第一部分

静态射血分数分型的心力衰竭：亚型内异质性表型的识别与表型特异性预后分析

 

第一节：超射血分数型心力衰竭的表型分型研究

 

背景和目的：超射血分数型心力衰竭（heart failure with supranormal ejection fraction, HFsnEF）作为心力衰竭（心衰）的特殊临床亚型，其治疗手段有限且预后不良。揭示该疾病的表型异质性对于理解疾病病理机制和优化患者管理至关重要。本研究旨在通过无监督聚类分析的方法识别HFsnEF中的潜在不同表型，描述HFsnEF的表型多样性，基于临床易获取变量构建实用分类器以实现表型聚类归属判定，并评估表型与预后的关系。

 

方法：本研究纳入住院HFsnEF患者（确诊心衰且基线超声心动图评估左心室射血分数 ≥ 65%的患者）进行分析。采用基于主成分的层次聚类（hierarchical clustering on principal components, HCPC）的方法对混合型数据变量（包括人口学特征、心衰病程、生命体征、人体测量学指标、吸烟/饮酒状况、心衰病因、临床合并疾病、实验室检查和超声心动图参数）进行无监督表型聚类；通过决策树模型识别能够区分不同表型的关键参数；并比较各表型间临床结局（全因死亡、心血管死亡及心血管再入院）的差异。

 

结果：研究纳入221名HFsnEF患者，识别出了三个独立表型。表型1（52.5%）的患者多以瓣膜性心脏病为主要病因，特征性表现为心腔扩大和心房颤动/扑动的高发生率；表型2（26.2%）以老年缺血性患者为主，该表型患者合并高血压、糖尿病、高脂血症及肾功能障碍的比例最高；表型3（21.3%）以肥厚型心肌病患者为主，40.4%的患者接受了室间隔切除术或酒精消融术。决策树模型证实，心衰病因与是否合并糖尿病是区分三组表型的核心指标。中位随访53.4个月期间，共记录到46例（20.8%）全因死亡（心血管死亡39例）和70例（31.7%）心血管再入院事件。生存分析显示，三个表型的死亡风险有显著差异，表型1的全因死亡风险显著高于其余表型（表型1 vs. 表型2：校正后风险比 [adjusted hazard ratio, aHR] = 3.32，95%置信区间 [confidence interval, CI] 1.19–9.25，P = 0.022；表型1 vs. 表型3：aHR = 3.81，95% CI 1.09–13.28，P = 0.036；表型2 vs. 表型3：aHR = 1.15，95% CI 0.23–5.70，P = 0.865）；表型1 的心血管死亡风险也显著高于其余表型（表型1 vs. 表型2：aHR = 3.73，95% CI 1.21–11.48，P = 0.022；表型1 vs. 表型3：aHR = 4.27，95% CI 1.26–14.50，P = 0.020；表型2 vs. 表型3：aHR = 1.15，95% CI 0.23–5.61，P = 0.870）；各表型间心血管再入院风险的差异无统计学显著性（表型1 vs. 表型2：aHR = 0.82，95% CI  0.39–1.70，P = 0.590；表型1 vs. 表型3：aHR = 1.04，95% CI  0.53–2.07，P = 0.900；表型2 vs. 表型3：aHR = 1.28，95% CI 0.54–3.02，P = 0.580）。

 

结论：本研究利用无监督机器学习方法（HCPC）在HFsnEF患者人群中识别出了三个具有显著临床特征差异和预后差异的表型，为理解HFsnEF的疾病异质性、改善HFsnEF患者的管理和指导未来研究提供了重要基础。

 

关键词：超射血分数；心力衰竭；表型聚类；临床结局

 

第二节：射血分数保留型心力衰竭的表型分型和LASSO-Cox回归指导的表型特异性预后预测

 

背景和目的：射血分数保留型心力衰竭（heart failure with preserved ejection fraction, HFpEF）具有显著的病理生理异质性，给预后评估与个体化治疗带来了严峻挑战。现有表型分型研究多基于欧美人群队列开展，缺乏跨人群队列的验证，且当前研究主要聚焦于病理生理学机制解析，表型分型成果向临床转化不足。本研究旨在用临床常规指标进行HFpEF表型分型，识别与既往研究报道表型特征高度一致的、稳健的表型；并通过表型特异性预后预测研究，揭示不同表型的风险驱动因素并构建预后预测模型，为个体化风险管理提供循证依据，增加表型的临床转化价值。

