CN114464325A - Intelligent dialectical treatment device for traditional Chinese medicine - Google Patents

Intelligent dialectical treatment device for traditional Chinese medicine Download PDF

Info

Publication number
CN114464325A
CN114464325A CN202111583091.2A CN202111583091A CN114464325A CN 114464325 A CN114464325 A CN 114464325A CN 202111583091 A CN202111583091 A CN 202111583091A CN 114464325 A CN114464325 A CN 114464325A
Authority
CN
China
Prior art keywords
data
module
cross
fgcnn
intelligent
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111583091.2A
Other languages
Chinese (zh)
Inventor
温川飙
黄宗海
陈菊
宋海贝
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chengdu University of Traditional Chinese Medicine
Original Assignee
Chengdu University of Traditional Chinese Medicine
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chengdu University of Traditional Chinese Medicine filed Critical Chengdu University of Traditional Chinese Medicine
Priority to CN202111583091.2A priority Critical patent/CN114464325A/en
Publication of CN114464325A publication Critical patent/CN114464325A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H20/00—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance
    • G16H20/10—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance relating to drugs or medications, e.g. for ensuring correct administration to patients

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • General Physics & Mathematics (AREA)
  • Public Health (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • Biomedical Technology (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Chemical & Material Sciences (AREA)
  • Medicinal Chemistry (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to the technical field of informatization of traditional Chinese medicine syndrome differentiation, in particular to an intelligent syndrome differentiation and treatment device of traditional Chinese medicine. The apparatus performs steps comprising: s1, acquiring symptom characteristic data; s2, inputting the symptom characteristic data into a Cross-FGCNN intelligent dialectical model, and outputting a syndrome type corresponding to the symptom characteristic data; the Cross-FGCNN intelligent dialectical model comprises a data embedding module, a data feature crossing module, an FGCNN module and a classification module, wherein the data embedding module is used for mapping symptom feature data into low-dimensional dense feature data and outputting the low-dimensional dense feature data to the data feature crossing module and the FGCNN module, and the data feature crossing module is used for extracting multi-order linear crossing features in the low-dimensional dense feature data; the FGCNN module is used for extracting nonlinear cross features in the low-dimensional dense feature data; the classification module is used for merging the multi-order linear cross features and the nonlinear cross features and outputting syndrome types corresponding to the symptom feature data. The accuracy of intelligent dialectical treatment is improved.

Description

Intelligent dialectical treatment device for traditional Chinese medicine
Technical Field
The invention relates to the technical field of informatization of traditional Chinese medicine syndrome differentiation, in particular to an intelligent dialectical treatment device of traditional Chinese medicine.
Background
Dialectics is one of the important classification tasks for TCM treatment, and is the key to TCM treatment of diseases. The symptoms can be generally obtained from inspection, auscultation, inquiry and palpation. The four diagnostic methods are the key to the syndrome differentiation of traditional Chinese medicine. Therefore, from the machine learning point of view, the TCM dialectics can be regarded as a complex function model with the input of all four diagnostic methods of the patient and the output of the syndrome of the patient. With the advancement of machine learning techniques, many researchers have worked to fit this model. Case association analysis was performed using the DFP-growth algorithm to explore the association between symptoms and syndrome types. However, the output result of the algorithm changes according to the threshold value, and is difficult to converge, and the time complexity and the space complexity are high. Wang Bo et al, improves the original DFP-growth algorithm, introduces a hypergraph algorithm with different similarities and a DFP-growth algorithm with attribute combination, reduces the complexity of the algorithm to a certain extent, but still has large cost. Yanghui Gu et al and WANG Yan et al use decision tree models for intelligent forensics of kidney and liver cirrhosis, both models can only be used in the data set architecture mode of the constructed model, and migration capabilities are poor. Wenjie Xu et al and Yufeng Zhao et al describe the dialectical process of traditional chinese medicine intuitively through the improved bayesian network, but this method highly depends on the integrity and validity of the original training data and is not very friendly to the medical field with great difficulty in acquiring the complete data. Tao Yang et al uses fuzzy recognition and a five-organ dialectic system to construct a dialectic model, but the method has certain difficulties in the determination of a fuzzy set, the measurement of similarity between a fuzzy matrix and the fuzzy set and the like. The use of ShuJie Xia et al and Yong-Zhi Li et al found that ML-KNN had good results in the search of the dialectical model of TCM, but this model was extremely easy to classify all data into the class with the most amount of data in case of data imbalance. The BP neural network was optimized by using the third order convergent Levenberg-Marquardt algorithm but this method still did not solve the accuracy degradation problem due to data imbalance. QIANG XU et al, used a 10-layer artificial neural network, which was somewhat overfitting due to the number of network layers compared to the data. With the progress of research, in order to obtain better fitting degree, researchers further promote the training of dialectical models from the aspect of training data preprocessing. Wang Yiqin et al, the four-diagnosis information is input into a neural network and a support vector machine according to five dimensions of tongue, face, voice, inquiry and pulse diagnosis, but the input data of the method is complete patient data, so the classification effect of missing value data still needs to be deeply researched. Five rules are formulated on the basis of the artificial neural network, so that the intelligent degree of the model is reduced to a certain extent.
In summary, although the existing model can better distinguish the corresponding data set, the model has high requirements on data. Patient information needs to be sufficiently complete, and from the real world perspective, there must be a large amount of missing data for patient information acquisition. Therefore, what is needed is a model that is closer to the real world and that can effectively resolve high-dimensional sparse data.
The input dimensionality in a Click Through Rate (CTR) task is large and sparse, and a Factorization Machine (FM) automatically performs feature engineering by obtaining the relation between two corresponding features through performing inner product on the weighted values of the two corresponding features. However, FM is limited to second-order cross multiplication in feature selection, and automatic selection of high-dimensional features still has difficulty. In order to automatically extract higher-order feature combinations, the deep FM classifies each bit feature of an input end into a domain on the basis of an original FM model, and a DNN is constructed in parallel to obtain high-order nonlinear features. The method can learn low-order and high-order features at the same time, but can learn two-dimensional and one high-dimensional feature only, and the coverage is not strong. Thus, the dcn (deep and Cross network) proposal can learn the Cross feature of multiple dimensions at the same time by using Cross layer instead of FM. The DNN part of deep fm and DCN is more concerned with the non-linear high dimensional features of global data generation, ignoring some local features. The Feature Generation informed neural network (FGCNN) extracts local features by using a convolutional neural network, and combines the advantages of the original multilayer perceptron in extracting global features, so that the high-dimensional Feature contained information of the model is richer.
Disclosure of Invention
The invention aims to solve the problem that the conventional CTR model succeeds in classifying two categories of high-dimensional sparse data, inspires that the invention constructs an improved multi-classification model, namely a Cross-FGCNN intelligent dialectical model, which is suitable for dialectical typing of high-dimensional sparse symptom data of traditional Chinese medicine, and provides an intelligent dialectical treatment device of traditional Chinese medicine.
In order to achieve the above purpose, the invention provides the following technical scheme:
an intelligent dialectical treatment device for traditional Chinese medicine, which executes the steps of:
s1, acquiring symptom characteristic data;
s2, inputting the symptom characteristic data into a Cross-FGCNN intelligent dialectical model, and outputting a syndrome type corresponding to the symptom characteristic data;
wherein the Cross-FGCNN intelligent dialectical model comprises a data embedding module, a data characteristic crossing module, an FGCNN module and a classification module,
