使用3D心声学预测扩张性心肌病的不良结果:受惩罚的考克斯回归与机器学习对比
Manman Yang1,2, Bingjie Qu1, Jiacheng Cai1,3
1Henan Institute of Interconnected Intelligent Health Management, Henan Key Laboratory of Chronic Disease Prevention and Therapy & Intelligent Health Management, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, Henan, 450052, China.
BMC cardiovascular disorders
|February 22, 2026
概括
对于扩张性心肌病 (DCM) 患者,机器学习模型提供了高分辨率但差的校准. 一个受惩罚的考克斯回归 (拉索-考克斯) 模型为临床风险分层提供了最好的平衡.
科学领域:
- 心脏病学 心脏病学
- 医疗信息学 医疗信息学
- 生物统计学 生物统计学
背景情况:
- 扩张性心肌病 (DCM) 的风险预测低于最佳.
- 机器学习 (ML) 方法正在探索用于预后建模.
- 先进的3D心声回声测量可以改善风险分层.
研究的目的:
- 将传统的考克斯回归,惩罚性的考克斯回归和ML模型进行DCM风险预测.
- 识别具有最佳区分,校准和可解释性的模型.
- 在预后模型中评估3D心声回声学参数的实用性.
主要方法:
- 对196名DCM患者进行了回顾性队列研究.
- 复合结局的随访:死亡率,心力衰竭再住院,或LVAD植入.
- 开发和评估12个预后模型 (Cox,Lasso-Cox,ML) 使用41个预测因素,包括3D心声学参数.
主要成果:
- ML模型 (随机森林,梯度提升) 在12个月时显示出最高的歧视 (AUC 0.990),但校准不佳.
- 拉索-考克斯模型在24个月内保持了可接受的歧视 (AUC 0.729),校准更好.
- 关键预测因素包括4D右心室射出分数 (4D-RVEF),左心房体积指数 (LAVI),肺动脉静脉压 (PASP) 和三环状平面静脉外流 (TAPSE).
结论:
- 机器学习模型在区分方面表现出色,但在DCM中与校准作斗争.
- 惩罚性考克斯回归 (拉索-考克斯) 提供了最佳的性能和可解释性平衡.
- 拉索-科克斯推用于DCM的临床风险分层和实施研究.
相关概念视频
01:30Cardiomyopathy II: Dilated Cardiomyopathy
648
Dilated cardiomyopathy, or DCM, is a progressive myocardial disorder characterized by ventricular chamber dilation and contractile dysfunction.EtiologyVarious factors can cause DCM, including hypertension and heavy alcohol intake, which contribute to the weakening and enlargement of the heart muscle. Viral infections, such as Coxsackievirus B, adenoviruses, and influenza, can lead to DCM by causing inflammation and damage to heart tissue. Certain chemotherapeutic agents, including daunorubicin,...
648
01:29Cardiomyopathy III: Hypertrophic Cardiomyopathy
553
Hypertrophic cardiomyopathy, or HCM, is an autosomal dominant genetic disorder characterized by asymmetric left ventricular hypertrophy without ventricular dilation. It is more common in men and is typically diagnosed in young, athletic adults.EtiologyHCM is primarily genetic and is caused by mutations in genes encoding sarcomeric proteins. Researchers have identified over 1400 mutations across at least 11 different genes. Among these, the most frequently occurring mutations are found in the...
553