基于3类属性预测颈动脉斑块的随机森林方法研究

Predicting carotid plaque by random forest approach with three kinds of attributes

目的:针对B超诊断颈动脉斑块的局限性,提出一种基于3类属性预测颈动脉斑块的随机森林方法。方法:基于某院5993例糖尿病患者的脱敏数据,在已有研究基础上,初步选择影响颈动脉斑块的10个属性(性别、年龄、糖尿病病程、甘油三酯、高密度脂蛋白、低密度脂蛋白、总胆固醇、空腹血糖、糖化血红蛋白和空腹胰岛素),分别利用决策树、逻辑模型树、自助聚合和随机森林等多种机器学习模型来拟合脱敏数据,并通过十折交叉验证来验证拟合性能,最后统计各属性在预测颈动脉斑块中的贡献度。结果:实验表明,随机森林在查准率(0.808)、查全率(0.806)、F1值(0.805)、AUC(0.897)方面均优于其他模型。年龄、低密度脂蛋白和糖化血红蛋白这3个属性对预测颈动脉斑块的贡献度最大,利用这3类属性训练而成的模型可达到不错的预测效果。结论:基于年龄、低密度脂蛋白和糖化血红蛋白等常规指标训练而成的随机森林模型简单、高效、成本低且预测效果好,理论上能够作为诊断颈动脉斑块的辅助方法。

Objective To propose a random forest approach based on three kinds of attributes for predicting carotid plaque so as to overcome the limitation of B-type ultrasonic diagnosis.Methods Totally 10 attributes(gender,age,duration of diabetes,triglycerides,HDL,LDL,total cholesterol,fasting glucose,hemoglobin A1c and fasting insulin)affecting carotid plaque were initially selected based on existing studies with the desensitization data of 5993 diabetic patients in some hospital,various machine learning models such as decision tree,logistic model tree,bootstrap aggregating and random forest were used to fit the desensitization data respectively,then the fitting performance was verified by ten-fold cross-validation,and finally the contribution of each attribute was counted in predicting carotid plaque.Results Experiments showed that random forest outperformed the rest on precision(0.808),recall(0.806),F1(0.805)and AUC(0.897).The three attributes of age,LDL and hemoglobin A1c contributed the most to the prediction of carotid plaque,and a model trained with these three attributes could achieve good prediction results.Conclusion A random forest model trained on conventional indicators such as age,LDL and hemoglobin A1c is simple,efficient and low cost,and theoretically can be used as an aid for diagnosing carotid plaque.