 0.736(0.708~0.764) 0.612(0.598~0.621) 0.671(0.656~0.687) 0.732(0.701~0.763)
GBDT 0.793(0.785~0.801) 0.717(0.632~0.802) 0.534(0.503~0.579) 0.611(0.566~0.655) 0.702(0.665~0.739)
Bagging 0.796(0.744~0.859) 0.716(0.688~0.744) 0.522(0.387~0.591) 0.615(0.575~0.656) 0.698(0.652~0.744)
Logistic回归 0.777(0.732~0.811) 0.689(0.505~0.819) 0.494(0.405~0.619) 0.588(0.450~0.702) 0.688(0.664~0.712)
SVM 0.795(0.758~0.836) 0.612(0.568~0.656) 0.512(0.459~0.565) 0.561(0.512~0.609) 0.687(0.646~0.728)
AdaBoost 0.773(0.702~0.812) 0.698(0.601~0.795) 0.533(0.501~0.565) 0.608(0.552~0.665) 0.668(0.624~0.712)
MLP 0.744(0.688~0.805) 0.637(0.459~0.781) 0.515(0.488~0.542) 0.575(0.488~0.656) 0.658(0.616~0.691)

XGBoost:极限梯度提升;CatBoost:类别提升;RF:随机森林;GBDT:梯度提升决策树;Bagging:引导聚集算法;SVM:支持向量机;AdaBoost:自适应提升; MLP:多层感知机;AUC:曲线下面积。

图2.

图2

9个机器学习模型的受试者操作特征曲线

Figure 2 Receiver operating characteristic (ROC) curves of the 9 machine learning models

SVM: Support vector machine; LR: Logistic regression; RF: Random forest; XGBoost: Extreme gradient boosting; CatBoost: Categorical boosting; AdaBoost: Adaptive boosting; Bagging: Bootstrap aggregation; GBDT: Gradient boosting decision tree; MLP: Multilayer perceptron; AUC: Area under the curve.

2.4. 特征重要性和模型可解释性分析

经过模型评估筛选出2种机器学习模型(XGBoost和CatBoost),2个模型内置特征重要性分析的前20个特征及特征重要性如图3所示,根据特征排序、数值以及在2个模型中的出现频次,可发现房颤病史时长、术前LVEF、术后心律、术前LAD、术前中性粒细胞与淋巴细胞比值(neutrophil-to-lymphocyte ratio,NLR)、术前心率和术前白细胞计数是房颤复发的重要风险因素,其他可能的风险因素还包括术前红细胞计数和术后右心房内径等。通过SHAP可解释性分析生成XGBoost模型和Catboost模型的SHAP总结图(图4)和SHAP实例图(图5)。对于房颤复发预测而言,红色分布区域样本表示可能增加房颤复发风险,如房颤病史时长增加、术后非窦性心律等特征对房颤复发有正向贡献,房颤复发风险增加。蓝色分布区域样本表示可能降低房颤复发风险,即LVEF、术前心率等特征对房颤复发有负向贡献,房颤复发风险降低。图5分别显示Cox迷宫IV手术后中基于XGBoost模型和CatBoost模型的各样本房颤复发高风险示例和低风险示例,可看到选定的单个患者样本中各个特征对于该患者房颤复发的贡献值和影响方向(包括正向影响和负向影响)以及预测值,从而判定该患者的房颤复发风险。f(x)值表示给定样本的房颤复发风险预测结果,低于预期值表示房颤复发风险降低,高于预期值表示房颤复发风险升高。每个特征箭头大小对应权重值,表示该特征对预测结果的影响程度。红色特征及红色箭头表示对房颤复发有正向贡献,蓝色特征及蓝色箭头表示对房颤复发有负向贡献。权重值的大小表示特征对该样本预测影响的大小。图5B中的f(x)值为0.81,位于预期值-1.412的右侧,高于预期值,表明该样本为高房颤复发风险样本,且该患者显示影响房颤复发结果权重值高的红色特征(包括术后心律为非窦性、术前NLR高、术前LVEF低等)和蓝色特征(术前LAD低),据此可以评估每个患者的房颤复发风险及主要影响房颤复发风险的特征。

图3.

