CN109087702B - A four-diagnosis representation information fusion method for TCM health status analysis - Google Patents

Abstract

The four-diagnosis representation information fusion method for TCM health status analysis collects information such as sight, smell, questioning, and diagnosis of clinical patients, and is used to generate multi-source information representation of patients, and mark the syndrome category to which they belong; The characteristic representation of each information source and its category information respectively analyze the health status of the tester, and obtain the auxiliary decision-making information for the tester from multiple information sources; build an information fusion model to maximize the consistency of decision-making, which is used to return the optimal health status. Status analysis results; evaluate the performance of the proposed algorithm by comparing the tester's actual health status with the corresponding predicted results. It can detect the tester's current health status and the nature of the disease, so that the tester can understand his physical condition and provide a reference for the formulation of intervention plans. It can provide high-precision health status analysis results and provide a basis for health care. It can integrate the four-diagnosis characterization information of clinical patients to obtain more accurate and reliable state analysis results.

Description

A four-diagnosis representation information fusion method for TCM health status analysis

technical field

The invention relates to multi-label learning, in particular to a four-diagnosis representation information fusion method for TCM health state analysis.

Background technique

State is the logical starting point of TCM health cognition theory. Health state refers to the comprehensive state of the human body's morphological structure, physiological function, psychological state, and ability to adapt to the external environment within a unit of time, which reflects the state and situation of health. Health status analysis is based on the theory of traditional Chinese medicine. It expresses the collected information such as sight, smell, questioning, and incision in the form of data, emphasizing the objective evaluation of human health status and the nature of the disease, and giving a generalization of the disease and syndrome. Judgment (Li Candong. State of Traditional Chinese Medicine [M]. Beijing: China Traditional Chinese Medicine Press, 2016).

Multi-label learning techniques are used to deal with objects with ambiguity in the real world, and have received extensive attention and applications in the fields of automatic image annotation, bioinformatics, information retrieval, and recommender systems. Specifically, the distribution of syndrome types in clinical patients is often multi-state. Therefore, based on artificial intelligence technology to solve the problem of TCM health status analysis, multi-label learning technology is introduced into TCM health status analysis.

According to the principle of "four diagnostics combined with reference" in traditional Chinese medicine, state analysis is based on the information of four diagnostics. Considering that the contribution of different information sources to the prediction is different, and the different information sources are related to each other, the overall information of clinical patients is collected through the four-diagnosis method, and then an information fusion model is constructed to analyze the health status of the patient. .

TCM health big data presents the characteristics of multi-modality and multi-marker, which makes traditional data analysis theories, methods and technologies face severe challenges such as validity, accuracy and computability. Therefore, studying the information fusion method of the four diagnostics for TCM health status analysis is conducive to the construction of a more accurate and reliable identification model, and is conducive to giving full play to the advantages of artificial intelligence technology to promote the common development and prosperity of interdisciplinary subjects.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a four-diagnosis representation information fusion method for TCM health state analysis, aiming at the multi-state coexistence of clinical patients and the diverse sources of diagnostic information.

The present invention includes the following steps:

1) Collect information such as seeing, smelling, asking, cutting and other information of clinical patients, which is used to generate multi-source information representation of patients, and mark the syndrome type to which they belong;

2) Use the characteristic representation of each information source and its category information to analyze the tester's health status respectively, and obtain the auxiliary decision-making information for the tester from multiple information sources;

3) Build an information fusion model to maximize the consistency of decision-making, which is used to return the optimized health state analysis results;

4) The performance of the proposed algorithm is evaluated by comparing the tester's actual health status with the corresponding predicted results.

In step 1), the described collection of information such as sight, smell, questioning, and incision of clinical patients is used to generate a multi-source information representation of the patient, and the specific method of marking the syndrome category to which it belongs can be:

(1) Extract the four-diagnosis representation information of clinical patients from the electronic medical records to form information source A; obtain the patient's tongue image by using the inspection instrument, realize the tongue image segmentation based on the U-Net network model, and then use HSV, LAB and RGB to describe The operator obtains multiple feature representations of the tongue image, and forms information source B, information source C and information source D respectively;

(2) The doctor marks the health status of the clinically treated patients as {l ₁ ,l ₂ ,...,l _q }, 1≤j≤q, where l _j is the jth syndrome of the clinically treated patient , q is the total number of category labels;

(3) The algorithm is verified by the ten-fold cross-validation method: the processed standardized data is divided into training data and test data according to the ratio of 9:1.

In step 2), the characteristic representation of each information source and its category information are used to analyze the health status of the tester respectively, and the specific method for obtaining the auxiliary decision-making information for the tester from multiple information sources may be:

(1) Using SVM to predict the health status of the tester, the calculation formula is:

表示在数据源A上第i个测试者关于第j个证型的预测结果，

表示第i个测试者在数据源A上的特征表征信息；in,

represents the prediction result of the i-th tester on the j-th syndrome on the data source A,

Represents the feature representation information of the i-th tester on data source A;

(2) Combine the feature representation and the corresponding prediction information to search the top-k neighbors of the tester in the training set. The selection of the neighbors is based on the similarity between the tester and the training samples. The calculation formula is:

包含测试者与训练样本在特征空间上的相似度，用余弦相似性方法计算得到；

由杰卡德相似性方法求得，包含测试者与训练样本在标记空间上的相似度；β为阈值，其取值范围为[0,1]；in,

Contains the similarity between the tester and the training sample in the feature space, calculated by the cosine similarity method;

Obtained by the Jaccard similarity method, including the similarity between the tester and the training sample in the label space; β is the threshold, and its value range is [0,1];

(3) Use the similarity relationship sim ^A to model the correlation between the syndromes to reconstruct the tester's labeling space:

表示在数据源A上第i个测试者关于第j个证型的状态分析结果，Y_zj表示第i个测试者的第z个近邻在第j个证型上的实际值；in,

Represents the state analysis result of the i-th tester on the j-th pattern on the data source A, and Y _zj represents the actual value of the i-th tester's z-th neighbor on the j-th pattern;

(4) Repeat steps (1) to (3) to obtain state analysis results based on information sources B to D, respectively.

In step 3), the information fusion model is constructed to maximize decision consistency, and the specific method for returning the optimized health state analysis result may be:

(1) Use multiple state results predicted by the four-diagnosis representation information of clinical patients to obtain the final result of the tester, and construct the following optimization objective function to solve:

表示第i个测试者在第j个证型上的优化结果，该优化结果通过融合多源的决策信息得到，W＝{w₁,w₂,...,w_M}为M个信息源的权重分布，其中M＝4，

另外，c_m表示(i,j)的集合，且(i,j)满足

α为阈值，其取值范围为[0,1]；in,

represents the optimization result of the i-th tester on the j-th card type, which is obtained by fusing multi-source decision-making information, W={w ₁ ,w ₂ ,...,w _M } is M information sources The weight distribution of , where M=4,

In addition, cm represents the set of (i, _j ), and (i, j) satisfies

α is the threshold, and its value range is [0,1];

设置：

(2) Initialize the weights, let

set up:

(3) Fix W, use the gradient descent method to solve Y ^* , the calculation formula is:

(4) Fix Y ^* , use the Lagrange multiplier method to solve W, and the calculation formula is:

(5) Steps (3) and (4) are repeated until the optimization objective converges, and the optimization result Y ^* of the tester's health state is returned.

In step 4), the specific method for evaluating the performance of the proposed algorithm by comparing the actual health state of the tester with the corresponding prediction result may be:

The proposed method is used to predict the class labels of testers in the test data, and the following five indicators are used to evaluate the performance of the proposed algorithm:

(1) Hamming loss: used to examine the misclassification of the sample on a single marker, the smaller the evaluation index, the better;

(2) 1-Error rate: It is used to investigate the situation in which the label at the front end of the sequence does not belong to the relevant label set in the class label sorting sequence of the sample. The smaller the evaluation index, the better;

(3) Coverage rate: It is used to investigate the search depth required to cover all relevant tags in the category tag sorting sequence of the sample. The smaller the evaluation index, the better;

(4) Sorting loss: It is used to investigate the case of sorting errors in the sorting sequence of the class labels of the samples. The smaller the evaluation index, the better;

(5) Average precision: It is used to investigate the situation where the tags before the relevant tags are still relevant tags in the category tag sorting sequence of the samples. The larger the evaluation index, the better.

Compared with the prior art, the present invention can detect the current health state and lesion nature of the tester, so that the tester can understand his own physical condition and provide a reference for formulating an intervention plan.

The present invention can provide high-precision health state analysis results and provide basis for health care.

The invention can integrate the four-diagnosis representation information of the clinical patients, so as to obtain more accurate and reliable state analysis results.

