【作者】 孙颖；

【导师】 张志国；

【作者基本信息】 吉林大学 ， 临床医学硕士（专业学位）， 2025， 硕士

【摘要】 目的:基于机器学习算法构建多参数融合的预测体系,评估不同算法对阵发性房颤(Paroxysmal atrial fibrillation,PAF)患者导管消融术(Catheter Ablation,CA)术后复发的预测效能,筛选最优预测模型,通过可解释性分析量化关键生物标志物的预测贡献度,最终为临床决策提供依据,优化个体化风险评估和术后随访策略。方法:回顾性纳入2019年6月1日至2023年6月1日在吉林大学第一医院首次接受CA的PAF患者。排除标准为:既往导管消融治疗史、严重结构性心脏病(左室射血分数<40%或重度瓣膜病变)以及恶性肿瘤病史。随访截至2024年12月31日,根据术后复发情况分为复发组和未复发组,最终纳入501例患者。收集并预处理患者的多维度临床特征,采用多阶段特征筛选策略(包括方差分析、递归特征消除和临床验证)确定预测变量,并基于这些变量分别构建逻辑回归(Logistic Regression,LR)、支持向量机(Support Vector Machine,SVM)、K邻近算法(K-Nearest Neighbor,KNN)、随机森林(Random Forest,RF)、极端梯度提升机(Extreme Gradient Boosting,XGBoost)和轻量梯度提升机(Light Gradient Boosting Machine,Light GBM)六种机器学习模型。使用分层随机抽样法按7:3比例将数据划分为训练集(模型构建与超参数优化)和测试集(模型验证与评估)。训练过程中通过网格搜索、随机搜索和5折交叉验证联合优化模型超参数。模型性能从区分度(Area Under the Receiver Operating Characteristic Curve,AUROC)、分类效能(准确率、精确率、召回率、特异度和F1分数)、校准度(校准曲线)及临床实用性(决策曲线分析)三方面进行全面评估,最终筛选出最优预测模型,并应用SHAP(SHapley Additive ex Planations)框架量化分析特征对模型预测的贡献程度。研究结果:1.本研究共纳入501例首次行CA的PAF患者,其中69例术后3月空白期后出现≥30s的房性心律失常,复发率为13.8%(95%CI:10.7-16.9%)。2.经特多阶段征筛选,房颤负荷(AF Burden)、二尖瓣反流程度、身体质量指数(Body Mass Index,BMI)、B型钠尿肽(B-type Natriuretic Peptide,BNP)、左房内径(Left Atrial Diameter,LAD)、左室舒张末期内径(Left Ventricular End-Diastolic Dimension,LVEDd)、P波时限(P wave Duration)、校正QT间期(QTc)、病程时长、房间传导阻滞(Intra-atrial Block,IAB)11项变量纳入最终模型。3.Light GBM为最优模型,其在训练集中交叉验证AUC为0.97,准确率、精确率、召回率、特异度、F1分数都为1.00;测试集中,AUC为0.97,准确率0.93,特异度0.97,召回率(0.72)与F1分数(0.78)为所有模型中最高。校准曲线与理想校准重合度较高,决策曲线分析显示其在广泛阈值区间都有较高的净收益。4.SHAP分析贡献度,房颤负荷为最重要的预测因素(SHAP值=1.50),远高于其他变量。其次是P波时限(SHAP值=0.66)、二尖瓣反流程度(SHAP值=0.52)、病程时长(SHAP=0.37)也对预测复发贡献度较大。QTc(SHAP值=0.29)、BMI(SHAP值=0.25)、BNP(SHAP值=0.23)是相对重要的影响因素;LVEF(SHAP值=0.10)、LAD(SHAP值=0.08)、LVEDd(SHAP值=0.08)、IAB(SHAP值=0.03)对复发的贡献度相对较小。结论:1、阵发性房颤(PAF)患者导管消融(CA)术后有一定的复发率,术后3月空白期后复发率达13.8%(95%CI:10.7-16.9%),提示临床需重视术后复发风险的早期识别与管理,以改善患者长期预后并降低医疗负担。2、通过多阶段特征筛选与SHAP可解释性分析,研究证实房颤负荷、P波时限、二尖瓣反流程度、LAD及QTc与复发相关,其中房颤负荷与P波时限的贡献度最为显著,揭示了电生理重构与结构重塑的协同作用机制。3.本研究构建的基于Light GBM算法的多维临床参数机器学习预测模型,在测试集中表现出优异的区分度(AUC=0.97)与临床适用性(决策曲线分析显示广泛阈值范围内具有净获益),其图形化、可解释的风险分层工具(如SHAP特征贡献图)为临床医生提供了直观的决策支持,有助于优化术后随访策略及个体化干预方案的制定。更多还原

【Abstract】 Objective:This study aimed to develop a multi-parameter predictive system based on machine learning algorithms to assess the predictive performance of different models for recurrence after catheter ablation(CA)in patients with paroxysmal atrial fibrillation(PAF).The optimal predictive model was selected,and key biomarkers contributing to recurrence were quantified using explainability analysis to provide evidence for clinical decision-making and optimize individualized risk assessment and postoperative follow-up strategies.Methods:A retrospective study was conducted to collect data from patients with paroxysmal atrial fibrillation(PAF)who underwent their first catheter ablation(CA)at the First Hospital of Jilin University between June 1,2019,and June 1,2023.Patients with a history of catheter ablation,severe structural heart disease(left ventricular ejection fraction<40%or severe valvular disease),or a history of malignant tumors were excluded.Follow-up continued until December 31,2024,and the patients were categorized into recurrence and non-recurrence groups,resulting in a final cohort of 501 patients.A comprehensive set of clinical features was collected,and data preprocessing was performed.Predictive variables were selected through a multi-stage feature selection process,including variance analysis,recursive feature elimination,and clinical validation.A predictive modeling framework was constructed using six machine learning algorithms:logistic regression(LR),support vector machine(SVM),k-nearest neighbor(KNN),random forest(RF),extreme gradient boosting(XGBoost),and light gradient boosting machine(Light GBM).The dataset was randomly split into a training set and a test set in a 7:3 ratio.The training set was used for model construction and hyperparameter tuning,while the test set was used for validation and performance evaluation.Hyperparameters were optimized through a combination of grid search,random search,and five-fold cross-validation.Model performance was assessed based on three key aspects:discrimination ability(AUROC),classification performance(accuracy,precision,recall,specificity,and F1-score),and calibration(calibration curve analysis).Clinical utility was further evaluated through decision curve analysis(DCA)to identify the optimal predictive model.Additionally,feature contributions were analyzed using the SHAP(SHapley Additive ex Planations)framework.Results:1.A total of 501 PAF patients undergoing CA were included,with 69 experiencing atrial arrhythmia recurrence(≥30s)beyond the three-month blanking period,resulting in a recurrence rate of 13.8%(95%CI:10.7–16.9%).2.After multi-stage feature selection,11 key variables were included in the final model:AF burden,mitral regurgitation severity,body mass index(BMI),B-type natriuretic peptide(BNP),left atrial diameter(LAD),left ventricular end-diastolic diameter(LVEDd),P wave duration,corrected QT interval(QTc),AF duration,and intra-atrial block(IAB).3.Light GBM was identified as the optimal model,with an AUROC of 0.97 in cross-validation,achieving perfect classification metrics(accuracy,precision,recall,specificity,and F1-score=1)in the training set.In the test set,Light GBM achieved an AUROC of 0.97,accuracy of 0.93,specificity of 0.97,recall of 0.72,and an F1-score of 0.78—the highest among all models.Calibration analysis demonstrated close alignment with the ideal calibration curve,while decision curve analysis indicated high clinical utility across a broad threshold range.4.SHAP analysis identified AF burden as the most significant predictor(SHAP value=1.50),followed by P wave duration(SHAP value=0.66),mitral regurgitation severity(SHAP value=0.52),and AF duration(SHAP value=0.37).QTc(SHAP value=0.29),BMI(SHAP value=0.25),and BNP(SHAP value=0.23)also played notable roles,whereas LVEF(SHAP value=0.10),LAD(SHAP value=0.08),LVEDd(SHAP value=0.08),and IAB(SHAP value=0.03)had relatively lower contributions.Conclusion:1.Catheter ablation(CA)for patients with paroxysmal atrial fibrillation(PAF)is associated with a certain rate of recurrence.The recurrence rate after the 3-month post-procedural blanking period reaches 13.8%(95%CI:10.7–16.9%),highlighting the importance of early identification and management of recurrence risk to improve long-term outcomes and reduce healthcare burden.2.Using multi-stage feature selection and SHAP-based interpretability analysis,the study identified atrial fibrillation burden,P-wave duration,degree of mitral regurgitation,left atrial diameter(LAD),and QTc interval as recurrence-related factors.Among these,atrial fibrillation burden and P-wave duration demonstrated the highest contributions,suggesting a synergistic mechanism involving both electrical and structural remodeling.3.The machine learning prediction model developed in this study,based on the Light GBM algorithm and multidimensional clinical parameters,demonstrated excellent discriminative performance in the test set(AUC=0.97)and strong clinical utility(as shown by decision curve analysis across a wide range of threshold probabilities).The graphical and interpretable risk stratification tools(e.g.,SHAP feature contribution plots)offer intuitive decision support for clinicians,facilitating optimization of post-ablation follow-up strategies and individualized intervention planning.更多还原