the data embedding module is used for mapping the symptom characteristic data into low-dimensional dense characteristic data and outputting the low-dimensional dense characteristic data to the data characteristic crossing module and the FGCNN module,
the data feature cross module is used for extracting multi-order linear cross features in the low-dimensional dense feature data; the FGCNN module is used for extracting nonlinear cross features in the low-dimensional dense feature data;
and the classification module is used for merging the multi-order linear cross characteristics and the nonlinear cross characteristics and outputting the syndrome type corresponding to the symptom characteristic data.
As a preferred scheme of the present invention, the data feature crossing module is configured to extract multi-order linear crossing features from the low-dimensional dense feature data, and the specific implementation method includes the following steps:
x of the l layer outputl+1Inputting original data x0And output data x of the previous layerlBy the formula xl+1＝f(xl,wl,bl)+xlCalculated wherein f (x)l，wl，bl) By the formula
Figure BDA0003426779650000041
Alternative, blIndicates a deviation, wlRepresents a weight, x0Representing the initial input vector, xlThe output vector representing the upper layer is obtained by a plurality of calculations to obtain the cross feature vector C.
As a preferred scheme of the present invention, the FGCNN module is configured to extract a nonlinear cross feature in the low-dimensional dense feature data, and the specific implementation method includes the following steps:
s1, inputting n into the convolution layerfX k x 1 matrix E, where nfIs the number of fields, k is the size of the embedding vector;
s2, performing convolution operation on the matrix E and a convolution kernel with the size of h multiplied by 1 multiplied by m, and acquiring a corresponding convolution value by using tanh as an activation function to acquire a convolution matrix;
s3, performing pooling operation through the maximum pooling layer to obtain pooled matrix SiPooling matrix SiAnd continuously downloading the input of the next convolution layer and recombining the input to obtain a reconstruction matrix, wherein the calculation formula of the reconstruction matrix is as follows: ri＝tanh(SiW+B)
tanh is used as an activation function, B is a bias execution matrix, and W is a weight matrix;
and S4, transmitting the reconstruction matrix into a multilayer perceptron to obtain a nonlinear high-dimensional feature vector F of local and global cross features.
As a preferred embodiment of the present invention, the classification module is configured to combine the multi-order linear cross features and the nonlinear cross features, and output a syndrome type corresponding to the symptom feature data, and specifically includes:
s100, the classification module combines the nonlinear high-dimensional feature vector F output by the FGCNN module and the cross feature vector C output by the data feature cross module to obtain a vector I1＝(C,F)；
S200, converting the vector I1Inputting the (C, F) into a multilayer sensor for calculation, wherein the calculation formula is as follows: o isi＝rel(uIiWi+Bi) Wherein W isiRepresenting a weight matrix of the ith hidden layer; b isiA bias execution matrix representing the ith hidden layer;
s300, last hidden layer nhAnd outputting the syndrome.
As a preferred embodiment of the present invention, in step S300, the last hidden layer nhThe calculation formula of the output syndrome is as follows:
Figure BDA0003426779650000051
wherein,
Figure BDA0003426779650000052
denotes the n-thh+1 weight matrix of hidden layer;
Figure BDA0003426779650000053
denotes the n-thhA bias execution matrix of +1 hidden layers,
Figure BDA0003426779650000054
is a predicted value.
As a preferred scheme of the invention, the loss function of the Cross-FGCNN intelligent dialectical model is as follows:
Figure BDA0003426779650000055
where N is the total number of input data sets, y is a real value,
Figure BDA0003426779650000056
is a predicted value.
As a preferred embodiment of the present invention, the output syndrome includes: deficiency of both liver and kidney, cold accumulation and blood stasis, cold-dampness stagnation, liver depression and damp-heat, deficiency of both qi and blood, qi stagnation and blood stasis, kidney deficiency and blood stasis, damp-heat stagnation, and yang deficiency and internal cold.
As a preferred scheme of the invention, in the Cross-FGCNN intelligent dialectical model, a data characteristic Cross module selects 6 layers of Cross networks, a FGCNN module selects convolution kernels with the depths of 14, 16 and 18 and the width of 4 to carry out three layers of convolution operation, the depth of a multilayer perceptron is 3 layers, a classification module selects 3 layers of neural networks, and the number of neurons is 1024, 512 and 128 respectively.
As a preferable scheme of the invention, in the Cross-FGCNN intelligent dialectical model, dropout is 0.2, the learning rate is 0.001, and 1000 iterations are performed.
Compared with the prior art, the invention has the beneficial effects that:
the device normalizes the input of an intelligent dialectical treatment algorithm through the traditional Chinese medicine theory, and reduces the influence of high-dimensional sparsity on model classification through extracting and fusing linear and nonlinear cross features of input sparse symptom features. The combined feature vector after fusion is further classified by using a classification model, so that the effect of intelligent dialectical treatment is achieved, and the accuracy of the intelligent dialectical treatment is improved.
Description of the drawings:
FIG. 1 is a flowchart showing the steps executed by an intelligent dialectical treatment apparatus in accordance with embodiment 1;
FIG. 2 is an example of electronic medical record preprocessing in embodiment 1;
FIG. 3 is a structural diagram of the Cross-FGCNN intelligent dialectical model in example 1;
FIG. 4 is a diagram showing a data embedding block in example 1;
FIG. 5 is a schematic diagram of multi-level linear cross feature extraction in embodiment 1;
fig. 6 is an improved FGCNN network structure in embodiment 1 of the present invention;
FIG. 7 is a Cross-FGCNN model accuracy-iteration time scatter diagram in example 2 of the present invention;
FIG. 8 is a schematic diagram of Cross-FGCNN confusion matrix in example 1 of the present invention;
FIG. 9 is ROC curves of the new model and the conventional CTR model in example 2 of the present invention;
FIG. 10 is a ROC curve of the conventional intellectual faculty model in example 2 of the present invention.
Detailed Description