图3

基于XGBoostCatBoost模型的特征重要性图

Figure 3 Feature importance plots based on the XGBoost and CatBoost models

A: Feature importance plot based on the XGBoost model. B: Feature importance plot based on the CatBoost model. A higher value on the X-axis indicates a higher importance of the feature in the corresponding model. AF: Atrial fibrillation; LVEF: Left ventricular ejection fraction; NLR: Neutrophil-lymphocyte ratio; LAD: Left atrial diameter; RAD: Right atrial diameter; WBC: White blood cell count; LOS: Length of hospital stay; PT: Prothrombin time; PHTN: Pulmonary hypertension; CBP: Cardiopulmonary bypass time; ALT: Alanine transaminase; TBIL: Total bilirubin; NEUT%: Percentage of neutrophils; PLT: Platelets.

图4.

图4

基于XGBoost模型(A)CatBoost模型(B)SHAP总结图

Figure 4 Summary plots of SHAP values based on the XGBoost (A) and CatBoost (B) models

The middle part of the plot is a cluster of colored dots, where red dots represent features that have a positive impact on the model output, blue dots represent features that have a negative impact, and purple dots represent features that have a neutral impact. The size of the dots represents the magnitude of the feature’s impact, with larger dots indicating a greater impact. AF: Atrial fibrillation; LVEF: Left ventricular ejection fraction; NLR: Neutrophil-lymphocyte ratio; LAD: Left atrial diameter; RAD: Right atrial diameter; WBC: White blood cell count; RBC: Red blood cell count; LOS: Length of hospital stay; PT: Prothrombin time; CBP: Cardiopulmonary bypass time; ALT: Alanine transaminase; TBIL: Total bilirubin; NEUT%: Percentage of neutrophils.

图5.

图5

对应于特定实例的各模型预测房颤复发风险评分的SHAP实例图

Figure 5 SHAP force plots of the predicted atrial fibrillation recurrence risk scores for each model corresponding to a specific instance

A and B: SHAP force plot of a low-risk example (A) and a high-risk example (B) based on the XGBoost model; C and D: SHAP force plot of a low-risk example (C) and a high-risk example (D) based on the CatBoost model. Red features and arrows indicate a positive contribution to atrial fibrillation recurrence, while blue features and arrows indicate a negative contribution. The f(x) value represents the predicted atrial fibrillation recurrence risk for a given sample, lower than the expected value indicating a decreased risk, and higher than the expected value indicating an increased risk. LEVF: Left ventricular ejection fraction; NLR: Neutrophil-lymphocyte ratio; AST: Aspartate aminotransferase; LOS: Length of hospital stay; BUN: Blood urea nitrogen; LVD: Left ventricle diameter; LAD: Left atrial diameter; TBIL: Total bilirubin.

3. 讨 论

本研究对瓣膜手术合并Cox迷宫IV手术的患者进行了回顾性队列研究,根据随访结果绘制出的患者的窦性心律维持率,表明Cox迷宫IV手术治疗房颤效果明显,这与既往的文献[32-33]报道一致。

本研究成功构建9个机器学习模型,其中XGBoost和CatBoost在预测和分类该数据集方面优于其他方法,模型评估展现出良好的性能。使用特征重要性分析和SHAP可解释性分析,根据每个模型中的重要性和出现次数对获得的特征进行排序,确定了一些房颤复发的风险因素。其中,最重要的风险因素是房颤病史时长和术前LVEF,其他重要的风险因素包括术后心律、术前LAD、术前NLR、术前心率和术前白细胞计数。

