摘要
目的 通过使用机器学习算法在早期识别2型糖尿病并发的糖尿病周围神经病变(Diabetic Peripheral Neuropathy, DPN),并探索DPN与各项检查指标之间的关系,为预防DPN提供依据。方法 选择2022—2023年上海同济医院确诊的糖尿病患者720例,根据纳排标准,纳入665例患者作为研究对象,其中并发DPN患者340例(DPN组),未并发DPN患者325例(非DPN组)。比较2组患者的临床资料、实验室检查指标及肌电图指标,探讨2型糖尿病患者并发DPN的影响因素。并通过超参数调优机器学习算法,建立2型糖尿病患者并发DPN的预测模型,采用准确度、精度、召回率、F1评分和受试者工作特征曲线下面积(AUC)验证模型的可靠性和临床实用性,采用Shapley加法解释(SHapley Additive exPlanations, SHAP)技术对最优模型输出进行解释。结果 通过筛选特征,构建了随机森林(RF)、Stacking模型、XGBoost、支持向量机等四种机器学习分类算法对DPN进行预测,并且所有分类器的AUC值都超过了可接受的70%阈值。通过分析模型性能显示,支持向量分类器的准确率最佳,为74.84%,其次分别为XGBoost、Stacking模型、RF,并且支持向量分类器在F1 Score和精确率中也表现最佳,分别为0.7665和80.0%;然而,分析ROC曲线显示,XGBoost的曲线下面积(AUC)最高为0.757,其次分别为RF、支持向量机、Stacking模型;Stacking模型在召回率中表现最佳,为88.1%,其次分别为XGBoost、RF、支持向量机。模型变量重要性以及支持向量机的SHAP结果表明,年龄、病程、肾功能指标(PRO、UACR、MALB、24hMALB、UCr)、Ddimer、空腹C肽、维生素D、心血管病变(下肢动脉和颈动脉斑块的长度和厚度乘积、下肢动脉斑块长度)、血小板计数、低密度脂蛋白胆固醇、空腹胰岛素、直接胆红素是DPN的重要预测因素。结论 本研究构建了4种类型的DPN风险预测模型并选出支持向量机模型作为最优模型,通过在机器学习框架内利用患者的临床资料和实验室指标进行糖尿病神经病变诊断,利用预测模型提前筛查患者的得病情况,辅助医生进行初期诊断,提升诊疗效率;利用模型进一步确定与DPN相关的影响因素,通过有效控制和监测这些变量,预防或延缓糖尿病患者周围神经病变的进展可能是可行的。基于各种指标比较了这些模型的预测效果,本研究显示了ML作为DPN强大预测工具的潜力。
Objective To identify diabetic peripheral neuropathy(DPN)complicated by type 2 diabetes mellitus at an early stage using machine learning algorithms,and to explore the relationship between DPN and various routine laboratory examinations,so as to provide a basis for the prevention of DPN.Methods A total of 720 diabetic patients diagnosed at Tongji Hospital,Shanghai from 2022 to 2023 were included as subjects of this study according to the inclusion and exclusion criteria.The patients were divided into a DPN group(340 cases)and a non-DPN group(325 cases).Clinical data,laboratory examination indices,and electromyography examinations were compared between the two groups to explore the influencing factors of DPN in patients with type 2 diabetes mellitus.A hyperparameter-tuned machine learning algorithm was used to establish a prediction model for DPN in patients with type 2 diabetes mellitus.The model's reliability and clinical utility were verified using accuracy,precision,recall,F1 score,and area under the receiver operating characteristic curve(AUC).The SHapley Additive exPlanations(SHAP)technique was employed to explain the output of the optimal model.Results Four machine learning classification algorithms,including random forest(RF),Stacking model,XGBoost,and support vector machine(SVM),were constructed to predict DPN,and all achieved AUC values exceeding the acceptable threshold of 70%.The SVM model demonstrated the highest prediction accuracy at 74.84%,followed by XGBoost(68.69%),Stacking model(65.66%),and RF(63.64%).The SVM model also performed best in F1 score(0.7665)and precision(80.0%).The ROC curve analysis showed that XGBoost had the highest AUC(0.757),followed by RF,SVM,and Stacking model.The Stacking model exhibited the best recall rate at 88.1%,followed by XGBoost,RF,and SVM.The importance of variables in the predictive model and the SHAP results of SVM indicated that age,disease duration,renal function indicators(PRO,UACR,MALB,24hMALB,UCr),D-dimer,fasting C-peptide,vitamin D,cardiovascular disease markers(the product of the length and thickness of lower extremity artery plaques,and the length of lower extremity arterial plaques),platelet count,low-density lipoprotein cholesterol,fasting insulin,and direct bilirubin were significant predictors of DPN.Conclusions Four types of DPN risk prediction models were established,and the support vector machine was selected as the optimal model to assist in the early diagnosis of DPN using clinical data and laboratory indices of patients with type 2 diabetes mellitus within the machine learning framework.The explainable prediction model can be utilized to screen the disease state in advance to assist physicians in early diagnosis,thereby improving the efficiency of medical services.Additionally,the risk factors associated with DPN can be determined by these models,making it possible to prevent or delay the progression of DPN in patients with type 2 diabetes mellitus through active control and monitoring.This study demonstrates the potential of machine learning as a powerful predictive tool for DPN based on the predictive efficacy of various indices compared by these machine learning models.
作者
董丹梦
郑冰妍
吴亚
赵煦
朱丽杰
刘阳
罗梓鸿
谢晓云
Dong Danmeng;Zheng Bingyan;Wu Ya;Zhao Xu;Zhu Lijie;Liu Yang;Luo Zihong;Xie Xiaoyun(Medical School of Anhui University of Science and Technology,Huainan,Anhui 232001,China;School of Mathematical Sciences,Shanghai Jiao Tong University,Shanghai 200240,China;Tongji University School of Medicine,Shanghai 200092,China;Department of Geriatrics,Tongji Hospital Affiliated to Tongji University,Shanghai 200065,China;Xi'an Jiaotong Liverpool University,Suzhou,Jiangsu 215123,China;Interventional Vascular Surgery Department,Shanghai Tenth People's Hospital Affiliated to Tongji University,Shanghai 200072,China)
出处
《齐齐哈尔医学院学报》
2025年第3期201-209,共9页
Journal of Qiqihar Medical University
基金
国家自然科学基金(8207070257)。
关键词
糖尿病周围神经病变
预测模型
2型糖尿病
机器学习
支持向量机模型
影响因素
Diabetic peripheral neuropathy
Prediction model
Type 2 diabetes mellitus
Machine learning
Support vector machine
Influence factor