The present invention will be described in further detail with reference to test examples and specific embodiments. It should be understood that the scope of the above-described subject matter is not limited to the following examples, and any techniques implemented based on the disclosure of the present invention are within the scope of the present invention.
Example 1
The invention provides a traditional Chinese medicine intelligent dialectical treatment device, which executes the steps of which the flow chart is shown in figure 1 and comprises the following steps:
and S1, acquiring symptom characteristic data.
And S2, inputting the symptom characteristic data into a Cross-FGCNN intelligent dialectical model, and outputting a syndrome type corresponding to the symptom characteristic data.
The Cross-FGCNN intelligent dialectical model comprises a data embedding module, a data feature crossing module, an FGCNN module and a classification module, wherein the data embedding module is used for mapping the symptom feature data into low-dimensional dense feature data and outputting the low-dimensional dense feature data to the data feature crossing module and the FGCNN module, and the data feature crossing module is used for extracting multi-order linear crossing features in the low-dimensional dense feature data; the FGCNN module is used for extracting nonlinear cross features in the low-dimensional dense feature data; and the classification module is used for merging the multi-order linear cross characteristics and the nonlinear cross characteristics and outputting the syndrome type corresponding to the symptom characteristic data.
First, data preparation
4000 high-quality electronic medical records of dysmenorrhea are collected from a Sichuan traditional Chinese medicine big data management platform, and 1273 cases of dysmenorrhea are obtained through a Chinese literature database. Only symptoms, syndrome type, disease name and other data relevant to the study were retained. The physicians of professional Chinese medicine standardize the syndrome type and symptom dimension according to the basic theory term of Chinese medicine of GB/T20348-2006 and GB/T16751.2-1997. Syndrome type and symptom cleaning are carried out through ' GB/T20348-2006 Chinese medicine basic theory term ' and ' GB/T16751.2-1997 Chinese medicine clinical diagnosis and treatment term ' syndrome part ' issued by the Chinese national traditional medicine administration so as to achieve dimensional standardization unification. All the symptoms related to the dialectical reasons of TCM are classified into four categories, namely inspection, smell, inquiry and cutting according to the subclass of diagnosis in four diagnostic methods in TCM diagnostics. Wherein the inspection includes look; skin color; a body; a gesture; a head; a face; a nose; an eye; an ear; a mouth; teeth; a neck portion; a chest; the abdomen; lumbar vertebrae; external genitalia; an anus; skin; sputum; saliva; vomit; defecating; urinate; the superficial index finger vein; the tongue quality; the tongue color; tongue shape; the nature of the tongue coating; the tongue coating is colored; sublingual vessels, etc. 30 symptoms. Auscultation comprises sound; breath sounds; snoring; a cough sound; burping; tone, etc. 6 symptom descriptions. The inquiry includes cold and heat; sweating; the location of the pain; the nature of the pain; head discomfort; physical discomfort; limb discomfort; the ears are not; eyes are absent; sleeping; a diet; thirst or not; external genital condition; anal condition; the length of the menstrual period; menstrual color; menstruation volume; menses; mood; family history; history of vaccination; physiological abnormalities, etc. 22 symptoms. The palpation contains two symptoms descriptions of pulse and pressure. The total number of 60 symptom descriptions is divided into domains according to the expressions contained by different symptoms, so that the purpose of normalizing the input symptoms is achieved. The normalized vector of the input is defined as F. Embedding the normalized vector by an embedding function to obtain an embedded vector E.
GB/T16751.2-1997, the term "clinical diagnosis and treatment of TCM-syndrome" is published, in part, by the national administration of medicine. All symptoms related to the dialectical methods of TCM are classified according to the diagnostic methods of TCM. For example, "white-coated" and "yellow-coated" are classified as apparent tongue coating colors, and label-coded according to different attributes.
The division of symptoms is shown in table 1. Each symptom represents one input dimension, so the overall input dimension is 60 dimensions. Obviously, all symptoms cannot occur simultaneously, so there must be one missing value, so we set the missing value to-1.
TABLE 1 Classification of symptoms
Figure BDA0003426779650000091
According to the dysmenorrheal data subjected to model verification after pretreatment, the main symptoms are liver and kidney deficiency, congealing cold and blood stasis, cold-dampness stagnation, liver depression and damp-heat, deficiency of both qi and blood, qi stagnation and blood stasis, kidney deficiency and blood stasis, damp-heat and blood stasis, and yang deficiency and internal cold. Therefore, in the classification model, our output is focused only on these 9 syndromes. The proportions of the 9 syndrome types are shown in Table 2. It can be seen that there is still an imbalance in our data.
TABLE 2 proportion of syndrome types
Figure BDA0003426779650000092
Through standardization and structuring operations, electronic medical data can be converted into structured data. This data is used as input to the intelligent forensic model. Fig. 2 shows the preprocessing result of an actual electronic product.
Two, Cross-FGCNN intelligent dialectical model
The Cross-FGCNN intelligent dialectic model can automatically learn the feature interaction of high-dimensional and sparse symptom data, extract high-order combination features expressed in a low-dimensional mode from a high-dimensional sparse vector and classify the high-order combination features to complete an intelligent dialectic task.
1) Cross-FGCNN intelligent dialectical model integral structure
The Cross-FGCNN intelligent dialectical model mainly comprises a data embedding module, a data characteristic crossing module, an FGCNN module and a classification module. The model reads the tag-encoded symptom data and performs one-hot encoding according to each field. The high-dimensional sparse data is then mapped onto the low-dimensional dense features using an embedding layer in the data embedding module. And inputting the embedded data as shared input of two parallel modules, namely a linear feature extraction data feature crossing module and a nonlinear feature extraction module. After the two modules generate corresponding features, the two features are merged and input into the classification module, and finally the result pattern is obtained. The overall structure of the Cross-FGCNN intelligent dialectical model is shown in FIG. 3.