2个机器学习模型均发现房颤病史时长是影响房颤复发最重要的特征,这与Chew等[34]的荟萃分析结果相似,此研究发现首次诊断房颤和消融手术之间的持续时间,与导管消融后的房颤复发有关。Andrade等[35]还发现房颤病史时长与房颤预后结果相关,较长的房颤病史时长与较高的房颤负荷有关。本研究发现房颤复发的风险随着房颤病史时长的延长而显著增加,术前LVEF也是房颤复发的重要风险因素,2个模型的SHAP总结图显示:随着术前LVEF的降低,房颤复发的风险增加,其特征重要性在2个模型中排名第二,LVEF对房颤复发有负相关性,且具有重要影响。Lerman等[36]的研究发现:心衰患者的房颤发生率显著高于非心衰患者(39.6% vs 8.8%)。但总体人群的LVEF与房颤的相关性在既往研究中未被发现,需要进一步的研究来证实LVEF在房颤复发中的作用。本研究构建的模型还发现LAD与房颤复发密切相关,通过SHAP解释器可发现房颤复发的风险随着术前或术后LAD的增加而增加,这与既往的研究[37-39]一致。XGBoost和Catboost模型的特征重要性分析发现联合三尖瓣手术也是房颤复发的风险因素,既往研究[40-41]发现长期房颤与功能性三尖瓣反流相关,重度三尖瓣反流可能与房颤相关,这提示术前三尖瓣关闭不全可能与房颤复发风险相关。目前已发现白细胞计数与心血管风险相关[42],本研究还发现术前NLR与Cox迷宫IV手术后房颤复发显著相关。NLR与心血管风险有关,可以预测各种临床疾病的预后[43-44],包括高血压和心力衰竭[45-46]。高NLR可能被用作房颤复发的预测因子[47-48],这与本研究中机器学习模型的结果一致,CatBoost和XGBoost的SHAP总结图显示随着NLR增加,房颤复发的风险也会增加。此外,本研究还发现术后早期房颤复发与最终房颤复发相关。具体而言,术后早期房颤复发可能与术后过度炎症反应有关[49-50],显著增加晚期房颤复发的风险,这一发现与Kim等[51]对射频导管消融的研究一致。凝血酶原时间和国际标准化比率用于衡量房颤患者使用抗凝剂的有效性,在本研究中不是显著的房颤复发风险因素。结合2个模型的分析结果,本研究未发现常用的手术评分工具包括HAS-BLED评分[52],EuroScore II评分[53]和CHA2DS2-VASc评分[54]与房颤复发风险的显著相关性。此外,性别、年龄和瓣膜手术的主要类型也未被识别为重要的风险因素。

本研究发现:2个模型在识别房颤复发最有影响的风险因素方面具有高度一致性,但模型之间特征重要性较低的风险因素的结果排序存在差异。此外,XGBoost、Catboost模型内置的特征重要性和SHAP分析特征重要性的计算方法有差异,SHAP考虑所有特征的交互作用,而XGBoost和CatBoost模型内置的特征重要性为每个特征单独对预测结果的贡献,因此模型内置的特征重要性分析与SHAP分析的特征重要性结果存在部分差异[55-56]。本研究结果表明机器学习模型在识别关键风险因素方面是有效的,但在评估整体风险因素方面可能还有进一步优化的空间,图5中基于每个样本的SHAP实例模型,可作为可视化房颤复发风险预测模型,可以展示每个患者的房颤复发风险以及权重,它有助于对每个患者进行房颤复发的个体化风险评估。既往已有多个Cox迷宫手术后房颤复发风险因素及预测模型的研究[14-15, 57],但均基于传统统计分析和模型研究,本研究较以往研究扩大了样本量和随访时长,并使用机器学习算法成功建立Cox迷宫IV手术后房颤复发风险预测模型,可视化地针对研究总体和个体患者分别展现了房颤复发风险因素和预测模型,且本研究筛选出的2个模型性能均优于传统logistic回归模型。

本研究发现Cox迷宫IV手术可有效治疗房颤并维持房颤和瓣膜病患者的窦性心律,成功开发多个机器学习模型,并确定影响房颤复发的几个风险因素,其中最重要的风险因素是房颤病史时长和术前LVEF,其他重要的风险因素包括术后心律、术前LAD、术前NLR、术前心率和术前白细胞计数等。在本数据集中,XGBoost模型和CatBoost模型预测房颤复发性能表现最佳。本研究证实机器学习可用于识别Cox迷宫手术后房颤复发风险因素,构建房颤复发风险预测模型,这可能有助于在临床实践中预测和预防房颤复发,从而进一步改善患者预后和手术疗效。

本研究存在局限性:首先,样本量可能不足以完全涵盖房颤复发的复杂性。其次,研究是在2所医院进行的,可能无法推广到其他环境。最后,机器学习技术的有效性取决于输入特征的质量和相关性,本研究仅考虑有限数量的特征,未进行更详细的术中数据分析或评估患者消融的透壁性,故可能还有其他尚未确定的房颤复发风险因素。今后可继续用更大规模的数据集和特征种类研究房颤复发的个体化风险因素,进一步提升模型预测效能,可开发基于数据自动更新的个体化的房颤复发评估系统,以帮助外科医师根据患者的个体数据评估和改善房颤手术治疗的预后。

基金资助

国家重点研发计划(2018YFC1311204)。

This work was supported by the National Key Research and Development Program of China (2018YFC1311204).

