Establishment Of TCM Artificial Intelligence Syndrome Differentiation For Depression Based On Various Physiological Information,Symptoms And Signs

Background and Objective:Depression,also known as depression disorder,refers to a kind of mood disorder caused by various reasons,characterized by persistent and significant low mood,often accompanied by anxiety,somatization symptoms,and decreased thinking and cognitive function.In traditional Chinese medicine,depression is classified as "depression disease/syndrome",which can be divided into liver qi stagnation,liver qi deficiency and spleen deficiency,kidney deficiency and liver stagnation,heart and spleen deficiency,and qi stagnation and fire syndrome.In terms of TCM syndrome differentiation of depression,the main methods of TCM syndrome classification are TCM syndrome differentiation by clinicians and diagnosis scale TCM syndrome classification.However,in general,the current syndrome classification of depression cannot be separated from the subjective judgment of traditional Chinese medicine doctors.In order to objectify the information of patients with depression from multiple aspects such as tongue diagnosis,pulse diagnosis,emotion recognition and medical record data,so as to assist in syndrome differentiation of depression syndrome in patients with depression,we designed this study.This study analyzed patient data from four aspects:tongue diagnosis,pulse diagnosis,emotion recognition and medical record data.And deep learning and machine learning algorithms are used for dialectical identification of the data.At the same time,it provided some reference for the establishment and optimization of TCM syndrome differentiation system based on artificial intelligence for other diseases.Methods:1.Establishment and verification of TCM artificial intelligence tongue diagnosis scheme for depressionPython web crawler technology was used to crawl the data related to tongue image on the Internet.We selected pictures that contained a complete and clear tongue image.The YOLO V3 target recognition algorithm was used to identify the target of the tongue image,and then intercepted the target zone to obtain the tongue image database.The color correction algorithm was used to correct the color of the target image.The target tongue image data set was artificially classified.The tongue images obtained were divided into light red tongue,light white tongue,red purple tongue,white tongue fur,yellow tongue fur,thick tongue fur and thin tongue fur by two TCM doctors.Next,the deep learning algorithm based on Keras framework was used to carry out deep learning on the obtained data set,and the accuracy and recall rates of the learning results were obtained After confirming that the results were good,the tongue images of patients with depression were selected,and the Keras deep learning model was used to judge the kind of tongue color and tongue coating,and the accuracy and recall rates of the final results were calculated by compared with the correct judgment results.2.Establishment and verification of TCM artificial intelligence pulse diagnosis scheme for depressionUsing the photoelectric pulse wave principle,the pulse wave of the finger tip and the pulse wave of the radial artery were collected.Fingertip pulse wave was collected for about 20s-2min each time,and radial pulse wave was collected for about 3min each time,to ensure relatively stable pulse wave data can be obtained.Pulse signs were identified by two TCM physicians.Keras deep learning algorithm was used to carry out deep learning on the obtained data set,and the accuracy and recall rates of the learning results were obtained.If the results are good,further verification can be carried out on patients with depression.3.Establishment and verification of TCM constitution and TCM syndrome differentiation scheme for depression based on deep learning emotion recognition algorithmHeart waves and pulse waves of patients with mild to moderate depression and healthy people were selected and collected.The subjects were selected to collect and TCM constitution score,TCM syndrome differentiation of depression score were calculated.Patients were then given a 3 0-minute emotion-arousing video,which included seven video clips that elicited 7 emotions:calm,happiness,sadness,fear,anger,surprise,and disgust.The data of ECG and pulse waves obtained during emotional stimulation were recorded by software.The deep learning algorithm was used to identify the emotion of patients,and the deep learning algorithm was used to identify the depression patients and healthy people respectively,establishing the relationship between the emotion of patients and TCM syndrome differentiation,and try to use the emotion recognition method with deep learning algorithm to carry out TCM syndrome differentiation and TCM constitution recognition on patients.4.TCM syndrome differentiation of patients with depression according to the medical records systemPatients with depression disorder in Shenzhen Hospital of Beijing University of Chinese Medicine(Longgang)since January 1,2016 and their medical records were included.Information was extracted from medical records,and the extracted information was corresponding to the corresponding syndrome type.Logistic regression model of TCM syndrome differentiation for depression was established to evaluate the advantages and disadvantages of the model.Machine learning algorithms also were used to differentiate TCM syndrome according to the extracted medical record information(such as the symptoms of patients).ResultsIn terms of artificial intelligence tongue diagnosis and recognition,a total of 2515 relevant tongue images were obtained by web crawler.After screening,a total of 1427 images with target areas were obtained.YOLO v3 was used for target recognition,and 266 labeling results were learned.The 266 labeled results were divided into training set and verification set according to the ratio of 1:1.After two batches of 73 rounds of training,the model was obtained.The accuracy,precise and recall rates of the model were 96.10%,96.61%and 98.85%respectively.Deep learning recognition of tongue color,coating thickness and coating color were carried out.Keras deep learning model combined with K-folds cross test was mainly used for recognition.The accuracy of tongue color recognition was about 75-82%,tongue coating color recognition was about 82-95%,and tongue coating thickness recognition was about 87-92%.In the aspect of artificial intelligence pulse recognition,we took the photoelectric volumetric pulse wave as the research object,and collected the photoelectric volumetric finger pulse wave and the photoelectric volumetric radial artery pulse wave respectively by using the wearable and the watch-type photoelectric volumetric pulse wave device,and then conducted machine learning and deep learning analysis.Because the photoelectric volume fingertip pulse wave collection location was far from the pulse location of the radial artery,and the wrist watch wearable pulse wave collection had certain position bias and tightness difference,the learning results were poor.In total,fingertip pulse waves were collected from 162 subjects and radial pulse waves from 43 subjects.Among them,when collecting radial artery pulse waves,we can see that there were abnormal conditions such as location bias,tightness difference,data overflow and so on,so the data collection was terminated in advance.When it comes to identifying fingertip pulse waves,the accuracy of the machine learning algorithm was between 31%and 40%,and the accuracy of the deep learning algorithm was around 44%.For radial pulse wave,the highest accuracy of machine learning algorithm was around 64%,and the accuracy of deep learning was around 22%.Taking a typical pulse wave for analysis,the highest accuracy of the machine learning algorithm was about 53%.Regarding the establishment and verification of TCM constitution of depression and TCM syndrome differentiation scheme based on deep learning emotion recognition algorithm,we firstly verified the emotion recognition algorithm based on heart waves using the previous two data sets(Ascertain data set and Dreamer data set).For the Ascertain data set,the accuracy of deep learning of four kinds of emotions recognition was about 30%.For the Dreamer data set,the accuracy of nine kinds of emotions recognition was around 16%.In terms of the data collected by our team,a total of 75 subjects participated in the collection of emotional arousal cardiac and pulse waves.There were 36 participants in experimental groups and 39 participants in control groups.In terms of baseline data,there were 36 males and 39 females,with a mean age of 44.08±12.19 years.There were significant difference in Chinese medicine constitution scores between the control group and the experimental group.Among them,the scores of yang-deficiency constitution,yin-deficiency constitution,qi-deficiency constitution,phlegm-dampness constitution,damp-heat constitution,blood-stasis constitution,special benefit constitution and qi-stagnation constitution in the test group were significantly higher than those in the control group,and the scores of peaceful constitution were significantly lower than those in the control group.Among the TCM syndrome differentiation scores of depression patients in the experimental group,liver stagnation and spleen deficiency were the highest,followed by the syndrome types of heart and liver fire,liver and kidney Yin deficiency,phlegm and turbidity,and the score of qi stagnation and blood stasis was the lowest.In terms of depression and anxiety scale scores,compared with the control group,the total score of HAMD along with seven factors of HAMD in the experimental group were significantly higher than those in the control group,and the total score of HAMA along with two factors of HAMA in the experimental group were significantly higher than those in the control group.Pearson correlation analysis showed that the absolute value of correlation coefficients between different constitutions and different emotional arousal scores was basically below 0.3.There was no absolutely obvious correlation between different constitutions and emotional arousal.The absolute value of the correlation coefficient between different syndrome types and different emotional arousal scores in patients with depression was also basically below 0.3.There was no absolutely obvious correlation between different TCM syndrome types and emotional arousal in patients with depression.In terms of machine learning,emotional heart waves such as calm,joy,sadness,fear,anger,surprise and disgust were used to distinguish the experimental group and the control group(depressed patients and normal subjects),with the highest accuracy of 70%,72%,73%,77%,82%,69%and 75%,respectively.The average accuracy of each algorithm was compared,and the average accuracy of Gaussian Bayesian algorithm was the highest,reaching 73%.Gaussian Bayesian algorithm was used to identify the whole segment of cardiac waves and subjects,and the accuracy on ECG long segments was about 70%,and the accuracy of distinguishing depressed patients and normal subjects was 73.6%.The emotion recognition accuracies of the different algorithms were up to 30%by using the seven-divided method.Since the accuracy of machine learning to distinguish patients with depression,the accuracy of emotion recognition,and the absolute value of correlation coefficient between depression scale scores,Chinese constitution scores and syndrome differentiation scores were not very high,we believed that the original hypothesis was not valid.Therefore,the data set was not used to identify the syndrome differentiation and constitution classification with machine learning algorithm.In the identification of TCM syndrome differentiation of patients with depression based on the selection of patients’ symptom characteristics in the medical record system,a total of 567 cases and 1128 medical records of treatment in Shenzhen Hospital of Beijing University of Chinese Medicine(Longgang)since January 1,2016 were included.Among them,280 were male and 287 were female.Among all the patients with depression,the deficiency of heart and spleen accounted for the largest proportion(16.98%),followed by kidney deficiency and blood stasis(9.57%),liver depression and spleen deficiency(4.94%).In terms of symptoms,poor mood,anxiety,fatigue,difficulty in sleeping,dreaminess,dizziness,backache,upset,palpitation are the most core symptoms of patients with depression.Through PCA analysis of the symptoms,we can better distinguish the different syndrome types such as deficiency of heart and spleen,deficiency of kidney and blood stasis,stagnation of liver and deficiency of spleen.Logistic regression was used to establish the prediction model,and the degree of differentiation of the prediction model was evaluated by the area under the ROC curve(C index).The results showed that the C index of the heart and spleen deficiency syndrome type was 0.77,the C index of the kidney deficiency and blood stasis syndrome type was 0.89,and the C index of the liver stagnation and spleen deficiency type was 0.85.The model had a good degree of differentiation.Machine learning algorithm was used to predict TCM syndromes according to symptoms,and five-classification machine learning algorithm was used to predict TCM syndromes with an accuracy rate of 72%.Machine learning algorithm was used to predict TCM syndrome elements.Using binary classification machine learning algorithm,the accuracy of syndrome element of liver stagnation was 86%,syndrome element of spleen deficiency was 86%,syndrome element of blood stasis was 78%,syndrome element of kidney deficiency was 88%,and syndrome element of qi deficiency was 97%.Conclusion1.Tongue diagnosis of depression and TCM syndrome differentiation and recognition based on the symptoms of patients in the medical record system achieved good results in our study,while pulse diagnosis of depression,syndrome differentiation and TCM constitution analysis based on emotion recognition did not achieve good results in our study.On the whole,the study of tongue diagnosis of depression and the TCM syndrome differentiation recognition based on the symptoms of patients in the medical record system do not have high requirements for tongue photos and case system data,which are easy to be realized and have good significance for in-depth research and extensive promotion.However,researches on pulse diagnosis of depression(especially those on wearable wrist devices),emotion recognition based on physiological signals,and then TCM syndrome differentiation and TCM constitution identification have high requirements on data collection equipment,and the results were not good,so there were still some problems to be confirmed,improved and solved.2.Using the algorithm of YOLO v3 to identify the tongue diagnosis pictures and intercepting the target recognition results for deep learning recognition,could better identify the color of tongue,color of tongue coating and thickness of tongue coating.The recognition model was applied to the tongue images of patients with depression,and the results were good.The model was easy to operate and had low requirements for photo collection settings and tongue-pose.3.Using machine learning algorithms to identify pulse based on fingertip pulse wave or radial pulse wave,the results were poor.The reasons may be related to the low correlation between fingertip pulse wave and Chinese pulse,the radial artery pulse wave data collected by the wristwatch wearable acquisition device was easily affected by the acquisition location,data overflow and other conditions.There was a strong correlation between artificial intelligence pulse diagnosis and acquisition equipment for patients with depression,which requires higher requirements for acquisition equipment.4.Machine learning algorithms,pulse waves and heart waves when watching emotional stimulation videos can be used to identify depression patients and normal people to a certain extent,and can also identify different emotions of subjects to a certain extent.However,according to the current equipment conditions,the accuracy of identify the depression patients,the accuracy of emotion recognition,it was difficult for our research team to complete TCM syndrome differentiation and TCM constitution identification for patients with depression.For the reason of the results,the identification results were also strongly correlated with the data collection results and emotional arousal degree and there were higher requirements for data collection equipment and collection environment.At the same time,this study failed to mark the specific time of emotional excitation on ECG and pulse wave fragments,resulting in the inability to extract the specific forms of heart wave and pulse wave during various emotional excitations.How to capture the precise ECG and pulse wave fragments during emotional stimulation is also worthy of further research.5.In the identification of TCM syndrome differentiation of patients with depression based on the selection of symptoms and characteristics of patients in the medical record system,good results can be obtained by using a prediction model established by Logistic regression and machine learning algorithm to predict TCM syndrome types and syndrome elements.The data quality of the medical record system selected in this study was relatively general and conformed to the general situation in the real world.Nevertheless,the study still obtained good results,and we believed that the method of establishing the model has good expansion and is worthy of further study and verification in a larger sample size of medical record system data.