中医学报

目的：应用机器学习(machine learning, ML)方法，纳入2型糖尿病(type 2 diabetes mellitus, T2DM)阴虚证症状为变量，构建T2DM阴虚证糖化血红蛋白(glycosylated hemoglobin, HbA1c)控制水平预测模型，为丰富T2DM阴虚证内涵及优化HbA1c管理提供参考。方法：收集2020年1月1日至2022年12月1日在北京中医药大学东直门医院东城院区肾病内分泌二科、通州院区肾病内分泌三科，湖南中医药大学、厦门大学附属第一医院就诊的T2DM阴虚证患者资料，包括基本信息及阴虚证症状。采用LASSO回归与逐步回归联合筛选阴虚证相关特征变量，基于筛选结果构建多种机器学习模型进行训练、验证与测试，并利用沙普利加性解释(Shapley additive explanations, SHAP)方法对最优模型进行解释。结果：最终纳入T2DM阴虚证患者1 163例，其中HbA1c≥7%者共747例，HbA1c<7%者共416例。经LASSO与逐步回归筛选T2DM阴虚证13个特征变量，最终纳入渴喜冷饮、口干咽燥、盗汗、怕热、心烦、大便偏干、两目干涩等7项症状变量。经内部验证及外部验证，综合多模型各项指标(准确率、精确率、特异度、召回率、F1分数及AUC)比较显示，随机森林(random forest, RF)模型表现最佳，其准确率为0.814,召回率为0.721,精确率为0.754,特异性为0.868,F1值为0.738,AUC达到0.865。SHAP分析显示，口干咽燥、渴喜冷饮、盗汗、大便偏干、心烦、怕热、两目干涩为影响HbA1c控制的重要特征，其中口干咽燥、渴喜冷饮、盗汗等3个特征变量对于HbA1c控制水平的重要性最高。结论：本研究基于阴虚证特征症状构建了T2DM患者HbA1c控制水平的预测模型，表现出良好的预测性能和临床适用性，有助于丰富T2DM阴虚证的现代内涵，并为个体化血糖管理和中医药方案优化提供参考。

Objective: Using machine learning(ML) methods and incorporating symptoms of Yin deficiency syndrome in type 2 diabetes mellitus(T2DM) as variables, this study aims to construct a prediction model for glycosylated hemoglobin(HbA1c) control levels in T2DM with Yin deficiency syndrome, providing reference to enrich the connotation of Yin deficiency syndrome in T2DM and optimize HbA1c management.Methods: Data of T2DM patients with Yin deficiency syndrome treated from January 1,2020,to December 1,2022,at the Nephrology and Endocrinology Department II of Dongcheng Campus of Dongzhimen Hospital and the Nephrology and Endocrinology Department III of Tongzhou Campus of Dongzhimen Hospital affiliated to Beijing University of Chinese Medicine, Hunan University of Chinese Medicine, and the First Affiliated Hospital to Xiamen University were collected, including basic information and symptoms of Yin deficiency syndrome.LASSO regression and stepwise regression were used jointly to screen features related to Yin deficiency syndrome.Based on the selected features, various machine learning models were constructed for training, validation, and testing.Shapley additive explanations(SHAP) were used to interpret the optimal model.Results: A total of 1,163 T2DM patients with Yin deficiency syndrome were included, of whom 747 had HbA1c ≥7% and 416 had HbA1c <7%.Through LASSO and stepwise regression, 13 characteristic variables of yin deficiency syndrome in T2DM were screened, ultimately including 7 symptom variables: craving for cold drinks, dry mouth and throat, night sweats, aversion to heat, irritability, dry stools, and dry eyes.Internal and external validations, comparing multiple models across various metrics(accuracy, precision, specificity, recall, F1 score, and AUC),showed that the Random Forest(RF) model performed the best, with an accuracy of 0.814,recall of 0.721,precision of 0.754,specificity of 0.868,F1 score of 0.738,and an AUC of 0.865.SHAP analysis indicated that dry mouth and throat, craving for cold drinks, night sweats, dry stools, irritability, aversion to heat, and dry eyes are significant features influencing HbA1c control, with dry mouth and throat, craving for cold drinks, and night sweats being the top three most important variables for HbA1c management.Conclusion: This study constructed a predictive model for HbA1c control in T2DM patients based on the characteristic symptoms of yin deficiency syndrome, demonstrating good predictive performance and clinical applicability, helping to enrich the modern understanding of yin deficiency in T2DM and providing references for personalized glycemic management and optimization of TCM treatment plans.