Research On Diagnosis Of Coronary Heart Disease Using Traditional Chinese And Western Medicine Assisted By Machine Learning Based On Tongue Feature

Background:Coronary heart disease(CHD)is the first cause of death globally.As a common disease,CHD has become a major social and public health problem and has significant research value.Early screening and diagnosis of CHD can prevent more harm.Accurate traditional Chinese medicine(TCM)syndrome differentiation for patients with CHD is a critical way to improve the effectiveness of integrated Chinese and Western medicine in the treatment of CHD.Currently,coronary angiography is the gold standard for the diagnosis of CHD,but this method has obvious shortcomings,such as invasiveness,high cost,and technical severe difficulty.While the combination of Chinese and Western medicine in the treatment of CHD has clear clinical efficacy advantages,the clinical application of Traditional Chinese Medicine is limited due to the difficulty in providing objective and clear diagnostic evidence in diagnosis.Therefore,finding a method suitable for early screening and diagnosis of large populations in the diagnosis of CHD is a critical problem to solve.To clarify the quantitative evidence of TCM diagnosis and realize accurate diagnosis based on syndrome differentiation is an important premise to better play the advantages of integrated TCM and Western medicine.Tongue is a terminal organ with abundant blood flow that contains a large amount of physiological and pathological information about the human body.Tongue examination is an essential part of TCM diagnosis.In recent years,tongue diagnosis in TCM has been developing towards objectification,standardization,and quantification,but an effective tongue-disease-syndrome relationship system based on quantitative features has yet to be established.Whether tongue imaging can be used as an essential basis for non-invasive diagnosis of CHD,whether tongue images can be used for automated syndrome differentiation of CHD and applied to clinical practice,and how to better explore and utilize tongue image information are still under further exploration.Objective:(1)Based on the tongue image features of coronary heart disease patients,analyze the quantitative features of tongue appearance for different types of syndromes,establish an effective correspondence for tongue-disease-syndrome.(2)Based on the objective features extracted from the tongue image analysis system,we extract more features from the tongue image at the same time,so as to make full use of the diagnostic function of the appearance of the tongue.(3)Establish a CHD diagnosis and syndrome differentiation model based on tongue image features,explore the value of tongue features for non-invasive diagnosis of CHD.We explored the feasibility of applying tongue in clinical assisted diagnosis,and provided methods and technical support for clinical transformation of objective study results of tongue imaging.Methods:(1)Objective tongue features of coronary heart diseaseWe recruited 315 patients with CHD and 211 healthy people as participants and collected their clinical data and tongue images.We used tongue image analysis system(TDAS 3.0)to analyze tongue images and obtained 27 features including color,texture and tongue coating area.Tongue features between patients with CHD and healthy people and between different syndrome types of CHD were compared by SPSS 23.0.(2)Noninvasive diagnosis and syndrome types differentiation of CHD based on tongue image.To fully utilize the information of tongue images and improve the classification performance of the model,my research extracted TDAS features and deep features from tongue images,and fused the two features.To accurately extract the deep features of tongue images,the PFLD model was first used for facial keypoint detection,followed by specific key points for tongue image rectangular area positioning and cropping,and the U-Net model for tongue segmentation,which constructed a tongue image database containing only tongue information.To quickly adapt to the segmentation task of this dataset,the transfer learning was used to retrain the pre-trained model on the PASCAL VOC dataset.In extracting deep features,the pre-trained ResNet-50 model on the ImageNet dataset was used and further retrained on our dataset.Seven machine learning models,including decision trees(DT),random forests(RF),K-nearest neighbors(KNN),logistic regression(LR),support vector machines(SVM),artificial neural networks(ANN),and XGBoost,were used to construct classifiers for the diagnosis and differential diagnosis of coronary heart disease.The parameter optimization of the classification model was carried out using the random search strategy,and the performance of the model was evaluated using metrics such as accuracy,precision,and recall.Finally,TDAS features,deep features and fusion features were used as inputs to compare the performance of different models,and the model with the best performance was selected to construct the final non-invasive diagnosis and differential diagnosis models of CHD.Results:(1)Objective tongue features of CHDThe tongue objective features are significant differences in patients with CHD and healthy individuals.The analysis results show that for tongue feature indicators,the CHD group was lower than the healthy group in color indicators TB-R,TB-G,TB-B,and TB-I(P<0.05).In comparison,the CHD group was significantly higher than the healthy group in color indicators TB-H,TB-L,and TB-b(P<0.05).For tongue coating