基于中医体质辨识的糖尿病风险评估模型建立与验证

背景　中医体质因素在糖尿病的发生发展过程中具有重要作用，但目前糖尿病的预测预警模型仅涵盖一般人口学资料、客观检查指标、生活方式等内容。在糖尿病风险评估模型中纳入中医体质辨识内容，对有针对性地防治糖尿病的发生发展具有重要意义。目的　根据健康体检数据建立基于中医体质辨识的糖尿病风险模型并对其进行验证。方法　于2016年1月-2018年12月，以2014-2015年某省级综合性医院健康管理中心体检数据为训练集数据（n=30 951），对是否患糖尿病进行单因素和多因素Logistic回归分析，纳入有意义的影响因素指标建立糖尿病风险评估模型；以2016-2017年的健康体检数据作为测试集数据（n=24 061），采用受试者工作特征曲线（ROC）对模型进行验证。结果　训练集人群中，患有糖尿病者1 315例（4.25%），未患糖尿病者29 636例（95.75%）。多因素Logistic回归分析结果显示，logit（P）（糖尿病患病情况）=-4.632-0.198×（女）+0.864×（年龄45~59岁）+1.684×（年龄≥60岁）+0.635×（高血压）+0.149×（超重）+0.376×（肥胖）-0.531×（偏轻）-0.234×（淋巴细胞百分数偏高）+0.279×（淋巴细胞百分数偏低）+0.304×（红细胞计数异常）-0.430×（红细胞比容偏低）+0.722×（平均红细胞血红蛋白浓度异常）+0.532×（血小板分布宽度异常）+1.016×（癌胚抗原异常）-0.406×（尿酸异常）+1.341×（肌酐偏低）+0.488×（血尿素氮偏高）+0.473×（三酰甘油异常）+0.257×（总胆固醇偏高）+0.544×（高密度脂蛋白偏低）+0.290×（总蛋白异常）+0.395×（丙氨酸氨基转移酶异常）+0.362×（谷氨酰转肽酶异常）+0.993×（阴虚质）+1.016×（气虚质）+0.601×（痰湿质）。模型验证结果显示，训练集ROC曲线下面积（AUC）为0.792，95%CI为0.779~0.816（P<0.05），最佳截断值为0.405，灵敏度为0.771，特异度为0.690；测试集验证准确率达到95.69%，Kappa=0.636（P<0.001）。结论　初步构建了糖尿病风险评估模型，且此模型具有较高诊断效应。中医体质辨识作为重要的影响因素纳入糖尿病发病风险评估的模型中来，可以提高其预测能力，为糖尿病的中医药早期防治提供一定的依据。

Establishment and Verification of a Diabetes Risk Assessment Model Based on TCM Constitution

Background　Physical constitution is supposed to play an important role in the occurrence and development of diabetes mellitus in traditional Chinese medicine（TCM）.However，current models for predicting and warning diabetes only cover general demographic data，objective inspection indicators，lifestyle and so on.It is of great significance to incorporate TCM constitution into the diabetes risk assessment model for the prevention and treatment of diabetes.Objective　To establish and verify a TCM constitution-based diabetes risk assessment model using the health check-up data.Methods　Health checkup data during 2014 to 2017 were obtained from the physical examination center of a provincial hospital from January 2016 to December 2018.In particular，the data of those（n=30 951） undergoing health checkup during 2014 to 2015 were used as the training data.Univariate and multivariate logistic regression were adopted to explore factors associated with diabetes in this group.Then a diabetes risk assessment model was developed with the identified risk factors for diabetes incorporated，and was verified using the data of those（test data，n=24 061） undergoing health check up during 2016 to 2017.ROC curve of the model in predicting diabetes was plotted and analyzed.Results　Of those undergoing health checkup during 2014 to 2015，1 315 were found with diabetes（4.25%），and 29 636（95.75%） without.Multivariate Logistic regression analysis showed that Logit（P）（prevalence of diabetes mellitus）=-4.632-0.198 ×（female）+0.864×（age 45-59）+1.684×（age≥60）+0.635×（hypertension）+ 0.149×（overweight）+0.376 ×（obesity）-0.531×（underweight）-0.234×（high lymphocyte percentage）+ 0.279×（low lymphocyte percentage）+0.304 ×（abnormal RBC count）-0.430×（low hematocrit）+0.722×（abnormal mean corpuscular hemoglobin concentration）+ 0.532×（abnormal platelet distribution width）+1.016×（abnormal carcinoembryonic antigen）-0.406×（abnormal uric acid）+1.341×（low creatinine）+0.488×（high blood urea nitrogen）+0.473×（abnormal triglyceride）+ 0.257×（high cholesterol）+0.544×（low HDL）+0.290×（abnormal TP）+ 0.395×（abnormal alanine aminotransferase）+ 0.362×（abnormal glutamyl transpeptidase）+0.993×（Yin deficiency）+ 1.016×（Qi deficiency）+0.601×（Phlegm dampness）．The verification results showed that the AUC of the model was 0.792〔95%CI（0.779-0.816，P<0.05）〕，the optimal cut-off value was 0.405 with a sensitivity of 0.771，and a specificity of 0.690.The accuracy of the model in identifying diabetes in the 24 061 physical examinees reached 95.69%，Kappa coefficient=0.636（P<0.001）．Conclusion　Our diabetes risk assessment model proves to be highly accurate，which may be attributed to the inclusion of TCM constitution，a major factor closely associated with diabetes，providing a basis for early TCM-based prevention and treatment of the disease.