利益冲突声明

作者声称无任何利益冲突。

作者贡献

蒋泽楠 研究设计,数据分析,模型构建和评估,论文撰写;宋珑、梁春水、张昊 病例数据收集,图表分析;刘立明 项目管理,研究设计,论文审阅和修改。所有作者阅读并同意最终的文本。

原文网址

http://xbyxb.csu.edu.cn/xbwk/fileup/PDF/202307995.pdf

参考文献

  • 1. Gillinov AM, Bagiella E, Moskowitz AJ, et al. Rate control versus rhythm control for atrial fibrillation after cardiac surgery[J]. N Engl J Med, 2016, 374(20): 1911-1921. 10.1056/NEJMoa1602002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Miyasaka Y, Barnes ME, Gersh BJ, et al. Secular trends in incidence of atrial fibrillation in Olmsted County, Minnesota, 1980 to 2000, and implications on the projections for future prevalence[J]. Circulation, 2006, 114(2): 119-125. 10.1161/CIRCULATIONAHA.105.595140. [DOI] [PubMed] [Google Scholar]
  • 3. Paciaroni M, Agnelli G, Micheli S, et al. Efficacy and safety of anticoagulant treatment in acute cardioembolic stroke: a meta-analysis of randomized controlled trials[J]. Stroke, 2007, 38(2): 423-430. 10.1161/01.STR.0000254600.92975.1f. [DOI] [PubMed] [Google Scholar]
  • 4. von Kummer R, Broderick JP, Campbell BC, et al. The Heidelberg bleeding classification: classification of bleeding events after ischemic stroke and reperfusion therapy[J]. Stroke, 2015, 46(10): 2981-2986. 10.1161/STROKEAHA.115.010049. [DOI] [PubMed] [Google Scholar]
  • 5. Ruddox V, Sandven I, Munkhaugen J, et al. Atrial fibrillation and the risk for myocardial infarction, all-cause mortality and heart failure: a systematic review and meta-analysis[J]. Eur J Prev Cardiol, 2017, 24(14): 1555-1566. 10.1177/2047487317715769. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Böhm M, Ezekowitz MD, Connolly SJ, et al. Changes in renal function in patients with atrial fibrillation: an analysis from the RE-LY trial[J]. J Am Coll Cardiol, 2015, 65(23): 2481-2493. 10.1016/j.jacc.2015.03.577. [DOI] [PubMed] [Google Scholar]
  • 7. García-Villarreal OA. The Cox-maze procedure: rules not to be broken[J]. Ann Thorac Surg, 2021, 112(6): 2109. 10.1016/j.athoracsur.2021.02.026. [DOI] [PubMed] [Google Scholar]
  • 8. Ruaengsri C, Schill MR, Khiabani AJ, et al. The Cox-maze IV procedure in its second decade: still the gold standard?[J]. Eur J Cardiothorac Surg, 2018, 53(suppl_1): 19-25. 10.1093/ejcts/ezx326. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Ad N, Henry L, Hunt S, et al. The outcome of the Cox Maze procedure in patients with previous percutaneous catheter ablation to treat atrial fibrillation[J]. Ann Thorac Surg, 2011, 91(5): 1371-1377; discussion1377. 10.1016/j.athoracsur.2011.01.026. [DOI] [PubMed] [Google Scholar]
  • 10. Robertson JO, Saint LL, Leidenfrost JE, et al. Illustrated techniques for performing the Cox-Maze IV procedure through a right mini-thoracotomy[J]. Ann Cardiothorac Surg, 2014, 3(1): 105-116. 10.3978/j.issn.2225-319X.2013.12.11. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Calkins H, Kuck KH, Cappato R, et al. 2012 HRS/EHRA/ECAS Expert Consensus Statement on Catheter and Surgical Ablation of Atrial Fibrillation: recommendations for patient selection, procedural techniques, patient management and follow-up, definitions, endpoints, and research trial design[J]. Europace, 2012, 14(4): 528-606. 10.1093/europace/eus027. [DOI] [PubMed] [Google Scholar]
  • 12. Winkle RA, Mead RH, Engel G, et al. Long-term results of atrial fibrillation ablation: the importance of all initial ablation failures undergoing a repeat ablation[J]. Am Heart J, 2011, 162(1): 193-200. 10.1016/j.ahj.2011.04.013. [DOI] [PubMed] [Google Scholar]
  • 13. Kim JS, Lee SA, Park JB, et al. Preoperative risk factor analysis of postoperative stroke after Cox-maze procedure with mitral valve repair[J]. BMC Cardiovasc Disord, 2014, 14: 116. 10.1186/1471-2261-14-116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Takagaki M, Yamaguchi H, Ikeda N, et al. Risk factors for atrial fibrillation recurrence after cox maze IV performed without pre-exclusion[J]. Ann Thorac Surg, 2020, 109(3): 771-779. 10.1016/j.athoracsur.2019.07.016. [DOI] [PubMed] [Google Scholar]
  • 15. Damiano RJ, Schwartz FH, Bailey MS, et al. The Cox maze IV procedure: predictors of late recurrence[J]. J Thorac Cardiovasc Surg, 2011, 141(1): 113-121. 10.1016/j.jtcvs.2010.08.067. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Bzdok D, Altman N, Krzywinski M. Statistics versus machine learning[J]. Nat Methods, 2018, 15(4): 233-234. 10.1038/nmeth.4642. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Rajula HSR, Verlato G, Manchia M, et al. Comparison of conventional statistical methods with machine learning in medicine: diagnosis, drug development, and treatment[J]. Medicina, 2020, 56(9): 455. 10.3390/medicina56090455. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Verma AA, Murray J, Greiner R, et al. Implementing machine learning in medicine[J/OL]. CMAJ, 2021, 193(34): E1351-E1357[2023-01-02]. 10.1503/cmaj.202434. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Peiffer-Smadja N, Rawson TM, Ahmad R, et al. Machine learning for clinical decision support in infectious diseases: a narrative review of current applications[J]. Clin Microbiol Infect, 2020, 26(5): 584-595. 10.1016/j.cmi.2019.09.009. [DOI] [PubMed] [Google Scholar]
  • 20. van Kooten RT, Bahadoer RR, Ter Buurkes de Vries B, et al. Conventional regression analysis and machine learning in prediction of anastomotic leakage and pulmonary complications after esophagogastric cancer surgery[J]. J Surg Oncol, 2022, 126(3): 490-501. 10.1002/jso.26910. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Wagner M, Brandenburg JM, Bodenstedt S, et al. Surgomics: personalized prediction of morbidity, mortality and long-term outcome in surgery using machine learning on multimodal data[J]. Surg Endosc, 2022, 36(11): 8568-8591. 10.1007/s00464-022-09611-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Cervantes J, Garcia-Lamont F, Rodríguez-Mazahua L, et al. A comprehensive survey on support vector machine classification: applications, challenges and trends[J]. Neurocomputing, 2020, 408: 189-215. 10.1016/j.neucom.2019.10.118. [DOI] [Google Scholar]
  • 23. Bisong E. Logistic regression[M]//Building machine learning and deep learning models on Google cloud platform. Berkeley, CA: Apress, 2019: 243-250. 10.1007/978-1-4842-4470-8_20. [DOI] [Google Scholar]
  • 24. Sagi O, Rokach L. Approximating XGBoost with an interpretable decision tree[J]. Inf Sci, 2021, 572: 522-542. 10.1016/j.ins.2021.05.055. [DOI] [Google Scholar]
  • 25. Denisko D, Hoffman MM. Classification and interaction in random forests[J]. Proc Natl Acad Sci USA, 2018, 115(8): 1690-1692. 10.1073/pnas.1800256115. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Prokhorenkova L, Gusev G, Vorobev A, et al. CatBoost: unbiased boosting with categorical features[C]//Proceedings of the 32nd International Conference on Neural Information Processing Systems. December 3-8, 2018, Montréal, Canada. New York: ACM, 2018: 6639-6649. 10.5555/3327757.3327770. [DOI] [Google Scholar]
  • 27. Wang WY, Sun DC. The improved AdaBoost algorithms for imbalanced data classification[J]. Inf Sci, 2021, 563: 358-374. 10.1016/j.ins.2021.03.042. [DOI] [Google Scholar]
  • 28. Lee TH, Ullah A, Wang R. Bootstrap aggregating and random forest[M]//Macroeconomic forecasting in the Era of big data. Cham: Springer, 2020: 389-429. 10.1007/978-3-030-31150-6_13. [DOI] [Google Scholar]
  • 29. Zhang CS, Zhang Y, Shi XJ, et al. On incremental learning for gradient boosting decision trees[J]. Neural Process Lett, 2019, 50(1): 957-987. 10.1007/s11063-019-09999-3. [DOI] [Google Scholar]
  • 30. Tang JX, Deng CW, Huang GB. Extreme learning machine for multilayer perceptron[J]. IEEE Trans Neural Netw Learn Syst, 2016, 27(4): 809-821. 10.1109/TNNLS.2015.2424995. [DOI] [PubMed] [Google Scholar]
  • 31. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach[J]. Biometrics, 1988, 44(3): 837-845. [PubMed] [Google Scholar]
  • 32. Khiabani AJ, MacGregor RM, Bakir NH, et al. The long-term outcomes and durability of the Cox-Maze IV procedure for atrial fibrillation[J/OL]. J Thorac Cardiovasc Surg, 2022, 163(2): 629-641. e7[2023-01-02]. 10.1016/j.jtcvs.2020.04.100. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Philpott JM, Zemlin CW, Cox JL, et al. The ABLATE trial: safety and efficacy of cox maze-IV using a bipolar radiofrequency ablation system[J]. Ann Thorac Surg, 2015, 100(5): 1541-1546. 10.1016/j.athoracsur.2015.07.006. [DOI] [PubMed] [Google Scholar]
  • 34. Chew DS, Black-Maier E, Loring Z, et al. Diagnosis-to-ablation time and recurrence of atrial fibrillation following catheter ablation: a systematic review and meta-analysis of observational studies[J/OL]. Circ Arrhythm Electrophysiol, 2020, 13(4): e008128[2023-01-02]. 10.1161/CIRCEP.119.008128. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Andrade JG, Deyell MW, Verma A, et al. Association of atrial fibrillation episode duration with arrhythmia recurrence following ablation: a secondary analysis of a randomized clinical trial[J/OL]. JAMA Netw Open, 2020, 3(7): e208748[2023-01-02]. 10.1001/jamanetworkopen.2020.8748. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36. Lerman BJ, Popat RA, Assimes TL, et al. Association of left ventricular ejection fraction and symptoms with mortality after elective noncardiac surgery among patients with heart failure[J]. JAMA, 2019, 321(6): 572-579. 10.1001/jama.2019.0156. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Minami K, Kabata D, Kakuta T, et al. U-shaped association between intraoperative net fluid balance and risk of postoperative recurrent atrial tachyarrhythmia among patients undergoing the cryo-maze procedure: an observational study[J]. J Cardiothorac Vasc Anesth, 2021, 35(8): 2392-2396. 10.1053/j.jvca.2020.10.023. [DOI] [PubMed] [Google Scholar]
  • 38. Hwang SK, Yoo JS, Kim JB, et al. Long-term outcomes of the maze procedure combined with mitral valve repair: risk of thromboembolism without anticoagulation therapy[J]. Ann Thorac Surg, 2015, 100(3): 840-844. 10.1016/j.athoracsur.2015.02.073. [DOI] [PubMed] [Google Scholar]
  • 39. Jin X, Pan JK, Wu HL, et al. Are left ventricular ejection fraction and left atrial diameter related to atrial fibrillation recurrence after catheter ablation?: a meta-analysis[J/OL]. Medicine, 2018, 97(20): e10822[2023-01-02]. 10.1097/MD.0000000000010822. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Muraru D, Caravita S, Guta AC, et al. Functional tricuspid regurgitation and atrial fibrillation: which comes first, the chicken or the egg?[J]. CASE, 2020, 4(5): 458-463. 10.1016/j.case.2020.04.011. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41. Bak M, Jeong DS, Park SJ, et al. Predictor of atrial fibrillation recurrence in patients who underwent a tricuspid valve operation with modified Cox maze procedure[J]. Echocardiography, 2022, 39(3): 447-456. 10.1111/echo.15315. [DOI] [PubMed] [Google Scholar]
  • 42. Horne BD, Anderson JL, John JM, et al. Which white blood cell subtypes predict increased cardiovascular risk?[J]. J Am Coll Cardiol, 2005, 45(10): 1638-1643. 10.1016/j.jacc.2005.02.054. [DOI] [PubMed] [Google Scholar]
  • 43. Khiabani AJ, Schuessler RB, Damiano RJ. Surgical ablation of atrial fibrillation in patients with heart failure[J]. J Thorac Cardiovasc Surg, 2021, 162(4): 1100-1105. 10.1016/j.jtcvs.2020.05.125. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. Yayla Ç, Akboğa MK, Gayretli Yayla K, et al. A novel marker of inflammation in patients with slow coronary flow: lymphocyte-to-monocyte ratio[J]. Biomark Med, 2016, 10(5): 485-493. 10.2217/bmm-2016-0022. [DOI] [PubMed] [Google Scholar]
  • 45. Fici F, Celik T, Balta S, et al. Comparative effects of nebivolol and metoprolol on red cell distribution width and neutrophil/lymphocyte ratio in patients with newly diagnosed essential hypertension[J]. J Cardiovasc Pharmacol, 2013, 62(4): 388-393. 10.1097/FJC.0b013e31829f716a. [DOI] [PubMed] [Google Scholar]
  • 46. Uthamalingam S, Patvardhan EA, Subramanian S, et al. Utility of the neutrophil to lymphocyte ratio in predicting long-term outcomes in acute decompensated heart failure[J]. Am J Cardiol, 2011, 107(3): 433-438. 10.1016/j.amjcard.2010.09.039. [DOI] [PubMed] [Google Scholar]
  • 47. Silberman S, Fink D. Neutrophil to lymphocyte ratio and risk of atrial fibrillation[J]. Am J Cardiol, 2019, 123(11): 1889. 10.1016/j.amjcard.2019.03.033. [DOI] [PubMed] [Google Scholar]
  • 48. Min K, Kwon S, Cho SY, et al. Atrial fibrillation is strongly associated with the neutrophil to lymphocyte ratio in acute ischemic stroke patients: a retrospective study[J/OL]. J Clin Lab Anal, 2017, 31(2): e22041[2023-01-02]. 10.1002/jcla.22041. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49. Wei Y, Bao YY, Lin CJ, et al. Early recurrence after cryoballoon versus radiofrequency ablation for paroxysmal atrial fibrillation: mechanism and implication in long-term outcome[J]. BMC Cardiovasc Disord, 2022, 22(1): 400. 10.1186/s12872-022-02816-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50. McCabe JM, Smith LM, Tseng ZH, et al. Protracted CRP elevation after atrial fibrillation ablation[J]. Pacing Clin Electrophysiol, 2008, 31(9): 1146-1151. 10.1111/j.1540-8159.2008.01155.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51. Kim YG, Boo KY, Choi JI, et al. Early recurrence is reliable predictor of late recurrence after radiofrequency catheter ablation of atrial fibrillation[J]. JACC Clin Electrophysiol, 2021, 7(3): 343-351. 10.1016/j.jacep.2020.09.029. [DOI] [PubMed] [Google Scholar]
  • 52. Pisters R, Lane DA, Nieuwlaat R, et al. A novel user-friendly score (HAS-BLED) to assess 1-year risk of major bleeding in patients with atrial fibrillation: the Euro Heart Survey[J]. Chest, 2010, 138(5): 1093-1100. 10.1378/chest.10-0134. [DOI] [PubMed] [Google Scholar]
  • 53. Nashef SAM, Roques F, Sharples LD, et al. EuroSCORE II[J]. Eur J Cardiothorac Surg, 2012, 41(4): 734-745. 10.1093/ejcts/ezs043. [DOI] [PubMed] [Google Scholar]
  • 54. European Heart Rhythm Association, European Association for Cardio-Thoracic Surgery, Camm AJ, et al. Guidelines for the management of atrial fibrillation: the task force for the management of atrial fibrillation of the European society of cardiology (ESC)[J]. Eur Heart J, 2010, 31(19): 2369-2429. 10.1093/eurheartj/ehq278. [DOI] [PubMed] [Google Scholar]
  • 55. Chen TQ, Guestrin C. XGBoost: a scalable tree boosting system[C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. August 13 - 17, 2016, San Francisco, California, USA. New York: ACM, 2016: 785-794. 10.1145/2939672.2939785. [DOI] [Google Scholar]
  • 56. Lundberg SM, Lee SI. A unified approach to interpreting model predictions[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. December 4-9, 2017, Long Beach, California, USA. New York: ACM, 2017: 4768-4777. 10.5555/3295222.3295230. [DOI] [Google Scholar]
  • 57. Kakuta T, Fukushima S, Minami K, et al. Novel risk score for predicting recurrence of atrial fibrillation after the Cryo-Maze procedure[J]. Eur J Cardiothorac Surg, 2021, 59(6): 1218-1225. 10.1093/ejcts/ezaa468. [DOI] [PubMed] [Google Scholar]

Articles from Journal of Central South University Medical Sciences are provided here courtesy of Central South University

RESOURCES