indicators,the CHD group was significantly lower than the healthy control group in color indicators TC-R and TC-B(P<0.05),while the CHD group was considerably lower than the healthy control group in color indicators TC-H and TC-b(P<0.05),and significantly higher than the healthy control group in texture indicators TC-CON and TC-MEAN(P<0.05).These findings suggest that compared with the healthy population,CHD patients have lower tongue brightness,more reddish-purple and darker tongue color,darker tongue coating color,and coarser tongue coating texture.Significant differences exist of tongue features in CHD with different syndrome types,and each type has its personalized expression.Syndrome of stagnation in heart channel is characterized by a red and purplish tongue,a slightly tender tongue body,and a thin and greasy tongue coating with a darker color.Syndrome of Qi stagnation and blood stasis is characterized by a little red and purplish dark tongue,a slightly tender tongue body,and a yellow and thin greasy tongue coating.The phlegm stasis syndrome is characterized by a tongue color that tends to be light red,an roughed tongue texture,and a thick white and roughed tongue coating.Heart pain with syndrome of yang deficiency and cold congelation is characterized by a tongue color that is dark red and purplish and a thin and roughed tongue coating.The Qi deficiency and Blood stasis syndrome are relatively in the middle among the nine groups regarding indicator values,with tongue manifestations similar to those of the healthy population.The Qi and Yin deficiency syndrome is characterized by the lightest tongue color,a slightly tender tongue body,and a thick greasy white tongue coating.Yin deficiency of heart and kidney syndrome have a dark red and purplish tongue,an roughed tongue body,and a thin yellow and roughed tongue coating.Yang deficiency of heart and kidney syndrome is characterized by a slightly lighter red tongue color,an roughed tongue body,and a thick white tongue coating.(2)Noninvasive diagnosis of CHD based on tongue image featuresOverall,common machine learning algorithms can be used to construct non-invasive diagnostic models for CHD based on tongue features.Except for LR,the accuracy of these algorithms is above 0.7,and the XGBoost algorithm performs the best with a diagnostic model accuracy of over 0.8.In comparison,models relying solely on TDAS features for CHD diagnosis have relatively lower accuracy,while diagnostic models based on fused features have the best performance.When sorting the input features by importance,the top features are TB-L,PC 9,PC 215,TB-Lb,PC 15,PC 128,TB-H,PC 11,and PC 83.The final non-invasive diagnostic model has an accuracy of 0.812,an AUC of 0.893,a precision of 0.818,a recall of 0.803,and an Fl-score of 0.810.(3)Study on syndrome differentiation of CHD based on tongue image featuresOn the whole,among common machine learning algorithms,XGBoost has the best performance in the construction of CHD syndrome differentiation model.When sorting the importance of input features,the top features are PC 61,PC 9,TB-S,PC 15,PC 34,PC 24,PC 86,PC 78,and TB-H.Top 1 is defined as only looking at the highest probability diagnosis result determined by the model,that is,the first diagnosis result is the actual diagnosis result.Top 2 is defined as looking at the results of the top two diagnosis results with the highest probabilities,that is,the top two diagnosis results contain the actual diagnosis result.The syndrome differentiation model constructed in this study performs well,witih accuracy is 0.472 for Top 1,AUC is 0.830,BACC is 0.443,and a diagnostic accuracy is 0.817 for Top 2,AUC is 0.834,and BACC is 0.808.Conclusion:(1)Tongue image indicators can sensitively reflect the tongue’s characteristics of CHD and its different syndrome types,and these features have good interpretability,indicating that objective and quantitative tongue features have good medicinal value.By studying objective tongue features of CHD,an effective correspondence between tongue-syndrome-disease can be established,which is consistent with clinical practice.This further elucidates the value of tongue in the diagnosis and syndrome differentiation of CHD,and the research results are expected to be applied to guide clinical practice.(2)A diagnostic model for CHD with good performance can be constructed based on tongue image features;tongue image features are practical information for diagnosing CHD;non-invasive diagnosis of CHD based on tongue features is feasible;tongue images can also be used for clinical auxiliary syndrome differentiation;machine learning is an excellent technical means to transform the research results of the objective tongue features into clinical applications.(3)The quantitative results of tongue features obtained by tongue feature extraction systems such as TDAS have good interpretability,but their ability to mine tongue information is still insufficient to meet diagnostic requirements in practical diagnosis problems.Deep learning frameworks perform well in mining hidden tongue features,and the fusion of deep features can significantly improve the performance of CHD diagnosis models.It can be seen that the TDAS tongue features based on TCM tongue diagnosis prior knowledge ignore the adequate information of tongue image diagnosis for CHD diagnosis.Therefore,only using TDAS to obtain tongue features is relatively insuficient,and deep features can serve as an excellent complement to TDAS to optimize tongue information mining and diagnosis models.