 

方法：本研究连续纳入左心室射血分数 ≥ 50%的住院心力衰竭（心衰）患者，采用基于主成分的层次聚类方法对患者进行无监督表型分型。按7:3比例将各表型人群分为训练集与测试集，在训练集内通过最小绝对收缩和选择算子（least absolute shrinkage and selection operator, LASSO）回归结合Cox回归筛选全因死亡风险预测变量并构建预后模型（包括全变量模型和逐步回归模型）。分别在每个表型的训练集和测试集中评价模型的1-5年全因死亡风险预测性能，评价指标包括区分度（Harrell’s C指数、Uno’s C指数、时间依赖的曲线下面积 [area under the curve, AUC]），校准度（校准曲线、观测风险与预测风险比值 [observed-to-expected ratio, OE比值]、校准斜率、校准误差指数），模型总体性能（Brier评分、预测准确性指数）和临床实用性（决策曲线分析 [decision curve analysis, DCA]）。

 

结果：本研究纳入2157名HFpEF患者（平均年龄60.9岁，男性59.9%），识别出两个独立表型。表型1患者以缺血性病因为主（52.7%），合并代谢异常比例较高，纽约心脏协会心功能分级多为I或II级（累计占比59.3%），N末端B型利钠肽原（N-terminal pro-brain natriuretic peptide, NT-proBNP）水平较低。表型2患者较为年长（平均年龄62.9岁），主要病因为瓣膜病，心功能和肝肾功能均较差，NT-proBNP及炎症标志物显著升高，常合并感染和心房颤动/扑动。通过比较所识别的表型和既往研究表型的特征，对表型的稳健性进行了验证。LASSO-Cox回归分析提示表型1全因死亡预测变量包括NT-proBNP、总三碘甲腺原氨酸、大内皮素-1等；根据以上变量构建的逐步回归模型区分度优异（训练集及测试集C指数、时间依赖的AUC均接近0.80），校准良好（OE比值接近1.0；校准斜率接近1.0），Brier评分显著低于基准模型。表型2全因死亡预测变量包括NT-proBNP、收缩压、血尿素氮、总胆红素和红细胞分布宽度标准差等；根据以上变量构建的全变量模型C指数接近0.70，时间依赖的AUC在多个时间点接近甚至超过0.80，OE比值接近1.0。DCA表明，针对两种表型构建的预后模型在广泛阈值概率范围内均展现出临床净获益。

 

结论：本研究基于国内HFpEF患者队列，通过无监督聚类方法识别出了两个具有显著临床特征差异的HFpEF表型，并通过分型对齐的方式验证了表型的稳健性。为增加表型分型的临床转化价值，本研究构建了表型特异性预后模型，为个体化风险管理提供了实用工具。

 

关键词：射血分数保留型心力衰竭；表型聚类；预后模型

 

第三节：射血分数降低型/轻度降低型心力衰竭的简易表型分型和

基于有监督机器学习的表型特异性预后预测

 

背景和目的：射血分数降低型/轻度降低型心力衰竭（heart failure with reduced or mildly reduced ejection fraction, HFrEF/HFmrEF）患者群体存在异质性，表现为不同的病理生理特征、不良事件风险及治疗反应差异。然而，相较于射血分数保留型心力衰竭（heart failure with preserved ejection fraction, HFpEF），HFrEF/HFmrEF的表型异质性研究相对匮乏，已有分型体系与HFpEF类似，分型复杂度高，临床推广受限；且缺乏针对特定表型的预后和干预研究，导致表型的临床转化价值有限。本研究旨在基于常规临床指标识别HFrEF/HFmrEF的异质性表型，并建立简易表型分型规则；此外，为了提高表型分型的临床转化价值，本研究拟探索不同表型的预后特征差异并进行表型特异性预后预测，为个体化风险管理提供依据。

 

方法：本研究连续纳入住院HFrEF/HFmrEF患者，采用基于主成分的层次聚类方法进行无监督表型分型，并通过随机森林算法筛选关键表型区分变量，建立简易分型规则，按照简易表型分型规则对患者重新进行分型。按7:3比例将各新表型人群分为训练集与测试集，在训练集中分别通过四种机器学习方法（随机生存森林 [random survival forests, RSF]、梯度提升机、生存支持向量机和监督主成分分析）构建生存预测模型，并在训练集和测试集中分别评估其在1-5年生存结局预测中的性能（通过动态C指数、时间依赖的曲线下面积、校准曲线及决策曲线分析等）。最后，筛选各表型最优机器学习模型的前20位重要性变量，构建表型特异性逐步回归Cox模型，并与ADHF/NT-proBNP风险评分进行性能对比。

 

结果：研究共纳入3711例住院HFrEF/HFmrEF患者（平均年龄54.6岁，男性76.1%），识别出三个独立表型。表型1患者较为年轻，主要病因为扩张型心肌病（dilated cardiomyopathy, DCM）；其N末端B型利钠肽原（N-terminal pro-brain natriuretic peptide, NT-proBNP）水平低，接受药物治疗比例高。表型2患者平均年龄大，以缺血性病因为主，合并代谢异常的比例高，左心重构程度低。表型3患者的主要病因为DCM，此表型患者心功能和肝肾功能较差，NT-proBNP水平高，常合并感染和心房颤动/扑动。根据随机森林的变量重要性排序，病因为区分各个表型的关键变量。因此，本研究通过病因分型（包括缺血性和DCM）的方式捕捉患者异质性。随访期间，在两个病因亚组中分别记录到477例（40.5%）和377例（28.8%）全因死亡事件。在缺血性表型中，RSF模型预测性能最优；NT-proBNP、血肌酐、年龄、血红蛋白等居RSF模型变量重要性前位。在DCM表型中，RSF同样表现卓越，在区分度、校准度、临床获益等方面优于其他模型或与其他模型相似；NT-proBNP、左心室舒张末期内径、基线是否使用血管紧张素转换酶抑制剂/血管紧张素II受体拮抗剂、红细胞分布宽度标准差等居变量重要性前位。在两个病因表型中，由RSF模型前20个重要性变量构建的逐步回归Cox模型在测试集中均表现出了较好的区分度、校准度、总体性能和临床实用性（优于ADHF/NT-proBNP风险评分）。

 

结论：HFrEF/HFmrEF内部存在异质性，病因是异质性的重要来源，可作为简易表型分型的依据。机器学习指导的预后预测和变量重要性排序揭示了不同病因表型预后相关预测因素的差异，其衍生的死亡风险预测模型为病因分型指导下的患者风险管理提供了循证支持。

 

关键词：射血分数降低型心力衰竭；射血分数轻度降低型心力衰竭；表型聚类；机器学习；预后模型

 

第二部分

动态射血分数分型的心力衰竭：射血分数改善型心力衰竭

 

射血分数改善型心力衰竭的发病率、疾病进展及预后——一项基于射血分数纵向评估的研究

 

背景和目的：修订版的《欧洲心脏病学会心力衰竭（心衰）通用定义和分型》确立了射血分数改善型心衰（heart failure with improved ejection fraction, HFimpEF）的独立亚型地位，对HFimpEF做出了标准化定义，并强调了纵向评估左心室射血分数（left ventricular ejection fraction, LVEF）的重要性。目前，基于这一最新定义，关于HFimpEF发病率、预测因素和预后的研究仍较少；基于多时间点LVEF评估的纵向研究也很有限，阻碍了对HFimpEF疾病进展的认识。本研究旨在探讨住院低LVEF（≤ 40%）心衰患者队列中HFimpEF的发病率和疾病进展，探索与HFimpEF及其进展有关的预测因素，评估HFimpEF及其进展与临床预后的关联。

 

方法：本研究回顾性纳入基线LVEF ≤ 40%且具备符合要求的超声随访记录的住院心衰患者。根据LVEF动态演变将人群分为：1）HFimpEF：定义为LVEF增加 ≥ 10个百分点且达到 > 40%；2）持续低射血分数型心衰：包括了随访期间始终未达到HFimpEF标准的患者；3）短暂型HFimpEF（LVEF恶化）：达到HFimpEF标准后，LVEF再次降至 ≤ 40%；4）持续型HFimpEF：定义为随访期间LVEF持续改善至 > 40%；5）LVEF再改善：LVEF恶化后再次达到HFimpEF标准。研究结局包括全因死亡、心血管死亡和心衰再住院。HFimpEF和LVEF动态变化的发生率通过累积发生率函数曲线展示。与HFimpEF和LVEF动态变化相关的变量通过决策树分析进行识别。HFimpEF、LVEF动态变化与结局之间的关系通过Cox比例风险模型和竞争风险模型进行评估。此外，本研究定量比较了不同生存结局患者的纵向LVEF轨迹，以识别可能预示不良预后的LVEF轨迹模式。按性别分组后重复以上分析，以探讨可能的性别差异。

 

结果：本研究共纳入923名患者。在中位随访47.9个月期间，517例达到HFimpEF标准，其中65.0%的LVEF改善发生于12个月内。与持续低射血分数型心衰患者相比，HFimpEF患者的全因死亡风险（校正后风险比 [adjusted hazard ratio, aHR] = 0.16，P < 0.001）、心血管死亡风险（aHR = 0.19，P < 0.001）和心衰再住院风险（aHR = 0.39，P < 0.001）均显著降低。然而，160名HFimpEF患者在随访期间经历了LVEF恶化，与持续型HFimpEF患者相比，经历LVEF恶化（短暂型HFimpEF）患者的不良事件风险更高（全因死亡aHR = 1.89，心血管死亡aHR = 2.13，心衰再住院aHR = 2.13，P均 < 0.05），且此类患者后续LVEF再改善可能性显著降低。关于上述所有分析，男性和女性之间的结果没有显著差异。对HFimpEF患者LVEF轨迹的纵向分析表明，倒“U”形的LVEF变化趋势——即缓慢轻度的升高后出现下降——预示着更高的死亡风险。

 

结论：在住院低LVEF（≤ 40%）心衰患者中，超过半数（56.0%）可在随访期间达到HFimpEF标准，但近1/3会出现LVEF恶化。LVEF恶化预示着更高的不良事件风险和更低的心脏逆重构可能。纵向评估LVEF演变轨迹具有重要临床价值，有助于识别HFimpEF患者、监测疾病进展、指导风险分层。

 

关键词：射血分数改善型心力衰竭；左心室射血分数；连续测量；预后评估

Part 1

Traditional Heart Failure Classification Based on Ejection Fraction: Identification of Phenotypic Heterogeneity within Subtypes and Phenotype-Specific Prognostic Analysis

 

Section 1:

Phenotypic Clustering in Heart Failure with Supranormal Ejection Fraction

 

Background and Aims: Heart failure with supranormal ejection fraction (HFsnEF) represents a distinct clinical entity characterized by limited treatment options and an unfavourable prognosis. Revealing its phenotypic diversity is crucial for understanding disease mechanism and optimizing patient management. We aim to identify phenotypic subgroups in HFsnEF using unsupervised clustering analysis, characterize the phenotypic diversity of HFsnEF, create classifiers using the most easily obtainable variables so that every individual can fall into a certain phenotype, and assess associations between phenotypes and outcomes.

 

Methods: Consecutive hospitalized patients with a diagnosis of heart failure (HF) and a left ventricular ejection fraction ≥ 65% at baseline echocardiographic evaluations were included for analysis. We conducted unsupervised hierarchical clustering on principal components (HCPC) analysis to identify HFsnEF phenotypes using mixed-data variables including demographics, HF duration, vital signs, anthropometrics, smoking/drinking status, HF etiology, comorbid diseases, laboratory tests, and echocardiographic parameters. We then employed decision tree modelling to identify parameters capable of distinguishing distinct phenotypes. Clinical outcomes, including all-cause death, cardiovascular (CV) death, and CV readmission for different phenotypes, were examined.

 

Results: Three mutually exclusive phenotypes were identified from the cohort of 221 HFsnEF patients. Phenotype 1 (52.5%) predominantly consisted of patients with valvular heart disease, who had larger cardiac chambers and a higher prevalence of atrial fibrillation/flutter. Phenotype 2 (26.2%) primarily comprised older ischemic patients with a higher prevalence of comorbidities such as hypertension, diabetes, hyperlipidemia, and renal dysfunction. Phenotype 3 (21.3%) were mainly hypertrophic cardiomyopathy patients, and 40.4% of these patients had received surgical septal myectomy or alcohol septal ablation. Two clinical variables were identified that could be used to group all HFsnEF patients into one of the phenotypes; they were HF etiology and comorbid diabetes. During the median follow-up of 53.4 months, 46 (20.8%) all-cause deaths occurred, among them 39 of CV causes. Seventy (31.7%) patients experienced CV readmissions. The three phenotypes showed distinct differences in mortality outcomes, with Phenotype 1 exhibiting the highest risk of all-cause mortality (Phenotype 1 vs. Phenotype 2: adjusted hazard ratio [aHR] = 3.32, 95% confidence interval [CI] 1.19–9.25, P = 0.022; Phenotype 1 vs. Phenotype 3: aHR = 3.81, 95% CI 1.09–13.28, P = 0.036; Phenotype 2 vs. Phenotype 3: aHR = 1.15, 95% CI 0.23–5.70, P = 0.865) and CV mortality (Phenotype 1 vs. Phenotype 2: aHR = 3.73, 95% CI 1.21–11.48, P = 0.022; Phenotype 1 vs. Phenotype 3: aHR = 4.27, 95% CI 1.26–14.50, P = 0.020; Phenotype 2 vs. Phenotype 3: aHR = 1.15, 95% CI 0.23–5.61, P = 0.870). CV readmission risk was comparable among the three phenotypes (Phenotype 1 vs. Phenotype 2: aHR = 0.82, 95% CI 0.39–1.70, P = 0.590; Phenotype 1 vs. Phenotype 3: aHR = 1.04, 95% CI 0.53–2.07, P = 0.900; Phenotype 2 vs. Phenotype 3: aHR = 1.28, 95% CI 0.54–3.02, P = 0.580).

 

Conclusion: In a heterogeneous HFsnEF cohort, three phenotypes were identified by unsupervised HCPC with distinct clinical characteristics and outcomes. This study provides a critical foundation for understanding the heterogeneity of HFsnEF, improving patient management, and guiding future research.

 

Keywords: Supranormal ejection fraction; Heart failure; Phenotypic clustering; Clinical outcomes

 

Section 2:

Phenotypic Clustering in Heart Failure with Preserved Ejection Fraction and LASSO-Cox-Based Phenotype-Specific Prognostic Prediction

 

Background and Aims: Heart failure with preserved ejection fraction (HFpEF) exhibits significant pathophysiological heterogeneity, posing challenges for prognosis assessment and individualized treatment. Existing phenotypic clustering studies were primarily conducted in European and American populations, with insufficient validation across diverse ethnic cohorts. Furthermore, current researches remain focused on pathophysiological mechanism elucidation and overlook clinical translational value of phenotypes. This study aims to identify robust phenotypes highly consistent with previous research findings. Additionally, by performing phenotype-specific prognostic analysis, this study aims to identify predictors for all-cause mortality and construct risk prediction models for each phenotype, providing evidence for individualized risk management, thereby enhancing the clinical utility of phenotyping.

 

Methods: This study included hospitalized HFpEF patients with left ventricular ejection fraction ≥ 50%. Unsupervised phenotyping was conducted using hierarchical clustering on principal components. Each phenotype was divided into a training set and a test set in a 7:3 ratio. In the training set, prognostic variables for all-cause mortality were selected using least absolute shrinkage and selection operator (LASSO)-Cox regression to construct models (including both full-variable and stepwise regression models). The model’s performance to predict all-cause mortality at 1-5 years was evaluated in both the training and test sets for each phenotype by metrics involving discrimination (Harrell’s C-index, Uno’s C-index, time-dependent area under the curve [AUC]), calibration (calibration curve, observed-to-expected ratio [OE ratio], calibration slope, calibration error indices), overall performance (Brier score, index of prediction accuracy), and clinical utility (decision curve analysis [DCA]).

 

Results: A total of 2157 HFpEF patients (mean age 60.9 years, 59.9% male) were included, two distinct phenotypes were identified. Phenotype 1 was characterized by ischemic etiology (52.7%) and a high prevalence of metabolic comorbidities. Patients were mostly classified as New York Heart Association functional class I or II (totaling 59.3%) and had low N-terminal pro-brain natriuretic peptide (NT-proBNP) levels at admission. Patients in Phenotype 2 were older (mean age 62.9 years), the primary underlying cause was valvular heart disease. They had worse cardiac, renal, and liver function, had higher levels of NT-proBNP and inflammatory biomarkers, and were more often comorbid with infections and atrial fibrillation/flutter. The robustness of the phenotypes was validated through characteristic comparison with existing phenotypes. LASSO-Cox regression identified predictors for all-cause mortality for Phenotype 1, such as NT-proBNP, triiodothyronine, and big endothelin-1. Stepwise regression model constructed with these variables showed excellent discrimination (C-indexes and time-dependent AUCs close to 0.80 in both training and test sets), good calibration (OE ratios close to 1.0, calibration slope close to 1.0), and good overall performance (significantly lower Brier scores compared to null model). Prognostic variables for all-cause mortality in Phenotype 2 included, but were not limited to, NT-proBNP, systolic blood pressure, blood urea nitrogen, bilirubin, and red blood cell distribution width-standard deviation. The derived full-variable model achieved C-indexes close to 0.70, with time-dependent AUCs at multiple time points approaching or exceeding 0.80 and OE ratios close to 1.0. DCA demonstrated that both prognostic models provided clinical net benefit across a wide range of threshold probabilities.

