基于XGBoost算法的增生型糖尿病视网膜病变预测模型的建立及可视化实现  

Development and visualization of a XGBoost model for proliferative diabetic retinopathy

作　　者：梁芳 朱广晶 吴婕 余丰 LIANG Fang;ZHU Guangjing;WU Jie;YU Feng(Department of Ophthalmology,Lixiang Eye Hospital of Soochow University,Suzhou,Jiangsu,215000,China)

机构地区：[1]苏州大学附属理想眼科医院眼科,江苏苏州215000

出　　处：《当代医学》2024年第2期8-12,共5页Contemporary Medicine

摘　　要：目的利用H2O机器学习平台,建立增生性糖尿病视网膜病变(PDR)预测模型,旨在为PDR的临床诊疗提供指导。方法选取2019年1月至2021年1月于苏州大学附属理想眼科医院体检中心及住院部连续就诊的350例糖尿病视网膜病变(DR)患者作为研究对象,根据DR类型的不同分为非增生性糖尿病视网膜病变(NPDR)组(n=256)与PDR组(n=94)。比较两组临床资料。尝试多种机器学习算法,建立预测模型,通过绘制ROC曲线、计算混淆矩阵,明确最佳模型,并通过可加性解释模型(SHAP)分析、局部可解析性算法(LIME)及部分依赖图(PDP)等可视化方法呈现重要特征在预测中的作用。结果两组糖尿病(DM)病程、体力活动、体重指数(BMI)、腰臀比、血压、高密度脂蛋白胆固醇、空腹血糖、空腹胰岛素、糖化血红蛋白、谷丙转氨酶(ALT)水平及高血压、脂肪肝、吸烟、饮酒占比比较差异有统计学意义(P<0.05)。机器学习最佳模型为XGBoost,该模型Gini值0.997,R^(2)为0.926,重要特征包括DM病程、高密度脂蛋白胆固醇及空腹胰岛素。结论本研究基于XGBoost算法建立PDR临床预测模型,通过可视化呈现重要特征在预测中的作用,为PDR临床诊疗提供了新的思路。Objective To develop a machine learning model for proliferative diabetic retinopathy(PDR)diagnosis on H2O platform,in order to provide guidance for the clinical diagnosis and treatment of PDR.Methods A total of 350 patients with diabetic retinopathy(DR)who received continuous treatment in the Physical Examination Center and Inpatient department of Lixiang Eye Hospital of Soochow University,Suzhou from January 2019 to January 2021 were selected as the research subjects,they were divided into the non-proliferative diabetic retinopathy(NPDR)group(n=256)and the PDR group(n=94)according to the different types of DR.The clinical data of the two groups were compared,various machine learning algorithms were tried to establish prediction models,and the optimal model was defined by drawing ROC curve and calculating confusion matrix.Visualization methods such as additive interpretation model(SHAP)analysis,local resolvability algorithm(LIME)and partial dependence graph(PDP)were used to show the role of important features in prediction.Results There were significant differences in the course of diabetes mellitus(DM),physical activity,body mass index(BMI),waist-to-hip ratio,blood pressure,high-density lipoprotein cholesterol,fasting blood glucose,fasting insulin,glycosylated hemoglobin,alanine aminotransferase(ALT)levels and the proportion of hypertension,fatty liver,smoking and drinking between the two groups(P<0.05).The best machine learning model was XGBoost model,Gini was 0.997 and R^(2) was 0.926,the important features included diabetes course,high density lipoprotein cholesterol and fasting insulin.Conclusion This study developed a XGBoost model for PDR diagnosis and visualized the key features,which offers insights for clinical practice.

关 键 词：增生性糖尿病视网膜病变 预测模型 XGBoost算法 机器学习 

分 类 号：R587.2[医药卫生—内分泌] R774.1[医药卫生—内科学]