2) Data embedding module
The data embedding module consists of many structures, as shown in FIG. 4. According to the symptom classification in the diagnosis of traditional Chinese medicine, the obtained symptoms are mapped to different fields and are subjected to one-hot coding. The purpose of embedding vectors by density is to reduce the dimensionality of the embedded vectors from the field mapping to the input model and to ensure the density of the vectors. For example, the symptoms of white tongue coating are obtained from an electronic medical record and mapped to a coating color field for encoding. The dimension of each field can be reduced to a specified dimension by an embedding operation of the data embedding module, and the dimension of each field in the embedding layer is the same.
3) Data feature intersection module for multi-order linear intersection feature extraction
And linear Cross feature extraction is carried out by using a Cross Network in the model, and the Cross feature of more Cross orders can be obtained along with the increase of the number of layers. The principle of multi-order linear cross feature extraction is shown in FIG. 5, for the output x of the l-th layerl+1From raw input data x0And the output x of the previous layerlSpecifically, the calculation formula is as follows:
xl+1＝f(xl,wl,bl)+xl (1)
Figure BDA0003426779650000111
the residual error of the previous layer is added to the output of the l-th layer, so that the robustness of the model is increased. Obtaining the linear cross feature C through multiple operations, wherein the specific process is as follows:
x of the l layer outputl+1Inputting original data x0And output data x of the previous layerlF (x) calculated by the formula (1)l，wl，bl) By the formula (2), blIndicates a deviation, wlRepresents a weight, x0Representing the initial input vector, xlRepresenting the output vector of the upper layer.
The advantage of equation (2) is that: the input and output dimensions of each layer are the same and the initial features are maintained in the operations of each layer, resulting in the cross feature vector C.
By crossing the initial input vector with the previous output vector, a higher order linear mixture characteristic is obtained. The output of the previous layer is added after the feature crossing, which, like the residual, effectively prevents the gradient from dissipating. As the number of layers of the cross-network increases, the degree of cross-over of features in different domains also increases.
4) FGCNN module for multi-order nonlinear cross feature extraction:
the capability of the model is limited due to the small number of parameters of the cross network. To capture the high-order nonlinear cross features, we introduce a deep network in parallel, as shown in fig. 6. The traditional neural network is difficult to learn local features, and the convolutional neural network can quickly acquire the local features through convolution operation and combine the local features to generate a new model. However, unlike a text or image classification model, the intelligent dialectic model has local relevance to input data, so that the input data is input into a multilayer perceptron model after a Convolutional Neural Network (CNN) outputs cross features of local features to acquire some global cross information. Input to the convolutional layer is a matrix E obtained from the embedded layer, which is nfX k x 1 matrix, where nfIs the number of fields and k is the size of the embedding vector. And then carrying out convolution operation on the convolution kernel with a convolution kernel with the size of h multiplied by 1 multiplied by m, and using tanh as an activation function to obtain a corresponding convolution value, wherein the activation function is shown as follows.
Figure BDA0003426779650000121
Figure BDA0003426779650000122
After the convolution matrix is obtained, performing pooling operation through the maximum pooling layer to obtain a pooled matrix Si. Pooling matrix SiThe convolution layer as the next input is recombined while continuing to be downloaded, and the reconstructed matrix R is obtained by using tanh as an activation function and B as a bias matrix through the following operation1,R2,...,Rn} (n is the number of convolution layers)
Ri＝tanh(SiW+B) (5)
And finally, stretching the reconstruction matrix and then transmitting the stretched reconstruction matrix into a multilayer perceptron to finally obtain a nonlinear high-dimensional feature vector F with local and global cross features.
The specific process is as follows:
on convolutional layers, we obtain the matrix E, i.e. n, from the embedded layersfX k x 1 matrices of nfIs the number of fields and k is the embedding size. Then using a size of h × 1 × m1The convolution kernel of (1) is subjected to convolution operation, and tanh is taken as an activation function to obtain corresponding convolution characteristics. Due to the convolution property, the cross feature of the adjacent h rows can be obtained by using a convolution kernel of h multiplied by 1, and the feature map of a specified number of channels is output. In the first convolution, the number of channels is m at the convolution kernel1First convolution C1The output convolution matrix has a size of nf×k×m1. After obtaining the convolution matrix, the first merged matrix S1Is obtained by max pooling of p 1, and after merging, the matrix size is
Figure BDA0003426779650000123
Merged matrix S1Passed on as input to the next convolution and then recombined. The recombination is a fully-connected operation, as shown in equations (3) and (4), with tanh as the activation function, BiAs ithRecombining the deviation matrices, WiAs ithAnd the first resulting high-order nonlinear feature size is
Figure BDA0003426779650000131
After this, after repeating the convolution, pooling and recombination n times, the entire convolution operation generates n reconstruction matrices, R ═ { R ═ R {1,R2，....Rn}. In order to better combine the characteristics of the output of the cross network, R is serially converted into a new matrix according to a second dimension according to the traditional convolutional network concept, then the reconstructed matrix is introduced into a multilayer perceptron, and finally a nonlinear high-dimensional characteristic vector F with local and global cross characteristics is obtained.
