Deep Learning-based Study On Facial Features Of Anxiety States And Correlation With The Five States Of Personality In Traditional Chinese Medicine

Objective: The aim of this study is to learn facial features from the perspective of facial diagnosis in Traditional Chinese Medicine,using deep learning,which is a rapidly developing technology in the field of computers,in order to objectify the facial diagnosis of anxiety states and to assist in the diagnosis of anxiety states.In addition,anxiety states are closely related to personality,and there is also an intrinsic link between personality and facial features.Therefore,the second part of this study will explore the correlation between anxiety states and the five states of personality in Traditional Chinese Medicine,to explore the front-end factors in the disease process of anxiety states,to understand the deeper internal logic of anxiety states.It also provides a basis for further refinement of the facial features of patients with different personality states,which will be of great significance in the clinical treatment of the disease.Methods:This study consists of two parts,the first part of the enrollment was from patients who attended the outpatient and ward of the Second Department of Brain Diseases at the Affiliated Hospital of Shandong University of Traditional Chinese Medicine from April 2022 to December 2022 and the healthy subjects recruited.Xception,a deep learning-lightweight network in the field of computer vision,was used to study the anxiety states facial feature classification model.The specific methods were as follows.1.Creation of a database The acquisition device is camera + lens(the camera is BOSMA G1 camera(resolution: 7680×4320(8k UHD)),the lens is OLYMPUS M.ZUIKO DIGITAL ED40-150 mm F2.8PRO telephoto zoom lens),and the facial images of two groups of subjects in anxiety states and healthy subjects are collected.First,the data are cleaned to eliminate unqualified pictures;A further pre-processing is operated,including scale normalization and data enhancement,where scale normalization consists of scaling the image to 299 x 299 and normalizing the values to be in the range [0,1] by using tf.keras.layers.Rescaling;The data enhancement adopts the method of flip conversion to establish the facial image database.2.Construction of anxiety states facial feature classification model based on the lightweight network Xception networkA facial feature classification model of anxiety states is studied by using lightweight network Xception.Firstly,the basic model is created from the pre-trained convolutional network,and the Xception network is used to create the basic model,which can be divided into the following steps:(1)each 299×299×3 image is converted into 10×10×2048 feature blocks through the feature extraction program;(2)Freezing the convolutional basis:By setting layer.trainable = False,freezing the convolutional basis can be achieved to avoid updating the trained weights of the original model during training.There are many layers in Xception,so setting trainable to False freezes all layers of the model.(3)adding classifier:The features are first averaged over 10 x10 spatial locations by employing the tf.keras.layers.Global Average Pooling2 D layer and then converted into one vector per image(containing 2048elements).The tf.keras.layers.Dense layer was applied to convert these features into one prediction per image.In order to avoid the phenomenon of overfitting during the training process of the network,the Dropout layer is added to the fully connected layer of the network.The selection range of Dropout value is(0,1).The Dropout layer of the model proposed in this study takes the value of 0.2.Next compile the model to use tf.keras.losses.Binary Crossentropy loss in conjunction with from＿logits=True.Finally,the model training and classification test were carried out,and the model training and test were carried out again after fine tuning.The subjects of the second part of this study are patients with anxiety states who meet the inclusion criteria of the first part of this study.The basic information of the study participants and the scores of the Five States of Personality Test were created into a database,which was statistically processed using SPSS 23.0 software.Results:1.A total of 362 subjects were included in the first part of this study,including 182 anxiety states subjects in group A and 180 healthy subjects in group B.After eliminating unqualified pictures,there are 2,718 pictures in group A and 2,691 pictures in group B.2.The training set and test set are randomly generated according to a ratio of 7:3.At this time,the total number of trained pictures is 3,786,and the total number of test pictures is1,623;Model training and test data using lightweight network Xception.In the model training,A total of 3,786 images were selected from the two groups of data A and B for training.0 and 1 were used as labels respectively to represent the category of images.The basic parameters of the model were set as follows:set epochs as 80 times(epoch:An epoch is equal to the use of training set of all samples training once),set batchsize as 32(batchsize:number of samples taken in the training set per training session),set the learning rate as 0.001(learning rate:monitoring,and guiding model to adjust the weight of the network);On the basis of the pre-training model,the parameters of the classifier are fine-tuned,and the final training model is obtained.The test results showed that 695 people actually belonged to Group A were predicted to be Group A,60 people actually belonged to Group A were wrongly predicted to be Group B,105 people actually belonged to Group B were wrongly predicted to be group A,and 763 people actually belonged to Group B were predicted to be Group B.The final accuracy was 88.19%,and the AUC value was 0.883.In the anxiety states group,the recognition accuracy was 87%,the recall rate was92%,and the F value was 0.89.In the healthy subjects group,the recognition accuracy was 93%,the recall rate was 88% and the F value was 0.90.2.In the second part of this study,there were 182 subjects,11 invalid questionnaires were excluded,and 171 valid questionnaires remained.(1)Compared with the norm,the five states of personality scores of patients with anxiety states were lower on the Yinyang peace dimension and higher on the Shaoyin and Taiyin dimension,with statistical significance(P<0.05).(2)Compared with the norm,male patients had higher scores in Shaoyin dimension;Compared with the norm,the scores of Yinyang peace dimension were lower in female patients,but the scores of Shaoyin and Taiyin dimension were higher.Compared with female patients,male patients had higher scores in the Taiyang and Shaoyin dimensions,and lower scores in the Taiyin dimension.(3)There was no significant difference in the scores of Taiyang,Shaoyang,Yinyang peace,Shaoyin and Taiyin in anxiety states among different ages(P>0.05).(4)Compared with different levels of education,the scores of the Taiyang and Taiyin had statistical significance(P<0.05),while the scores of Shaoyang,Yinyang peace,and Shaoyin had no statistical significance(P>0.05).Conclusion:1.Anxiety states is not only an established disease outcome,but also an influencing factor,which is associated with and influenced by various clinical diseases.In Traditional Chinese Medicine,observating,smelling,asking and touching are mostly used for the diagnosis of anxiety states,which lacks objective indicators.Therefore,an interdisciplinary integration of Artificial Intelligence and facial visualization in Traditional Chinese Medicine was carried out to construct a classification model of facial features of anxiety states and explore the path of objectifying facial visualization for anxiety states.2.The deep learning-lightweight network Xception constructing classification model in the field of computer vision explores the application of Artificial Intelligence in the field of facial observation in Traditional Chinese Medicine,and lays a certain foundation for the study of facial observation in other diseases and other fields of research in Traditional Chinese Medicine.3.The occurrence and development of anxiety states has a certain correlation with Shaoyin and Taiyin personality,so targeted treatment and suggestions can be given in combination with the personality characteristics of patients in clinical diagnosis and treatment.In addition,personality is closely related to facial features,especially facial expressions.The study of the correlation between anxiety states and personality can lay a foundation for further refining the classification model of facial features in anxiety states.