Abstract: Background　There are a large number of type 2 diabetes mellitus（T2DM）patients in China at present，it is urgent to develop a simple and effective risk assessment tool for subclinical atherosclerosis in T2DM. Objective　To construct a multi-layer artificial neural network classification model for predicting subclinical atherosclerosis in T2DM patients and verify its prediction accuracy based on multiple indicators. Methods　A total of 3 627 T2DM patients who were hospitalized in the Third Affiliated Hospital of Anhui Medical University from January 2020 to December 2016 were selected. All of them underwent color Doppler ultrasound of bilateral carotid arteries，including 2 196 cases detected subclinical atherosclerosis（observation group）and 1 431 cases did not detected（control group）. The general information，laboratory examination indicators and fatty liver occurrence of the two groups were compared and a multi-layer artificial neural network classification model was constructed accordingly. A total of 3 027 patients were randomly selected from the 3 627 T2DM patients as the training set，and the remaining 600 patients as the test set to verify the prediction accuracy of the multi-layer artificial neural network classification model. Results　There were no significant differences of BMI，DBP，proportion of people with smoking history，proportion of people with alcohol consumption history，alcohol consumption，DBiL，total protein，AST，SUA，TG，LDL-C/HDL-C ratio，TSH，FT3，FT4，HbA1c，FBG，fasting C-peptide，HOMA-C-peptide index，proportion of severe fatty liver between two groups（P>0.05）；but compared with control group，observation group showed higher female ratio，SBP，proportion of hypertension history，globulin，total bile acid，BUN，Scr，cystatin C，UARE，TC，LDL-C，HDL-C，WBC and neutrophil count，older age，larger smoking amount，longer course of disease，smoking time，drinking time（P<0.05），lower proportion of family history of diabetes，TBiL，IBiL，albumin，ALT，GFR，TG/HDL-C ratio，lymphocyte count，red blood cell count，Hb and incidence of fatty liver（P<0.05）. Combining clinical practice，the above 49 indicators are used as input variables to construct the multi-layer artificial neural network classification model；in the testing set，the accuracy of Logistic model for predicting subclinical atherosclerosis in T2DM was 59%，that of multi-layer artificial neural network classification model was 76% when the number of plies was 3. Conclusion　The multi-layer artificial neural network classification model successfully constructed in this study has a high accuracy in predicting subclinical atherosclerosis in T2DM patients，and can be used as a risk assessment tool for subclinical atherosclerosis in T2DM patients.

Key words: Diabetes mellitus, type 2, Atherosclerosis, Subclinical atherosclerosis, Neural networks, computer, Deep learning, Models, theoretical

摘要： 背景　现阶段我国2型糖尿病（T2DM）患者数量较多，亟须开发简单、有效的亚临床动脉粥样硬化发生风险评估工具。目的　依据多项指标构建预测T2DM患者亚临床动脉粥样硬化的多层人工神经网络分类模型并验证其预测准确性。方法　选取2010年1月至2016年12月在安徽医科大学第三附属医院住院的T2DM患者3 627例，均行双侧颈动脉彩色多普勒超声检查，其中检出亚临床动脉粥样硬化者2 196例（观察组），未检出亚临床动脉粥样硬化者1 431例（对照组）。比较两组患者一般资料、实验室检查指标及脂肪肝发生情况并据此构建多层人工神经网络分类模型。从3 627例T2DM患者中随机选取3 027例患者作为训练集，其余600例患者作为测试集，验证多层人工神经网络分类模型的预测准确性。结果　两组患者体质指数、舒张压、有吸烟史者所占比例、有饮酒史者所占比例、饮酒量、直接胆红素、总蛋白、天冬氨酸氨基转移酶、血尿酸、三酰甘油、低密度脂蛋白胆固醇/高密度脂蛋白胆固醇比值、促甲状腺激素、游离三碘甲状腺原氨酸、游离甲状腺素、糖化血红蛋白、空腹血糖、空腹C肽、HOMA-C肽指数、严重脂肪肝所占比例比较，差异无统计学意义（P>0.05）；观察组患者女性所占比例、收缩压、有高血压病史者所占比例、球蛋白、总胆汁酸、尿素氮、血肌酐、胱抑素C、尿微量白蛋白排泄率、总胆固醇、低密度脂蛋白胆固醇、高密度脂蛋白胆固醇、白细胞计数、中性粒细胞计数、糖化血红蛋白、空腹血糖高于对照组，年龄、吸烟量大于对照组，病程、吸烟时间、饮酒时间长于对照组，有糖尿病家族史者所占比例、总胆红素、间接胆红素、白蛋白、丙氨酸氨基转移酶、肾小球滤过率、三酰甘油/高密度脂蛋白胆固醇比值、淋巴细胞计数、红细胞计数、血红蛋白、脂肪肝发生率低于对照组（P<0.05）。结合临床实际，将上述49项指标作为输入变量构建多层人工神经网络分类模型；在测试集上，Logistic模型预测T2DM患者亚临床动脉粥样硬化的准确率为59%，而多层人工神经网络分类模型隐藏层数为3时预测T2DM患者亚临床动脉粥样硬化的准确率为76%。结论　本研究构建的多层人工神经网络分类模型对T2DM患者亚临床动脉粥样硬化的预测准确率较高，可作为T2DM患者亚临床动脉粥样硬化发生风险评估工具。

关键词: 糖尿病，2型, 动脉粥样硬化, 亚临床动脉粥样硬化, 神经网络，计算机, 深度学习, 模型，理论