Liang Guizi, Peng Ziyu, Hu Jie, Liang Limei, Xie Dongwei, Chen Xing, Zhi Rentao, Guan Xuechun, Deng Yan
Department of Ultrasound Medicine, The First Affiliated Hospital of Guangxi Medical University, Nanning, Guangxi, China.
Department of Radiology, The First Affiliated Hospital of Guangxi Medical University, Nanning, Guangxi, China.
Ultrasound Med Biol. 2025 Nov;51(11):1991-2000. doi: 10.1016/j.ultrasmedbio.2025.07.010. Epub 2025 Aug 5.
This study aims to construct a low-cost echocardiographic myocardial texture-based screening model to identify patients with myocardial fibrosis and filter out non-fibrosis patients, reducing unneeded contrast use and offering an efficient screening method for resource-limited hospitals.
This study enrolled 149 patients with hypertrophic cardiomyopathy (HCM), 105 (70%) of whom were late gadolinium enhancement (LGE) positive. Using stratified random sampling, we divided the patients into training and test sets at a 6:4 ratio. Logistic regression was subsequently used to identify clinical and echocardiographic features. Based on the radiomics score (RadScore), 10 machine learning models were constructed, and cost-sensitive learning was applied to the training set to address class imbalance, with fivefold cross-validation optimizing hyperparameters. Finally, radiomics, echocardiographic (E), and combined models were constructed using an extreme tree algorithm, and their performance was evaluated using receiver operating characteristic, calibration and decision curve analyses.
In the test set, the combined model outperformed both the radiomic and E models, with an area under the receiver operating characteristic curves of 0.91 (95% CI: 0.83, 0.97), a sensitivity of 0.81, a specificity of 0.83 and an F1 score of 0.85 for detecting LGE-positive patients. Calibration and decision curve analyses demonstrated good calibration for the combined model and greater net benefit than the other models did.
A machine learning model based on myocardial texture features has the potential to be a non-invasive tool for predicting myocardial fibrosis in patients with HCM and is particularly suitable for screening.
本研究旨在构建一种基于超声心动图心肌纹理的低成本筛查模型，以识别心肌纤维化患者并筛选出非纤维化患者，减少不必要的造影剂使用，并为资源有限的医院提供一种高效的筛查方法。
本研究纳入了149例肥厚型心肌病（HCM）患者，其中105例（70%）延迟钆增强（LGE）呈阳性。采用分层随机抽样，将患者按6:4的比例分为训练集和测试集。随后使用逻辑回归来识别临床和超声心动图特征。基于影像组学评分（RadScore）构建了10个机器学习模型，并将成本敏感学习应用于训练集以解决类别不平衡问题，采用五折交叉验证优化超参数。最后，使用极端树算法构建影像组学、超声心动图（E）和联合模型，并使用受试者工作特征、校准和决策曲线分析来评估它们的性能。
在测试集中，联合模型的表现优于影像组学模型和E模型，检测LGE阳性患者的受试者工作特征曲线下面积为0.91（95%CI：0.83，0.97），灵敏度为0.81，特异度为0.83，F1评分为0.85。校准和决策曲线分析表明联合模型校准良好，净效益高于其他模型。
基于心肌纹理特征的机器学习模型有可能成为预测HCM患者心肌纤维化的非侵入性工具，特别适用于筛查。
