医疗卫生装备

目的：基于梯度提升决策树的集成学习(eXtreme gradient boosting,XGBoost)算法进行2型糖尿病(type 2 diabetes mellitus,T2DM)合并颈动脉斑块的风险预测及影响因素研究，以期为该疾病的发现、预防和治疗提供有价值的指导。方法：收集2019—2023年上海市某三甲医院3 707例T2DM患者的临床数据，包括2 359例横断面数据和1 348例纵向随访数据。采用方差分析法和Pearson相关系数法对数据进行交叉验证和变量筛选。基于横断面数据和纵向随访数据，采用XGBoost算法构建预测模型，并与人工神经网络、支持向量机以及高斯朴素贝叶斯3种机器学习模型进行对比以验证XGBoost模型的性能；采用沙普利加法解释(Shapley additive explanations,SHAP)分析模型的可解释性，采用特征权重图分析模型中各特征的重要性，并基于横断面数据和纵向随访数据对影响T2DM合并颈动脉斑块的因素进行对比。结果：XGBoost模型在横断面数据中准确率为0.77、精确率为0.62、召回率为0.59、F₁分数为0.51，在纵向随访数据中准确率为0.75、精确率为0.57、召回率为0.68、F₁分数为0.62。综合2类数据下的性能表现，XGBoost模型整体预测效果优于人工神经网络、支持向量机以及高斯朴素贝叶斯3种机器学习模型。SHAP值散点图和特征权重图显示，口干、高血压和年龄是影响T2DM合并颈动脉斑块的主要因素。结论：通过XGBoost算法可以实现对T2DM合并颈动脉斑块的预测。临床实践中应对影响该疾病的相关因素加强监测、评估和干预，以降低罹患该疾病的可能性。

Objective To explore the risk prediction and influencing factors of type 2 diabetes mellitus(T2DM) combined with carotid plaque based on XGBoost algorithm, so as to guide its diagnosis, prevention and treatment. Methods The clinical data from 2019 to 2023 were collected from 3 707 T2DM patients in some Tertiary Grade A hospital of Shanghai, including 2 359 cross-sectional cases and 1 348 longitudinal follow-up cases. Data cross-validation and variable screening were performed using analyses of variance(ANOVA) and Pearson correlation coefficients. A prediction model was constructed based on the cross-sectional and longitudinal follow-up data and the eXtreme gradient boosting(XGBoost) algorithm of the gradient boosting decision tree(GBDT), which had its performance validated by the comparison with three machine learning models of artificial neural network, support vector machine and Gaussian naive Baye; the interpretability of the model established was analyzed using Shapley additive explanations(SHAP) analysis model, the importance of each feature in the model was analyzed using feature weight plots, and the factors affecting combined T2DM and carotid plaque were compared with the cross-sectional data and longitudinal follow-up data. Results The XGBoost model had an accuracy of 0.77, precision of 0.62, recall of 0.59 and F₁ score of 0.51 for the cross-sectional data, and an accuracy of 0.75, precision of 0.57, recall of 0.68 and F₁ score of 0.62 for the longitudinal followup data, for the two kinds of data which gained advantages over the three machine learning models of artificial neural network, support vector machine and Gaussian naive Baye. Scatter plots of SHAP values and feature weight plots showed that dry mouth, hypertension and age were the main factors affecting combined T2DM and carotid plaque. Conclusion The XGBoost algorithm enables prediction of T2DM combined with carotid plaque, and the influencing factors have to underwent monitoring, evaluation and intervention during clinical operations so as to decrease the morbidity of combined T2DM and carotid plaque. [Chinese Medical Equipment Journal,2025,46(12):9-14]