Stacking集成学习算法验证动脉损伤对糖尿病早期检测的意义

背景:糖尿病可引起广泛的动脉结构和功能病理变化，导致动脉僵硬度增加、顺应性降低和动脉弹性降低。本研究从动脉损伤的角度，实现对尚未出现临床表现但有动脉损伤的糖尿病患者的早期检测。方法:动脉损伤会导致血管的力学参数发生变化，而脉搏信号的波形变化与心血管系统的力学参数变化密切相关。通过9级小波对糖尿病患者脉搏信号进行分解，提取cD8、cD7、cD6系数（中高频成分，代表信号细节特征），作为能够反映动脉损伤程度的特征，将特征矩阵输入到10折交叉验证模型的Stacking集成学习模型中，设置第一层的4个基学习器为SVM、Random Forest、XGBoost、Extra Trees，第二层的元学习器是KNN。结果:单个机器学习模型可以达到90%以上的准确率。Stacking集成学习算法的准确率比单一机器学习模型高4%~5%，ROC曲线下面积提高1%~6%。结论:小波分解得到的脉搏信号cD8、cD7、cD6系数可以有效反映糖尿病引起的动脉损伤程度，因此动脉损伤对糖尿病的早期检测具有一定的指导意义。Stacking 集成学习算法将多个模型的优势结合起来生成一个新模型，可以获得比单一模型更好的性能。

Verifying the significance of arterial injury for early detection of diabetes by Stacking ensemble learning algorithm

Abstract: Background Diabetes can cause extensive pathological changes in the structure and function of arteries, leading to increased arterial stiffness, decreased compliance, and decreased arterial elasticity. From the perspective of arterial injury, this study aims to realize the early detection of diabetes in patients who have not yet appeared clinical manifestations of diabetes but have arterial injury. Methods Arterial injury leads to mechanical parameters changes in the vascular system. The waveform changes of pulse signals are closely related to mechanical parameters changes in the cardiovascular system. By decomposing the pulse signals of diabetic patients with 9-level wavelet, cD8, cD7 and cD6 coefficients (medium-high frequency components, representing the features in signal details) were extracted as features that reflect the degree of arterial injury. The feature matrix was input into the Stacking ensemble learning algorithm of the 10-fold cross-validation model, with SVM, Random Forest, XGBoost and Extra Trees as the 4 base-learners of the first layer, and KNN as the meta-learner of the second layer. Results A single machine learning model could achieve an accuracy higher than 90%. Stacking ensemble learning algorithm was 4%-5% higher than a single machine learning model in accuracy, and 1%-6% higher in area under the ROC curve (AUC). Conclusion The cD8, cD7, and cD6 coefficients of pulse signals obtained by wavelet decomposition can effectively reflect the degree of arterial injury caused by diabetes. Therefore, arterial injury has certain guiding significance for the early detection of diabetes. Stacking ensemble learning algorithm that combines the advantages of multiple models to generate a new model can achieve better performance than single models.