5) Classification module
Combining the linear and nonlinear vectors automatically constructed by FGCNN and Cross Network to obtain a vector I1And (C, F) as an input vector of the multilayer perceptron. i.e. ithThe output of the ith hidden layer is OiThe calculation method is as follows
Oi＝relu(IiWi+Bi) (6)
Wherein WiRepresenting a weight matrix of the ith hidden layer; b isiA bias execution matrix representing the ith hidden layer.
Last hidden layer (n)h) A final decision is then made:
Figure BDA0003426779650000132
in the whole Cross-FGCNN model, the Cross entropy is selected as the loss function of the whole model:
Figure BDA0003426779650000133
where N is the total number of input data sets, y is a real value,
Figure BDA0003426779650000134
is a predicted value.
Example 2
Experiments were performed using the Cross-FGCNN and dysmenorrhoea data mentioned in example 1. 75% of the data were randomly selected as training data and 25% were selected as test data. A 60-dimensional tag encoder vector is input and a 6-level cross network is selected. In FGCNN, convolution kernels with depths of 14, 16, 18 and a width of 4 are selected for three-layer convolution operation, with MLP depth of 3 layers. A3-layer neural network is selected in the classification module, and the number of the neurons is 1024, 512 and 128. To maintain the robustness and optimization efficiency of the algorithm, we choose to discard 0.2, set the learning rate to 0.001, and perform 1000 iterations, where the purpose of setting the dropout parameter is to discard a part of the node connections and prevent overfitting, and the discard is set to 0.2, which means that the discarded nodes account for 20% of the total nodes.
Comparison between Cross-FGCNN model and other models
Six traditional intelligent dialectical models were selected: a Bayes classifier, an ML-KNN classifier, a 10-layer artificial neural network, a decision tree, spectral clustering and a support vector machine. Meanwhile, in order to prove the superiority of the algorithm in sparse symptom classification, three traditional CTR models, namely DNN, FGCNN and DCN, are selected to be applied to intelligent syndrome classification.
1. Results of Cross-FGCNN experiment
First, we calculate the accuracy of the model, which can directly represent the reliability of the model.
Figure BDA0003426779650000141
P represents the correct number of model predictions and the total number represents the total number of input data from the model.
The accuracy of the model Cross-FGCNN was 96.21%. In Table 1, the amount of data between classes of the entire dataset is unbalanced, thus introducing F1-Score and the confusion matrix.
Figure BDA0003426779650000142
P and R represent accuracy and recall, respectively. Cross-FGCNN had an f1 score of 0.9621. It can be seen that Cross-FGCNN is very good at the intelligent identification method of traditional Chinese medicine. FIG. 7 shows a scatter plot of model accuracy change iteration time. The accuracy of the model remained around 96% over approximately 200 iterations.
FIG. 8 shows a classification confusion matrix for Cross-FGCNN, where the model divides all classes as much as possible into the correct classes. In short, Cross-FGCNN can show great strength in intelligent forensic tasks.
2. Comparison between models
Table 3 shows the Cross-FGCNN model compared to other models in terms of accuracy and f1 score. It can be easily found that although the 10-layer artificial neural network has good classification effect in the traditional model, the accuracy of the artificial neural network is still 5% lower than that of Cross-FGCNN, and meanwhile, the CTR model can also display some potential intelligent dialectical tasks, and the accuracy of FGCNN can even exceed that of the traditional intelligent network, but the artificial neural network still has a difference with the Cross-FGCNN model. To demonstrate how each model fits into the imbalance class, log-loss and receiver performance curves were introduced:
Figure BDA0003426779650000151
in the calculation method of the logarithmic loss expressed in the equation, n corresponds to the number of samples or the number of instances of input, and i corresponds to a certain sample or instance; m represents the number of possible classes in the sample, j represents a certain class; y isijIndicating for a sample i, the label that it belongs to class j. Thus, the smaller the log loss, the better the fit of the reaction model, and Cross-FGCNN still performs well in this respect. Compared with other models, Cross-FGCNN has great potential in the aspect of intelligent identification, and is a high-dimensional sparse vector multi-classification task.
TABLE 3 result indices for the respective models
Figure BDA0003426779650000161
The abscissa of the ROC curve is the false positive rate, and the ordinate is the true positive rate. The ROC curve remains constant as the distribution of positive and negative samples varies in the test set. Some syndromes are rare for the dialectics of TCM. Therefore, it is necessary to evaluate the smart forensic model by ROC. Intuitively, the closer the ROC curve is to the upper left corner, the better the classification effect of the model is. Fig. 9 and 10 show ROC curves of the new model, the conventional CTR model, and the conventional intellectual faculty model, respectively. It is clear that Cross-FGCNN has an effect on the classification of different syndromes relative to other models. Secondly, the area under the ROC curve can also be used as one of the indexes of the model classification effect. By comparing the areas under the macro-average ROC curve, the Cross-FGCNN intelligent dialectical model still shows great advantages. For the classification comparison of single syndrome types, the classification of each syndrome type can be easily seen, and Cross-FGCNN can achieve better effect.
While there have been shown and described what are at present considered the fundamental principles and essential features of the invention and its advantages, it will be apparent to those skilled in the art that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference sign in a claim should not be construed as limiting the claim concerned.
Furthermore, it should be understood that although the present description refers to embodiments, the embodiments do not include only one independent technical solution, and such description is only for clarity, and those skilled in the art should take the description as a whole, and the technical solutions in the embodiments may be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims (9)

1. An intelligent dialectical treatment device for traditional Chinese medicine, which is characterized by comprising the following steps:
s1, obtaining symptom characteristic data;
s2, inputting the symptom characteristic data into a Cross-FGCNN intelligent dialectical model, and outputting a syndrome type corresponding to the symptom characteristic data;
wherein the Cross-FGCNN intelligent dialectical model comprises a data embedding module, a data characteristic crossing module, an FGCNN module and a classification module,
the data embedding module is used for mapping the symptom characteristic data into low-dimensional dense characteristic data and outputting the low-dimensional dense characteristic data to the data characteristic crossing module and the FGCNN module,
the data feature cross module is used for extracting multi-order linear cross features in the low-dimensional dense feature data; the FGCNN module is used for extracting nonlinear cross features in the low-dimensional dense feature data;
and the classification module is used for merging the multi-order linear cross characteristics and the nonlinear cross characteristics and outputting the syndrome type corresponding to the symptom characteristic data.
2. The intelligent traditional Chinese medicine dialectical treatment device according to claim 1, wherein the data feature intersection module is configured to extract multi-order linear intersection features from the low-dimensional dense feature data, and the specific implementation method includes the following steps:
x of the l layer outputl+1Inputting original data x0And output data x of the previous layerlBy the formula xl+1＝f(xl,wl,bl)+xlCalculated wherein f (x)l，wl，bl) By the formula
Figure FDA0003426779640000011
Alternative, blIndicates a deviation, wlRepresents a weight, x0Representing the initial input vector, xlThe output vector representing the upper layer is obtained by a plurality of calculations to obtain the cross feature vector C.
3. The intelligent dialectical treatment device of traditional Chinese medicine according to claim 2, wherein the FGCNN module is configured to extract nonlinear cross features from the low-dimensional dense feature data, and the specific implementation method includes the following steps:
s1, inputting n into the convolution layerfX k x 1 matrix E, where nfIs the number of fields, k is the size of the embedding vector;
s2, performing convolution operation on the matrix E and a convolution kernel with the size of h multiplied by 1 multiplied by m, and acquiring a corresponding convolution value by using tanh as an activation function to acquire a convolution matrix;
s3, performing pooling operation through the maximum pooling layer to obtain pooled matrix SiPooling matrix SiAs the input of the next convolution layer, the convolution layer is recombined while continuing to be downloaded to obtain a reconstruction matrix,
the calculation formula of the reconstruction matrix is as follows: ri＝tanh(SiW+B)
tan h is used as an activation function, B is a bias execution matrix, and W is a weight matrix;
and S4, transmitting the reconstruction matrix into a multilayer perceptron to obtain a nonlinear high-dimensional feature vector F of local and global cross features.
4. The apparatus according to claim 3, wherein the classification module is configured to combine the multi-step linear cross features and the non-linear cross features, and output a syndrome type corresponding to the symptom feature data, and specifically includes:
s100, the classification module combines the nonlinear high-dimensional feature vector F output by the FGCNN module and the cross feature vector C output by the data feature cross module to obtain a vector I1＝(C,F)；
S200, converting the vector I1Inputting the (C, F) into a multilayer sensor for calculation, wherein the calculation formula is as follows: o isi＝relu(IiWi+Bi) Wherein W isiRepresenting a weight matrix of the ith hidden layer; b isiA bias execution matrix representing the ith hidden layer;
s300, last hidden layer nhAnd outputting the syndrome.
5. The apparatus according to claim 4, wherein in step S300, the last hidden layer n is a layerhThe calculation formula of the output syndrome is as follows:
Figure FDA0003426779640000031
wherein,
Figure FDA0003426779640000032
denotes the n-thh+1 weight matrix of hidden layer;
Figure FDA0003426779640000033
denotes the n-thhA bias execution matrix of +1 hidden layers,
Figure FDA0003426779640000034
is a predicted value.
6. The apparatus according to any one of claims 1 to 5, wherein the loss function of said Cross-FGCNN intelligent dialectical model is:
Figure FDA0003426779640000035
where N is the total number of input data sets, y is a real value,
Figure FDA0003426779640000036
is a predicted value.
7. The intelligent apparatus for treatment and diagnosis of traditional Chinese medicine according to claim 6, wherein the outputted syndrome type includes: deficiency of both liver and kidney, cold accumulation and blood stasis, cold-dampness stagnation, liver depression and damp-heat, deficiency of both qi and blood, qi stagnation and blood stasis, kidney deficiency and blood stasis, damp-heat stagnation, and yang deficiency and internal cold.
8. The intelligent dialectical treatment device of traditional Chinese medicine according to claim 7, wherein in the Cross-FGCNN intelligent dialectical model, the data feature Cross module selects 6 layers of Cross networks, the FGCNN module selects convolution kernels with depths of 14, 16 and 18 and widths of 4 to perform three layers of convolution operation, the depth of the multi-layer perceptron is 3 layers, the classification module selects 3 layers of neural networks, and the number of neurons is 1024, 512 and 128 respectively.
9. The apparatus according to claim 8, wherein during the training of said Cross-FGCNN model, dropout is 0.2, learning rate is 0.001, and 1000 iterations are performed.
CN202111583091.2A 2021-12-22 2021-12-22 Intelligent dialectical treatment device for traditional Chinese medicine Pending CN114464325A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111583091.2A CN114464325A (en) 2021-12-22 2021-12-22 Intelligent dialectical treatment device for traditional Chinese medicine

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111583091.2A CN114464325A (en) 2021-12-22 2021-12-22 Intelligent dialectical treatment device for traditional Chinese medicine

Publications (1)

Publication Number Publication Date
CN114464325A true CN114464325A (en) 2022-05-10

Family

ID=81405008

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111583091.2A Pending CN114464325A (en) 2021-12-22 2021-12-22 Intelligent dialectical treatment device for traditional Chinese medicine

Country Status (1)

Country Link
CN (1) CN114464325A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116784804A (en) * 2023-07-19 2023-09-22 湖南云医链生物科技有限公司 No-disease analysis and evaluation system

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116784804A (en) * 2023-07-19 2023-09-22 湖南云医链生物科技有限公司 No-disease analysis and evaluation system

Similar Documents

Publication Publication Date Title
Hassantabar et al. CovidDeep: SARS-CoV-2/COVID-19 test based on wearable medical sensors and efficient neural networks
Liu et al. An attention-based CNN-BiLSTM hybrid neural network enhanced with features of discrete wavelet transformation for fetal acidosis classification
Dai et al. Analyzing tongue images using a conceptual alignment deep autoencoder
WO2016192612A1 (en) Method for analysing medical treatment data based on deep learning, and intelligent analyser thereof
CN109544517A (en) Multi-modal ultrasound omics analysis method and system based on deep learning
Zhao et al. Graph convolutional network analysis for mild cognitive impairment prediction
CN109378066A (en) A kind of control method and control device for realizing disease forecasting based on feature vector
Jiang et al. A hybrid intelligent model for acute hypotensive episode prediction with large-scale data
Zhang et al. Hybrid graph convolutional network for semi-supervised retinal image classification
CN109360658B (en) Disease pattern mining method and device based on word vector model
Wang et al. An end-to-end neural network model for blood pressure estimation using ppg signal
Kamada et al. Knowledge extraction of adaptive structural learning of deep belief network for medical examination data
Dai et al. An inception convolutional autoencoder model for chinese healthcare question clustering
CN114464325A (en) Intelligent dialectical treatment device for traditional Chinese medicine
Krishna et al. Standard fetal ultrasound plane classification based on stacked ensemble of deep learning models
Tian et al. A review of traditional Chinese medicine diagnosis using machine learning: Inspection, auscultation-olfaction, inquiry, and palpation
CN112168142B (en) Dysmenorrhea traditional Chinese medicine syndrome differentiation system based on DAELA-LSTM neural network
CN110853764B (en) Diabetes syndrome prediction system
Mendoza et al. Application of data mining techniques in diagnosing various thyroid ailments: a review
Xiao et al. AHE detection with a hybrid intelligence model in smart healthcare
CN116936108A (en) Unbalanced data-oriented disease prediction system
Ning et al. Research on a vehicle-mounted intelligent TCM syndrome differentiation system based on deep belief network
Goyal et al. Musculoskeletal abnormality detection in medical imaging using GnCNNr (group normalized convolutional neural networks with regularization)
Huang et al. Indexing Biosignal for integrated health social networks
Marulkar et al. Nail Disease Prediction using a Deep Learning Integrated Framework

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination