摘要

第一部分 NT-proBNP在植入型心律转复除颤器患者预后危险分层中的价值

研究背景：N末端B型利钠肽前体（NT-proBNP）在心力衰竭（心衰）预后评价中的价值已得到证实。然而，NT-proBNP在植入型心律转复除颤器（ICD）患者预后危险分层中的价值尚不明确。

研究方法：纳入2013年1月至2020年9月所有在阜外医院行ICD植入术的缺血性心肌病或扩张型心肌病患者，收集基线信息、程控资料以及生存状态。根据基线NT-proBNP水平从低到高将患者等分为4组，以水平最低组作为参照，分别采用Cox回归以及竞争分析回归分析其余三组与ICD植入者全因死亡、首次恰当放电治疗的关系。另外，采用限制性立方样条函数分析NT-proBNP作为连续变量处理时与结局的非线性关联，并使用Wald检验验证假设。

研究结果：最终纳入500例患者，平均年龄60.2±12.0岁，其中男性415例（83.0%），一级预防136例（27.2%），非缺血性心肌病患者231例（46.2%）。中位随访时长4.1（2.8-5.7）年后，106例（21.2%）患者死亡。在校正混杂因素后，相比NT-proBNP最低组，另外三组的相对风险比（HR）分别为1.77（95% 置信区间[Confidence Interval，CI] 0.71-4.43），3.98（95% CI 1.71-9.25），5.90（95% CI 2.43-14.30），且风险呈线性上升趋势（线性趋势检验P < 0.001）。限制性立方样条显示，随着NT-proBNP水平上升，全因死亡风险随之上升。但是当NT-proBNP水平到达3231.4 pg/mL时，全因死亡风险不再继续上升，展现出天花板效应（非线性检验P < 0.001）。中位程控随访时长1.7（0.8-3.5）年后，共有89例（17.8%）患者因持续性室性心动过速/心室颤动接受了首次ICD恰当电击治疗。无论将NT-proBNP作为分类变量或连续变量，Fine-Gray竞争风险模型均显示NT-proBNP水平与放电治疗无关联（所有检验P > 0.05），且不存在非线性关联（P = 0.666）。

研究结论：随着NT-proBNP水平的升高，心衰ICD植入者的全因死亡风险随之增高，但是达到一定水平后风险就不再相应增加，显示出天花板效应。NT-proBNP与ICD植入者后续接受放电治疗无关。据此推断，NT-proBNP水平越高，ICD植入者获益的可能性越低。

第二部分 竞争风险列线图在植入型心律转复除颤器的扩张型心肌病患者预后危险分层中的价值

研究背景：扩张型心肌病（DCM）患者安装植入型心律转复除颤器（ICD）能否获益存在争议。建立预测安装ICD的DCM患者在接受恰当放电治疗前就发生死亡或接受心脏移植的风险评分模型有助于筛选合适的ICD植入者。

研究方法：以接受ICD恰当放电治疗前发生死亡或接受心脏移植作为研究终点，回顾纳入2010年1月至2019年12月在阜外医院行ICD植入术的218例DCM患者。首先，以Cox比例风险回归寻找死亡及心脏移植的独立预测因子；然后，利用上述变量和ICD植入指征建立预测终点事件的Fine-Gray竞争风险模型；最后，在此基础上创建可视化的列线图。模型评价采用内部验证的方法，评价指标包括受试者特征曲线（ROC）的曲线下面积（AUC）、Harrell’s C统计量、校准曲线以及决策曲线（DCA）。

研究结果：终点事件1、3、5年的发生率分别为5.3%（95% 置信区间[Confidence Interval，CI] 2.9-9.9%）、16.6%（95% CI 11-25.0%）、25.3%（95% CI 17.2-37.1%）。模型最终纳入了以下5个预测指标：植入指征、左室舒张末期内径、N末端B型利钠肽前体、血管紧张素转换酶抑制剂和血管紧张素受体阻滞剂、胺碘酮（P均<0.05）。在此基础上建立的列线图区分度高，预测1、3、5年终点事件的AUC分别为0.83（95% CI 0.73-0.94，P < 0.001）、0.84（95% CI 0.75-0.93，P < 0.001）、0.85（95% CI 0.77-0.94，P < 0.001），C指数达到0.788（95% CI，0.697-0.877，P <0.001），校正后的C指数为0.762。模型的校准度佳，校准曲线的斜率达到0.896。另外，DCA分析显示模型的临床实用性好。根据列线图计算得到评分，可将患者发生终点事件的风险分为高中低三组（累计发生率函数差异P < 0.001），其中高风险组占比17.9%，该组几乎不能从ICD植入获益。

研究结论：基于竞争风险回归建立的列线图是一个简单、实用的可用于筛选潜在合适ICD植入者的危险分层工具，有望用于指导医患共同决策。

第三部分 无监督机器学习在植入型心律转复除颤器心脏性猝死二级预防心衰患者预后危险分层中的价值

研究背景：此前研究未能成功实现心脏性猝死二级预防安装植入型心律转复除颤器（ICD）心衰患者的危险分层。该研究旨在评估基于常规临床资料的无监督聚类能否将上述患者成功分为具有不同特征以及预后的亚组。

研究方法：纳入389例植入ICD的心脏性猝死二级预防心衰患者，共收集44个基线特征以及患者发生死亡和接受ICD恰当放电治疗的情况。选择层次K均值聚类作为聚类算法，并采用混合数据的因子分析（FAMD）进行数据预处理。聚类分组的有效性通过比较组间基线和结局（包括全因死亡和首次接受ICD恰当放电治疗）的差异验证。

研究结果：程控中位随访时间为2.7年，生存中位随访时间为5.1年，共有142（36.5%）例患者接受了恰当放电治疗，113（29.0%）例患者死亡。FAMD提取出的前12个主成分特征根大于1，共包含原始变量60.5%的总变异，被保留用于后续分析。基于主成分的聚类分析最终识别出了3个亚组。组1的患者几乎均为缺血性心肌病，且年龄最大，患糖尿病、高血压、高脂血症的比例也最高，但是心脏的结构和功能在三组中最佳。组2的患者年龄最小，多数为非缺血性心肌病，具有最小比例的合并症，心脏状况介于另外两组之间。组3的患者心衰的进展最严重。Kaplan-Meier生存曲线显示三组在接受放电治疗（P = 0.002）以及全因死亡（P < 0.001）方面均具有显著差异。经过校正药物治疗后，相比于组1，组2和组3患者接受放电治疗的风险比（HR）依次增加，分别为1.54（95% 置信区间[Confidence Interval，CI] 1.03–2.28，P = 0.033）、2.21（95% CI 1.42-3.43，P < 0.001），线性趋势检验P < 0.001。对于死亡风险而言，组3相比于组1患者校正后的HR为2.25（95% CI 1.45-3.49，P < 0.001），而组2相比组1不具有统计学差异（P = 0.124）。

研究结论：无监督聚类可以识别出具有不同临床特征及预后的二级预防ICD植入者，有望用于这些患者的精准危险分层。

第四部分 有监督机器学习在植入型心律转复除颤器患者预后危险分层中的价值

研究背景：当前预防心脏性猝死安装植入型心律转复除颤器（ICD）的指南已不能满足精准医学的需求。本研究旨在建立能处理生存资料的机器学习模型，比较其表现是否优于传统的Cox比例风险回归（CPH），评估模型的可解释性，并最终建立二维风险预测模型。

研究方法：共纳入887例植入ICD的患者，收集45个基线特征以及患者发生死亡和接受ICD恰当放电治疗的情况。将患者随机分为训练集（n=665）与测试集（n=222）。共采用4种机器学习算法：Cox弹性网络回归（EN-Cox）、随机生存森林（RSF）、生存支持向量机（SSVM）、基于极端梯度提升的生存分析（XGBoost），利用网格搜索确立最优超参数，模型的准确性以C指数评估。在测试集中比较机器学习算法与经典CPH算法的表现差异。模型的意义采用Shapley值（SHAP）解释。

研究结果：共有199例患者死亡（事件发生率5.0/100人年），265例患者接受了首次恰当放电治疗（事件发生率12.4/100人年）。在预测死亡的机器学习算法中，XGBoost表现最佳，且优于CPH模型（C指数：0.794 vs. 0.760，P < 0.001）。其余算法表现也均不低于CPH模型。在预测首次恰当放电治疗机器学习算法中，SSVM模型的表现在数值上优于CPH模型，但未到达显著性差异（C指数：0.621 vs. 0.611，P = 0.243）。其余算法表现与之相当或稍低。SHAP算法显示CPH与机器学习模型中变量的预测价值符合既往发现。利用筛选出的最优机器学习算法，在全集中建立可以同时评估死亡与放电治疗风险的二维预测模型。该模型可有效将患者分为9个不同风险组别，最终结合实际可给与3类不同的ICD植入推荐（推荐植入、医患共同决策、不推荐植入）。

研究结论：生存分析机器学习算法可用于ICD植入者的精准危险分层，并且解释性算法可以有效阐述模型中各变量的意义。

ABSTRACT (KEYWORDS)

Part I N-terminal pro-B-type natriuretic peptide in risk stratification of heart failure patients with implantable cardioverter-defibrillator

Background: Prognostic value of N-terminal pro-B-type natriuretic peptide (NT-proBNP) in heart failure is well-established. However, whether it could facilitate the risk stratification of heart failure patients with implantable cardioverter-defibrillator (ICD) is still unclear.

Methods: All patients with ischemic or non-ischemic dilated cardiomyopathy disease implanted with ICD between January 1, 2013 and September 1, 2020 at Fuwai hospital were enrolled. Patients’ baseline characteristics, ICD interrogation information, and survival status were collected. NT-proBNP levels were categorized into quartiles and the first quartile was set as the reference group to evaluate its association with the outcomes of all-cause mortality and first appropriate ICD shock due to sustained ventricular tachycardia/ventricular fibrillation in ICD recipients. Restricted cubic splines and Wald tests were used to find the nonlinear relationships.

Results: NT-proBNP was measured before ICD implant in 500 patients (mean age 60.2±12.0 years; 415 (83.0%) male; 231 (46.2%) non-ischemic dilated cardiomyopathy; 136 (27.2%) primary prevention). Over a median survival follow-up of 4.1 (interquartile range [IQR]: 2.8-5.7) years, 106 (21.2%) patients died. After adjusting for confounding factors, multivariable Cox regression showed a rise in NT-proBNP was associated with an increased risk of all-cause mortality. Compared with the lowest quartile, the hazard ratios with 95% confidence interval across increasing quartiles were 1.77 (0.71, 4.43), 3.98 (1.71, 9.25), and 5.90 (2.43, 14.30) for NT-proBNP (P for trend < 0.001). Restricted cubic spline demonstrated a similar pattern with an inflection point found at 3231.4 pg/mL, beyond which the increase in NT-proBNP was not associated with increased mortality (P for nonlinearity < 0.001). The median interrogation follow-up was 1.7 (IQR 0.8-3.5) years, and 89 (17.8%) patients had their first appropriate shock. Fine-Gray regression was used to evaluate the association between NT-proBNP and first appropriate shock accounting for the competing risk of death. In the unadjusted, partial, and fully adjusted analysis, however, no significant association could be found regardless of NT-proBNP as a categorical variable or log-transformed continuous variable (all P > 0.05). No nonlinearity was found, either (P = 0.666).

Conclusion: In heart failure patients with ICD, the rise in NT-proBNP is independently associated with increased mortality until it reaches the inflection point. However, its association with first appropriate shock was not found. Patients with higher NT-proBNP levels might derive less benefit from ICD implant.

Part II Competing risk nomogram predicting death and heart transplantation prior to appropriate ICD shock in dilated cardiomyopathy

Background: It’s still under debate whether non-ischemic dilated cardiomyopathy (DCM) patients would benefit from implantable cardioverter-defibrillators (ICD). Developing a simple risk score for predicting death and heart transplantation (HT) before receiving appropriate shock may help classifying potential ICD recipients.  

Methods: The primary endpoints included all-cause mortality and HT (whichever came first) without former appropriate shock. A total of 218 consecutive DCM patients implanted with ICD between 2010 and 2019 at Fuwai Hospital were retrospectively enrolled. Cox proportional-hazards model was primarily built to identify variables associated with death and HT. Then, a Fine-Gray model, accounting for the appropriate shock as a competing risk, was constructed using these selected variables along with implantation indication. Finally, a nomogram based on the Fine-Gray model was established to predict 1-, 3-, and 5-year probabilities of primary endpoints. The area under the receiver operating characteristic (ROC) curve (AUC), Harrell's C-index, and calibration curves were used to evaluate and internally validate the performance of this model. The decision curve analysis was applied to assess its clinical utility.

Results: The 1-, 3-, and 5-year cumulative incidence of all-cause mortality and HT without former appropriate shock were 5.3% (95% confidence interval [CI] 2.9-9.9%), 16.6% (95% CI 11-25.0%) and 25.3% (95% CI 17.2-37.1%), respectively. Five variables including implantation indication, left ventricular end-diastolic diameter, N-terminal pro-brain natriuretic peptide, angiotensin-converting enzyme inhibitor/angiotensin receptor blocker, and amiodarone treatment were independently associated with it (all P < 0.05) and were used for constructing the nomogram. The 1-, 3-, and 5-year AUC of the nomogram were 0.83 (95% CI 0.73-0.94, P < 0.001), 0.84 (95% CI 0.75-0.93, P < 0.001), and 0.85 (95% CI 0.77-0.94, P < 0.001), respectively. The Harrell's C-index was 0.788 (95% CI, 0.697-0.877, P < 0.001; 0.762 for the optimism-corrected C-index), showing the good discriminative ability of the model. And the calibration was acceptable (optimism-corrected slope 0.896). Decision curve analysis identified our model was clinically useful within the entire range of potential treatment thresholds for ICD implantation. Three risk groups stratified by scores were significantly different between cumulative incidence curves (P < 0.001). The identified high-risk group composed 17.9% of our population and did not derive long-term benefit from ICD.

