基于随机森林算法的糖尿病舌象特征分析和诊断模型研究

Study on the feature analysis and diagnosis model of diabetic tongue based on random forest algorithm

目的:基于随机森林算法联合舌象特征建立一个糖尿病诊断模型并对舌象特征重要性进行进一步分析。方法:使用TFDA-1型舌诊仪采集舌象,使用舌诊分析系统处理舌象图像获得舌象特征,结合基本信息和实验室检查等信息,使用随机森林算法对特征重要性进行分析并建立糖尿病诊断模型。结果:糖尿病组与健康组比较,舌象纹理特征中,Per-all、Per-part、TB-CON、TB-ASM、TB-MEAN、TC-CON、TC-ENT、TC-ASM、TC-MEAN差异有统计学意义(P<0.05,P<0.01);舌象颜色特征中,TB-Cr、TC-Cr、TC-a差异有统计学意义(P<0.05,P<0.01)。根据随机森林特征基尼系数,模型2与模型3前15位共有的舌象特征为TC-CON、TC-ASM、Per-part。模型1、模型2、模型3的准确率分别为0.70、0.68、0.74;灵敏度分别为0.70、0.69、0.78;特异度分别为0.71、0.68、0.71。结论:舌象的颜色和纹理特征在建立糖尿病诊断模型中具有一定作用,加入舌象特征可以提升糖尿病诊断模型的性能。

Objective: To establish a diabetes diagnosis model based on random forest algorithm combined with tongue features and to further analyze the importance of tongue features. Methods: TFDA-1 tongue diagnostic instrument was used to collect the tongue image, and tongue diagnosis analysis system was used to process tongue images to obtain tongue features.Laboratory examinations and other information were used to analyze the importance of features and establish a diabetes diagnosis model using the random forest algorithm. Results: When the diabetic group was compared with the healthy group, in the tongue texture features, the differences of Per-all, Per-part, TB-CON, TB-ASM, TB-MEAN, TC-CON, TC-ENT, TC-ASM and TCMEAN were statistical significance(P<0.05, P<0.01);In the tongue color features, the differences of TB-Cr, TC-Cr and TC-a were statistically significant(P<0.05, P<0.01). According to the Gini coefficient of features calculated by random forest, the top 15 tongue features of model 2 and model 3 are TC-CON, TC-ASM, and Per-part. The accuracy of model 1, model 2, and model 3 were 0.70, 0.68, and 0.74;the sensitivity were 0.70, 0.69, and 0.78;and the specificity were 0.71, 0.68, and 0.71, respectively.Conclusion: The color and texture features of tongue images play a certain role in the establishment of diabetes diagnosis models.Adding tongue features can improve the performance of diabetes diagnosis models.