基于冠状动脉CT血管成像的力学组学预测心肌桥近端斑块形成 Translated title: Coronary CT Angiography-Based Mechanomics Predicts Atherosclerotic Plaque Formation in Regions Proximal to Myocardial Bridging

Abstract

目的

利用机器学习的方法评估基于冠状动脉CT血管成像（coronary CT angiography, CCTA）的力学组学对左冠状动脉前降支心肌桥近端斑块形成的预测价值。

方法

回顾性搜集2007年1月–2021年4月在我院行至少2次CCTA检查示左冠状动脉前降支心肌桥且基线左冠状动脉前降支无粥样硬化斑块的患者，两次CCTA检查间隔3个月以上。心肌桥近端粥样硬化斑块的形成为主要终点事件。记录患者的人口学特征和临床危险因素并将不同组别患者按照年龄性别进行 1∶1 匹配。基于CCTA进行计算流体动力学分析。在CCTA图像上测量心肌桥的位置、长度、深度和收缩期狭窄指数，以及提取左冠状动脉前降支近段的力学组学参数。利用多因素Cox回归筛选有意义特征，采用随机森林算法挑选力学组学特征并进行后续建模，为每个患者的力学组学特征进行赋值评分。采用对数秩检验方法和Kaplan-Meier图探讨力学组学模型对未来斑块形成的预测价值。采用受试者工作特征曲线评估不同心肌桥组对斑块形成的预测价值。

结果

该研究共纳入104例左冠状动脉前降支心肌桥患者，其中52例患者心肌桥近端斑块形成，中位随访时间为3.0年。总人群的平均年龄为（54.56±10.56）岁，75.00%（78/104）为男性。除吸烟史外（21.15% vs. 5.77%, P=0.04），其余的临床及解剖学特征在有无斑块形成组间差异无统计学意义（所有 P>0.05）。入组患者按照7∶3分为训练集（ n=74）和验证集（ n=30），随机森林算法构建的力学组学模型按照赋分≥0.46和<0.46为赋分较高组和赋分较低组，力学组学模型在验证组的敏感性、准确性分别为0.87（0.58～0.98）和0.63（0.44～0.79）。多因素Cox回归模型中，力学组学（危险比=10.58；95%置信区间：3.23～34.64， P≤0.001）与斑块形成呈正相关。通过对数秩检验力学组学赋分较高组相对于赋分较低组在心肌桥近端更容易形成斑块（ P<0.001）。全部人群、训练集、验证集、表浅心肌桥组、长心肌桥组和短心肌桥组力学组学预测斑块形成曲线下面积分别为0.88（0.82～0.95）、0.89（0.82～0.96）、0.86（0.74～0.99）、0.92（0.86～0.97）、0.86（0.74～0.98）和0.91（0.83～0.98）。

结论

力学组学对左冠状动脉前降支心肌桥近端动脉粥样硬化斑块形成有一定预测价值。

Translated abstract

Objective

To assess with machine learning the predictive value of mechanomics derived from coronary CT angiography (CCTA) for atherosclerotic plaque formation in regions proximal to myocardial bridging (MB) in the left anterior descending coronary artery (LAD).

Methods

This retrospective study included a cohort of patients with MB in LAD and no atherosclerotic plaque formation in LAD as confirmed by two CCTA conducted between January 2007 and April 2021 at our hospital. The interval between the two CCTA examinations was more than 3 months. The primary endpoint was the formation of atherosclerotic plaques in regions proximal to the myocardial bridging. Patient demographic characteristics and clinical risk factors were documented. Then, the patients were matched by age and sex in a 1-to-1 ratio and divided into two groups, those with plaque formation and those without plaque formation. Computational fluid dynamics analysis was performed based on CCTA. Key anatomical parameters of MB, including location, length, depth, and systolic compression index, were meticulously measured on the CCTA images. Mechanomic data were extracted from the region proximal to the MB. A multivariate Cox regression analysis was performed to identify significant features. A random forest algorithm was used to select mechanomic features for subsequent modeling and to assign scores for each patient's mechanomic features. The log-rank test and Kaplan-Meier curves were used to investigate the mechanomic model's predictive performance concerning plaque formation. Additionally, the operator characteristic curves were applied to evaluate how well the model could predict plaque formation across various myocardial bridge subgroups.

Results

A total of 104 patients with LAD MB were recruited. The mean age of the subjects were (54.56±10.56) years and 75.00% (78/104) of them were male. Among them, 52 developed plaque formation over a median follow-up period of 3.0 years. Apart from a smoking history, which was more prevalent in the group with plaque formation than that in the group without plaque formation (21.15% vs. 5.77%, P=0.04), no significant differences between the groups were observed in terms of the other clinical or anatomical characteristics (all P≤0.05). The participants were divided into a training set ( n=74) and a validation set ( n=30) at a 7∶3 ratio. With the mechanomics model constructed using the random forest algorithm, the patients were classified into a high-score group (≥0.46) and a low-score group (<0.46) based on a cutoff score of 0.46. The mechanomics model achieved a sensitivity of 0.87 (0.58-0.98) and an accuracy of 0.63 (0.44-0.79) in the validation set. The multivariate Cox regression model revealed a strong positive association between mechanomics and plaque formation (hazards ratio [HR]: 10.58; 95% confidence interval [CI]: 3.23-34.64, P<0.001). The log-rank test showed that the high-score group in the mechanomics model was more likely to develop plaques at the proximal regions of the myocardial bridge compared to the low-score group ( P<0.001). The area under the curve (AUC) for plaque formation, as predicted by the model, was 0.88 (95% CI: 0.82-0.95) for the entire population, 0.89 (95% CI: 0.82-0.96) for the training set, 0.86 (95% CI: 0.74-0.99) for the validation set, 0.92 (95% CI: 0.86-0.97) for the superficial MB group, 0.86 (95% CI: 0.74-0.98) for the long MB group, and 0.91 (95% CI: 0.83-0.98) for the short MB group.

Conclusion

The mechanomic assessment holds substantial potential as a predictive tool for atherosclerotic plaque formation in regions proximal to MB in LAD.