 

Conclusion: In a domestic HFpEF cohort, this study identified two distinct phenotypes using unsupervised clustering analysis. The robustness of the phenotypes was verified through phenotype alignment. To enhance the clinical utility of phenotype classification, phenotype-specific prognostic models were constructed, offering a practical tool for individualized risk management.

 

Keywords: Heart failure with preserved ejection fraction; Phenotypic clustering; Prognostic model

 

Section 3:

Simplified Phenotypic Clustering in Heart Failure with Reduced/Mildly Reduced Ejection Fraction and Machine Learning-Based Phenotype-Specific Prognostic Prediction

 

Background and Aims: Heart failure with reduced or mildly reduced ejection fraction (HFrEF/HFmrEF) exhibits clinical heterogeneity in pathophysiology, outcomes, and treatment responses. However, there is a lack of phenotypic clustering studies in HFrEF/HFmrEF. The current multidimensional phenotyping framework fails to reflect distinct pathophysiological characteristics of HFrEF/HFmrEF, resulting in overly complex and clinically impractical phenotypes. Moreover, the current research on prognostic prediction and intervention strategies across various phenotypes remains limited. This study aims to identify phenotypes in HFrEF/HFmrEF and establish a simple phenotypic classification rule. In addition, to enhance clinical translational value of phenotyping, this study aims to explore predictors for mortality across different phenotypes and perform phenotype-specific prognostic analysis.

 

Methods: Phenotypic clustering was performed by unsupervised hierarchical clustering on principal components in a hospitalized HFrEF/HFmrEF cohort. Key variables for differentiating phenotypes were identified by random forest analysis and were used to classify HFrEF/HFmrEF patients into new phenotypes. The new phenotypes were divided into training and testing sets in a 7:3 ratio. In training set, four machine learning methods (random survival forests [RSF], gradient boosting machine, survival support vector machine, and supervised principal components) were separately used to construct models for prediction of survival outcomes. The models’ performance in predicting all-cause mortality at 1-5 years was evaluated in training and test sets by dynamic C-index, time-dependent area under the curve, calibration curves, and decision curve analysis. For each phenotype, the top 20 important variables identified from the optimal model were used to build stepwise regression Cox models; the predictive performance of the models were compared with that of ADHF/NT-proBNP risk score.

 

Results: The study included 3711 HFrEF/HFmrEF patients (mean age 54.6 years, 76.1% male) and identified three phenotypes. Phenotype 1 comprised younger patients with a primary diagnosis of dilated cardiomyopathy (DCM), they had lower N-terminal pro-brain natriuretic peptide (NT-proBNP) levels and higher proportion receiving pharmacotherapy. Phenotype 2 primarily included older ischemic patients with metabolic comorbidities and the least degree of left heart remodeling. Phenotype 3 patients were mostly of DCM etiology, they had poorer cardiac, renal, and liver function, higher NT-proBNP levels, and higher prevalence of infections and atrial fibrillation/flutter. Variable importance ranking from random forest analysis confirmed etiology as the primary phenotypic discriminator. Therefore, this study captured patient heterogeneity through etiological (ischemic and DCM) classification. During follow-up, 477 (40.5%) all-cause mortality events occurred in ischemic phenotype, and 377 (28.8%) in DCM phenotype. In ischemic phenotype, RSF model showed optimal predictive performance; NT-proBNP, serum creatinine, age, and hemoglobin were among the most important variables. In DCM phenotype, RSF model also performed the best, showing superior or comparable performance to other models; Key variables included NT-proBNP, left ventricular end-diastolic diameter, use of angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers, and red blood cell distribution width-standard deviation. In both etiological phenotypes, stepwise Cox regression model derived from the top 20 important variables outperformed ADHF/NT-proBNP risk score in discrimination, calibration, overall performance, and clinical utility.

 

Conclusion: HFrEF/HFmrEF exhibits heterogeneity, etiology is an important source of heterogeneity and may serve as an indicator for simple phenotypic classification. Machine learning-guided prognostic prediction and variable importance ranking reveal differences in risk factors for mortality across different etiological phenotypes, and the derived risk prediction models provide basis for etiological phenotype-specific risk management.

 

Keywords: Heart failure with reduced ejection fraction; Heart failure with mildly reduced ejection fraction; Phenotypic clustering; Machine learning; Prognostic models

 

Part 2

Dynamic Ejection Fraction Classification in Heart Failure: Heart Failure with Improved Ejection Fraction

 

Incidence, Disease Progression, and Clinical Outcomes of Heart Failure with Improved Ejection Fraction: A Study Based on Longitudinal Assessment of Ejection Fraction

 

Background and Aims: The revised universal definition and classification of heart failure (HF) recognizes HF with improved ejection fraction (HFimpEF) as a distinct subtype of HF. It establishes a standardized definition for HFimpEF and emphasizes the importance of longitudinally assessing left ventricular ejection fraction (LVEF). Currently, there is a lack of research exploring the incidence, predictors, and clinical outcomes of HFimpEF based on this updated definition. Longitudinal studies with serial assessments of LVEF are also scarce, rendering the disease progression of HFimpEF remaining largely unknown. We aim to investigate the incidence and disease progression of HFimpEF, and study the predictors and outcomes related to HFimpEF and its dynamic progression in a longitudinal cohort of hospitalized HF patients with lower LVEF (≤ 40%).

 

Methods: We retrospectively included HF patients with baseline LVEF ≤ 40% and satisfactory echocardiographic follow-ups. Patients were classified according to LVEF changes during follow-up, including: i) HFimpEF: defined as a ≥ 10-point increase in LVEF to > 40%; ii) HF with persistently low LVEF: includes patients who did not meet the HFimpEF criteria during follow-up; iii) Transient HFimpEF (LVEF worsening): defined as a recurrent LVEF ≤ 40% after achieving HFimpEF; iv) Persistent HFimpEF: defined as sustained improvement in LVEF to > 40% during follow-up; v) LVEF re-improvement: defined as an LVEF improvement to > 40% with a ≥ 10% increase following a prior decline to ≤ 40%. Clinical outcomes were all-cause death, cardiovascular death, and HF rehospitalization. The incidence of HFimpEF and subsequent dynamic LVEF changes was estimated and presented by cumulative incidence function curves. Variables associated with HFimpEF and subsequent dynamic LVEF changes were identified through decision tree analysis. The relationships between HFimpEF, subsequent dynamic LVEF changes, and outcomes were assessed using Cox proportional hazards models and competing risk models. Moreover, the study quantitatively compared the longitudinal LVEF trajectories of patients with different survival outcomes to identify LVEF trajectory patterns potentially indicative of poor prognosis. All analyses were repeated in sex subgroups to explore potential sex-specific disparities.

 

Results: During a median follow-up of 47.9 months, 517 of 923 patients met HFimpEF criteria; 65.0% of the HFimpEF patients improved LVEF within 12 months. Compared to patients with persistently low LVEF, HFimpEF patients had lower risks of all-cause death (adjusted hazard ratio [aHR] = 0.16, P < 0.001), cardiovascular death (aHR = 0.19, P < 0.001), and HF rehospitalization (aHR = 0.39, P < 0.001). However, 160 HFimpEF patients experienced LVEF worsening during follow-up; their risks for adverse events were higher (aHR = 1.89 for all-cause death, aHR = 2.13 for cardiovascular death, aHR = 2.13 for HF rehospitalization, P < 0.05 for all) compared to persistent HFimpEF patients, and their capability of LVEF re-improvement was diminished. After repeating above analyses in sex subgroups, no significant differences were observed between men and women. An inverted U-shaped LVEF profile for HFimpEF—characterized by a slow, modest increase followed by a decline—portended a higher mortality risk.

 

Conclusion: LVEF improvement was observed in 56.0% of hospitalized HF patients with LVEF ≤ 40% during follow-up, but nearly one-third of these HFimpEF patients experienced subsequent LVEF worsening. LVEF worsening not only indicates a higher risk of adverse events but also suggests a lower likelihood of cardiac reverse remodeling. Longitudinal assessment of LVEF holds significant clinical value in facilitating the identification of HFimpEF patients, monitoring of disease progression, and guiding risk stratification.

 

Keywords: Heart failure with improved ejection fraction; Left ventricular ejection fraction; Serial measurement; Prognosis