天津医药

天津医药 ›› 2024, Vol. 52 ›› Issue (7): 775-780.doi: 10.11958/20231584

• 应用研究 • 上一篇    下一篇

基于机器学习对2型糖尿病肾病预测模型的构建及验证

王娴1(), 刘霞明1, 陈曼玉1, 赵君1, 王立东2,∆()   

  1. 1 承德医学院研究生院（邮编067000）
    2 承德市中心医院内分泌科
  • 收稿日期:2023-10-27 修回日期:2024-02-06 出版日期:2024-07-15 发布日期:2024-07-11
  • 通讯作者: ∆E-mail：whlfdjs@126.com
  • 作者简介:王娴（1996），女，硕士在读，主要从事糖尿病发病机制的基础及临床方面研究。E-mail：1772575329@qq.com

Construction and verification of prediction model of type 2 diabetic nephropathy based on machine learning

WANG Xian1(), LIU Xiaming1, CHEN Manyu1, ZHAO Jun1, WANG Lidong2,∆()   

  1. 1 Graduate School of Chengde Medical University, Chengde 067000, China
    2 Department of Endocrinology, Chengde Central Hospital
  • Received:2023-10-27 Revised:2024-02-06 Published:2024-07-15 Online:2024-07-11
  • Contact: ∆E-mail：whlfdjs@126.com

RichHTML

0

PDF (PC)

133

摘要：

目的 寻找2型糖尿病（T2DM）患者糖尿病肾病（DKD）的独立预测因子，构建DKD发病风险的最佳机器学习（ML）模型并进行验证。方法 以2019年10月—2020年9月于承德市中心医院内分泌科住院治疗的528例T2DM患者为研究对象，随机分为训练集（370例）和验证集（158例），训练集依据是否合并DKD分为DKD组（89例）和非DKD组（281例）。单因素分析患者的一般资料和辅助检查，将其中有意义的变量通过最小绝对收缩和选择算法（LASSO）回归筛选最佳预测因子，将LASSO回归筛选出的最佳预测因子通过Logistc回归（LR）、K近邻（KNN）、支持向量机（SVM）、决策树（DT）、随机森林（RF）、朴素贝叶斯（NB）、人工神经网络（ANN）和极致梯度提升（XGBoost）8种ML算法，经3重交叉验证建立预测模型，通过比较受试者工作特征（ROC）曲线、Delong检验、GiViTI校准曲线选择最佳预测模型。采用决策曲线分析（DCA）评估模型的临床实用性。结果 年龄、丙氨酸转氨酶（ALT）、肌酐（Cr）、甘油三酯（TG）、胱抑素C（Cys-C）、25羟基维生素D[25（OH）D]、单核细胞计数（Mon）是DKD的独立预测因子。8种ML模型中，ANN模型表现最佳。GiViTI校准曲线提示模型具有较好的准确度（P>0.05），决策曲线显示预测模型曲线在0.027~0.612的阈值概率区间具有临床实用价值。结论 本研究构建的预测DKD发生风险的ANN模型有助于早期识别DKD的高危T2DM患者。

关键词: 糖尿病, 2型, 糖尿病肾病, 机器学习, 单核细胞, 神经网络, 计算机, 预测模型

Abstract:

Objective To search for independent predictive factors of diabetic kidney disease (DKD) in patients with type 2 diabetes mellitus (T2DM), construct and validate an optional machine learning (ML) model for the risk of DKD. Methods A total of 528 patients with T2DM, hospitalized in the Endocrinology Department of Chengde Central Hospital from October 2019 to September 2020, were selected as the study objects, and patients were randomly divided into a training set (370 cases), and a validation set (158 cases). The training set was divided into the DKD group (89 cases) and the non-DKD group (281 cases) according to whether DKD existed. The general data and diagnostic examination of patients were performed by univariate analysis, in which variables with statistical differences were used to screen the best predictors by least absolute shrinkage and selection operator (LASSO) regression analysis. The best predictors were used to establish eight ML algorithms by three cross-validation methods, including Logistic regression (LR), K-nearest neighbor (KNN), support vector machine (SVM), decision tree (DT), random forest (RF), naive Bayes (NB), artificial neural network (ANN), and extreme gradient lift (XGBoost). The optimal prediction model was selected by receiver operating characteristic (ROC) curve, Delong test and GiViTI calibration curve. Decision curve analysis (DCA) was used to evaluate the clinical practicability of the model. Results Age, alanine aminotransferase, creatinine, triglyceride, cystatin C, 25-hydroxy vitamin D and monocyte count were independent predictive factors of DKD. Eight ML models were established based on the above 7 predictors, and the ANN model performed best in the 8 ML models. The GiViTiI calibration curve indicated that the model had good accuracy (P>0.05), and the DCA showed that the prediction model curve had clinical practical value in the threshold probability range of 0.027-0.612. Conclusion In this study, the ANN model constructed in this study to predict the risk of DKD is helpful for early discrimination of high-risk T2DM patients with DKD.

Key words: diabetes mellitus, type 2, diabetic nephropathy, machine learning, monocytes, neural networks, computer, prediction model

中图分类号: