Lu Ree, Lumish Heidi S, Hasegawa Kohei, Maurer Mathew S, Reilly Muredach P, Weiner Shepard D, Tower-Rader Albree, Fifer Michael A, Shimada Yuichi J
Division of Cardiology, Department of Medicine, Columbia University Irving Medical Center, New York, NY, USA.
Department of Emergency Medicine, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA.
Eur J Heart Fail. 2025 Feb;27(2):275-284. doi: 10.1002/ejhf.3546. Epub 2024 Dec 18.
Atrial fibrillation (AF) is the most common sustained arrhythmia among patients with hypertrophic cardiomyopathy (HCM), leading to increased symptom burden and risk of thromboembolism. The HCM-AF score was developed to predict new-onset AF in patients with HCM, though sensitivity and specificity of this conventional tool are limited. Thus, there is a need for more accurate tools to predict new-onset AF in HCM. The objective of the present study was to develop a better model to predict new-onset AF in patients with HCM using machine learning (ML).
In this prospective, multicentre cohort study, we enrolled 1069 patients with HCM without a prior history of AF. We built a ML model (logistic regression with Lasso regularization) using clinical variables. We developed the ML model using the cohort from one institution (training set) and applied it to an independent cohort from a separate institution (test set). We used the HCM-AF score as a reference model. We compared the area under the receiver-operating characteristic curve (AUC) between the ML model and the reference model using the DeLong's test. Median follow-up time was 2.1 years, with 128 (12%) patients developing new-onset AF. Using the ML model developed in the training set to predict new-onset AF, the AUC in the test set was 0.84 (95% confidence interval [CI] 0.77-0.91). The ML model outperformed the reference model (AUC 0.64; 95% CI 0.54-0.73; DeLong's p < 0.001). The ML model had higher sensitivity (0.82; 95% CI 0.65-0.93) than that of the reference model (0.67; 95% CI 0.52-0.88). The ML model also had higher specificity (0.76; 95% CI 0.71-0.81) than that of the reference model (0.57; 95% CI 0.41-0.70). Among the most important clinical variables included in the ML-based model were left atrial volume and diameter, left ventricular outflow tract gradient with exercise stress and at rest, late gadolinium enhancement on cardiac magnetic resonance imaging, peak heart rate during exercise stress, age at diagnosis, positive genotype, diabetes mellitus, and end-stage renal disease.
Our ML model showed superior performance compared to the conventional HCM-AF score for the prediction of new-onset AF in patients with HCM.
心房颤动（AF）是肥厚型心肌病（HCM）患者中最常见的持续性心律失常，会导致症状负担加重和血栓栓塞风险增加。HCM-AF评分用于预测HCM患者新发AF，不过这种传统工具的敏感性和特异性有限。因此，需要更准确的工具来预测HCM患者的新发AF。本研究的目的是使用机器学习（ML）开发一个更好的模型来预测HCM患者的新发AF。
在这项前瞻性、多中心队列研究中，我们纳入了1069例无AF病史的HCM患者。我们使用临床变量构建了一个ML模型（带Lasso正则化的逻辑回归）。我们使用来自一个机构的队列（训练集）开发ML模型，并将其应用于来自另一个机构的独立队列（测试集）。我们将HCM-AF评分作为参考模型。我们使用德龙检验比较了ML模型和参考模型之间的受试者工作特征曲线下面积（AUC）。中位随访时间为2.1年，有128例（12%）患者发生新发AF。使用训练集中开发的ML模型预测新发AF，测试集中的AUC为0.84（95%置信区间[CI]0.77-0.91）。ML模型的表现优于参考模型（AUC 0.64；95%CI 0.54-0.73；德龙检验p<0.001）。ML模型的敏感性（0.82；95%CI 0.65-0.93）高于参考模型（0.67；95%CI 0.52-0.88）。ML模型的特异性（0.76；95%CI 0.71-0.81）也高于参考模型（0.57；95%CI 0.41-0.70）。基于ML的模型中纳入的最重要临床变量包括左心房容积和直径、运动应激和静息时的左心室流出道梯度、心脏磁共振成像上的延迟钆增强、运动应激时的峰值心率、诊断时的年龄、阳性基因型、糖尿病和终末期肾病。
在预测HCM患者新发AF方面，我们的ML模型表现优于传统的HCM-AF评分。
