Diabetes Risk Assessment And Syndrome Diagnosis Model Based On Tongue Image Dat

Objectives1.Look for key tongue features that are highly related to diabetes risk and clarify the clinical significance of these features changes.2.Establish and strengthen the relationship between objective tongue and diabetes risk,and construct a non-invasive diabetes risk prediction model.3.Establish the relationship between the objective tongue and the TCM syndromes of diabetes,and construct the TCM syndrome diagnosis model.MethodsThe subjects came from Shuguang Hospital affiliated to Shanghai University of Traditional Chinese Medicine.Basic information and blood test results were provided by Shuguang Hospital.TFDA-1 tongue diagnosis instrument or TDA-1 tongue diagnosis instrument were used to collect the tongue images of subjects,and the tongue features were calculated by the Tongue diagnosis analysis system(TDAS).1.Analysis of key tongue features related to diabetes riskDimensionality reduction of the original features through principal component analysis(PCA),and logistic regression to analyze the principal components(PCs)of the pre-diabetic population(healthy vs.pre-diabetes)and the diabetic population(healthy vs.diabetes),and establish the relationship between the original features and the PCs through the factor loading method,and screening of key tongue features related to diabetes risk.2.Construct a diabetes risk prediction modelFirst,based on tongue features,combined with different types of features such as basic features and blood features,machine learning algorithms were used to construct regression and classification models for diabetes risk prediction.The model performance was evaluated through classification accuracy(CA),ROC curve and P-R curve.Next,the Res Net-50 deep neural network was used to extract the depth features of the tongue image,and the GA＿XGBT algorithm was used to fuse the color-texture features and the depth features,and finally a non-invasive diabetes risk prediction model was constructed entirely based on the tongue features.3.Construct a diagnosis model of TCM syndromes of diabetesThe tongue images of 339 subjects were collected using the TFDA-1 tongue diagnosis instrument and the objective tongue calculated by TDAS.A four-diagnosis scale survey was conducted on 158 subjects and information on symptoms and signs was extracted.The random forest(RF)model was trained using the color and texture features of 339 tongue images,and the objective tongue was transformed into probability features.The mutual information(MI)algorithm was used to transform the subjects’ symptoms and signs into MI features.Finally,the logistic regression classifier was used to fuse the probability feature and the MI feature and calculate the final syndrome type of the subject.Result1.Tongue features that are highly correlated with diabetes riskIn the pre-diabetic group,TB-ASM,TC-ASM and TB-Cr are protective factors.TBCON,TC-CON,TB-MEAN,TC-MEAN,TB-ENT and TC-ENT are risk factors.In the diabetes group,TB-a,TB-S,TB-Cr,TB-b and TC-b are protective factors,and Per-all and TB-Cb are risk factors.2.Construct a diabetes risk prediction model based on tongue featuresThe best fasting blood glucose(FBG)prediction model was the RF model.In the Clark error grid analysis(EGA),89.68% of the points fell in the A area,9.63% of the points fell in the B area,MSE was 0.601,and R-squared was 0.606.The best prediction model for glycosylated hemoglobin was the GBDT model.In the scatter plot analysis,the correlation between the predicted value and the true value was 0.59(P<0.001),the MSE was 0.272,and the R-squared was 0.539.When testing the STACK model on the test set of dataset 1,the non-invasive STACK model performed well,with a micro-average AUROC 0.866,a macro-average AUROC0.841,and a micro-average AUPRC 0.773.When verifying three deep learning models on the validation set of dataset 2,the Res Net-50 model performed best,with a microaverage AUROC 0.839,a macro-average AUROC 0.827,and a micro-average AUPRC0.727.When testing three deep learning models on the test set of dataset 2,the Res Net-50 model performed best,with a micro-average AUROC 0.832,a macro-average AUROC0.817,and a micro-average AUPRC 0.710.The cross-validation CA of the model based on tongue features fusion reached 82.1%,and the AUROC reached 0.924.The CA of the model on the test set is 81%,and the AUROC is 0.918.When the model detects pre-diabetes on the test set,AUROC is 0.914,Precision is 0.69,Recall is 0.952,and F1-score is 0.8.When detecting diabetes,AUROC was 0.984,Precision was 0.929,Recall was 0.951,and F1-score was 0.94.3.Construction of TCM syndrome diagnosis model of diabetesIn the 5-fold cross-validation,the optimal performance of the TCM syndrome diagnosis model was CA 74.2%,Micro AUROC 0.879,Macro AUROC 0.824,AUROC0.69 for the diagnosis of phlegm-heat interaction,AUROC 0.879 for the diagnosis of heatexcessive and invigorating body fluid,and AUROC 0.833 for the diagnosis of deficiency of both qi and yin.The logistic regression model fused with prior knowledge and objective tongue achieved the best performance,with an average CA of 65.2%.Conclusion1.The first change in the pre-diabetes period is the texture of the tongue,which suggests that the body fluid is the first to be lost.In the diabetic period,the color features and the proportion of the tongue coating begin to change,suggesting that the disease is progressing toward deficiency syndrome,phlegm syndrome,and blood stasis syndrome.2.The diabetes risk prediction model constructed based on the objective tongue has achieved excellent performance,showing the unique advantages and important effects of the objective tongue applied to diabetes risk prediction,which further proves the scientific nature of the tongue diagnosis.3.Combine the prior knowledge and the objective tongue to establish a TCM syndrome diagnosis model for diabetes.Prior knowledge can avoid the blindness of machine learning,and the objective tongue can correct the subjectivity of the prior knowledge.The successful construction of a TCM syndrome diagnosis model proved the clinical practical value of objective tongue.