基于可视化临床影像模型预测冠状动脉易损斑块的价值

目的  探讨基于临床影像征象和影像组学联合模型对冠状动脉易损斑块发生的预测价值，并通过Shapley算法对模型进行可视化分析。方法  回顾性收集2016~2020年南方医科大学顺德医院确诊冠心病并且行2次CCTA检查的患者383例，提取相应区域的影像组学特征。使用多步联合方法筛选出各区域最佳特征后进行联合建模。通过Logistic回归方法筛选重要临床影像征象，最后构建可解释的XGBoost临床影像模型。利用Shapley算法对模型分别进行可视化和特征贡献度解释。结果  相比单区域影像组学模型，多区域影像组学模型展现出更高的预测性能（AUC=0.701）。结合重要临床影像征象的联合模型性能进一步提高（AUC=0.885）。利用Shapley分析算法对特征重要性进行解析，前6个组学特征对模型结果预测具有贡献度，Shapley热图算法实现了易损斑块发生的预测推演可视化过程。结论  临床影像组学联合模型对冠状动脉易损斑块的预测具有较高的准确性和泛化性。可解释机器学习算法的可视化保障了模型的实用性，为临床制定针对性治疗方案提供了一种无创工具。

Value of using visual clinical imaging models to predict vulnerable coronary artery plaques

Objective To explore the predictive value of a combined model based on clinical imaging features and radiomics for the occurrence of vulnerable coronary artery plaques, and visualize the model through Shapley algorithm for further analysis. Methods A retrospective study was conducted on 383 patients diagnosed with coronary heart disease and who underwent two CCTA examinations at Shunde Hospital of Southern Medical University from 2016 to 2020. Radiomics features were extracted from the corresponding regions of interest. A multi-step combined method was used to select the best features from each region for joint modeling. Logistic regression was employed to select important clinical imaging features, and an interpretable XGBoost clinical imaging model was constructed. The Shapley algorithm was utilized to visualize the model and interpret the feature contributions. Results Compared with single-region radiomics models, multi-region radiomics models demonstrated higher predictive performance (AUC=0.701). Combining important clinical imaging features with the joint model improved the performance even further (AUC=0.885). By analyzing the feature importance using the Shapley analysis algorithm, it was found that the first six radiomics features contributed significantly to the model's predictive results. The Shapley heatmap algorithm visualized the prediction process of vulnerable plaque occurrence. Conclusion The clinical radiomics combined model shows high accuracy and generalizability in predicting vulnerable coronary artery plaques. The visualization of interpretable machine learning algorithms ensures the practicality of the model, providing a non-invasive tool for the development of targeted treatment plans in clinical practice.