Conclusion: The proposed nomogram is a simple, useful risk stratification tool for selecting potential ICD recipients in DCM patients. It might facilitate the shared decision-making between patients and clinicians.

Part III Unsupervised learning in the risk stratification of patients with heart failure and secondary prevention implantable cardioverter-defibrillator implantation

Background: Previous studies have failed to implement risk stratification in patients with heart failure (HF) who are eligible for secondary implantable cardioverter-defibrillator (ICD) implantation. We aimed to evaluate whether machine learning-based phenomapping using routinely available clinical data can identify subgroups that differ in characteristics and prognoses. 

Methods: A total of 389 patients with chronic HF implanted with an ICD were included, and their clinical outcomes and forty-four baseline variables were collected. Phenomapping was performed using hierarchical k-means clustering based on factor analysis of mixed data (FAMD). The utility of phenomapping was validated by comparing the baseline features and outcomes of the first appropriate shock and all-cause death among the phenogroups.

Results: During a median follow-up of 2.7 years for device interrogation and 5.1 years for survival status, 142 (36.5%) first appropriate shocks and 113 (29.0%) all-cause deaths occurred. The first 12 principal components extracted using the FAMD, explaining 60.5% of the total variability, were left for phenomapping. Three mutually exclusive phenogroups were identified. Phenogroup 1 comprised the oldest patients with ischemic cardiomyopathy; had the highest proportion of diabetes mellitus, hypertension, and hyperlipidemia; and had the most favorable cardiac structure and function among the phenogroups. Phenogroup 2 included the youngest patients, mostly those with non-ischemic cardiomyopathy, who had intermediate heart dimensions and function, and the fewest comorbidities. Phenogroup 3 had the worst HF progression. Kaplan–Meier curves revealed significant differences in the first appropriate shock (P = 0.002) and all-cause death (P < 0.001) across the phenogroups. After adjusting for medications in Cox regression, phenogroups 2 and 3 displayed a graded increase in appropriate shock risk (hazard ratio [HR] 1.54, 95% confidence interval [CI] 1.03–2.28, P = 0.033; HR 2.21, 95% CI 1.42-3.43, P < 0.001, respectively; P for trend <0.001) compared to phenogroup 1. Regarding mortality risk, phenogroup 3 was associated with an increased risk (HR 2.25, 95% CI 1.45-3.49, P < 0.001). In contrast, phenogroup 2 had a risk (P = 0.124) comparable with phenogroup 1.

Conclusion: Machine-learning-based phenomapping can identify distinct phenotype subgroups in patients with clinically heterogeneous HF with secondary prophylactic ICD therapy. This novel strategy may aid personalized medicine for these patients.

Part IV: Supervised learning in the risk stratification of patients with implantable cardioverter-defibrillator implantation

Background: Current guideline-based implantable cardioverter-defibrillator (ICD) implant fails to meet the demands for precision medicine. We aimed to develop explainable machine learning (ML) models predicting mortality and the first appropriate shock and compare these to standard Cox proportional hazards (CPH) regression in ICD recipients. We also aimed to bring up a new bi-dimensional model predicting both death and shock risk.

 Methods: A total of 887 adult patients were finally enrolled and randomly split into a training set (n=665, 75%) and a test set (n=222, 25%). Forty-five routine clinical variables were collected. Four fine-tuned ML approaches (elastic net Cox regression, random survival forests, survival support vector machine, and XGBoost) were applied and compared with the CPH model on the test set using Harrell’s C-index. Shapley Additive exPlanation (SHAP) values were used to explain each variable’s contribution to prediction.

Results: 199 patients died (5.0 per 100 person-years) and 265 first appropriate shocks occurred (12.4 per 100 person-years) during the follow-up. Among ML models predicting death, XGBoost achieved the highest accuracy and outperformed the CPH model (C-index: 0.794 vs. 0.760, P < 0.001). Other ML models did not perform worse than CPH model. For appropriate shock, survival support vector machine showed the highest accuracy, although not statistically different from the CPH model (0.621 vs. 0.611, P = 0.243). The performance of other ML models was either similar or slightly inferior to it. Feature contribution of ML models assessed by SHAP values at individual and overall levels was in accordance with established knowledge. Accordingly, a bi-dimensional risk matrix of nine scenarios integrating death and shock risk was built. This risk stratification framework ultimately classified patients into three recommendation situations (for/against/shared decision) due to different likelihood of benefiting from ICD implant.

Conclusion: Explainable survival ML models offer a promising tool for risk stratification and feature interpretation in ICD-eligible patients and may aid clinical decision making in the future.