Description of drawings

Figure 1 is a schematic diagram of tongue image segmentation.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

This embodiment includes the following steps:

1) Collect information such as sight, smell, question, and cut of 729 clinical patients, which are used to generate multi-source information representation of patients, and mark the syndrome category to which they belong, with a total of 339 syndrome categories;

(1) Extract the four-diagnosis characteristic information of clinical patients from the electronic medical records to form an information source A; obtain the patient's tongue image by using the inspection instrument, and realize the tongue image segmentation based on the U-Net network model, as shown in Figure 1. Then, HSV, LAB and RGB descriptors are used to obtain multiple feature representations of the tongue image, which form information source B, information source C and information source D respectively;

(2) The doctor marks the health status of the clinical patient, which is marked as {l ₁ ,l ₂ ,...,l _q }(1≤j≤q). Among them, l _j is the jth syndrome type of the clinical patient, and q is the total number of category markers;

(3) The algorithm is verified by the ten-fold cross-validation method: the processed standardized data is divided into training data and test data according to the ratio of 9:1.

2) Use the characteristic representation of each information source and its category information to analyze the health status of the tester respectively, and obtain the auxiliary decision-making information for the tester from multiple information sources;

(1) Using SVM to predict the health status of the tester, the calculation formula is:

表示在数据源A上第i个测试者关于第j个证型的预测结果，

表示第i个测试者在数据源A上的特征表征信息；in,

represents the prediction result of the i-th tester on the j-th syndrome on the data source A,

Represents the feature representation information of the i-th tester on data source A;

(2) Combine the feature representation and the corresponding prediction information to search the top-k neighbors of the tester in the training set. The selection of nearest neighbors is based on the similarity between testers and training samples, and the calculation formula is:

包含测试者与训练样本在特征空间上的相似度，用余弦相似性方法计算得到；

由杰卡德相似性方法求得，包含测试者与训练样本在标记空间上的相似度；β为阈值，其取值范围为[0,1]；in,

Contains the similarity between the tester and the training sample in the feature space, calculated by the cosine similarity method;

Obtained by the Jaccard similarity method, including the similarity between the tester and the training sample in the label space; β is the threshold, and its value range is [0,1];

(3) Use the similarity relationship sim ^A to model the correlation between the syndromes to reconstruct the tester's labeling space:

表示在数据源A上第i个测试者关于第j个证型的状态分析结果，Y_zj表示第i个测试者的第z个近邻在第j个证型上的实际值；in,

Represents the state analysis result of the i-th tester on the j-th pattern on the data source A, and Y _zj represents the actual value of the i-th tester's z-th neighbor on the j-th pattern;

(4) Compare the prediction results on information source A with BSVM (M.R.Boutell,J.Luo,X.Shen,C.M.Brown,Learning multi-label scene classification,Pattern Recognition,2004,37(9):1757–1771) Compared with LIFT (M. Zhang, L. Wu, LIFT: Multi-label learning with label-specific features, IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(1): 107–120) method, the experimental results are shown in the table 1 shown. Algorithm 1 corresponds to the verification result of the algorithm proposed in the present invention; Algorithm 2 corresponds to the verification result of LIFT; Algorithm 3 corresponds to the verification result of BSVM. It can be seen from Table 1 that the present invention is better than other algorithms in most of the evaluation indexes by considering the tag-related performance.

Table 1

algorithm Hamming loss 1 - Error rate coverage sorting loss average precision 1 0.0122±0.0008 0.1701±0.0330 0.3251±0.0435 0.0617±0.0081 0.6405±0.0265 2 0.0133±0.0009 0.1990±0.0377 0.4168±0.0408 0.0882±0.0085 0.5910±0.0260 3 0.0123±0.0009 0.1358±0.0263 0.3453±0.0436 0.0665±0.0076 0.6355±0.0177

(5) Steps (1) to (3) are repeated to obtain the health state analysis results based on the information sources B to D, respectively.

3) Build an information fusion model to maximize the consistency of decision-making, which is used to return the optimized health state analysis results;

(1) Use multiple state identification results predicted by the four-diagnosis characterization information of clinical patients to obtain the final result of the tester, and construct the following optimization objective function to solve:

表示第i个测试者在第j个证型上的优化结果，该结果通过融合多源的决策信息得到。W＝{w₁,w₂,...,w_M}为M个信息源的权重分布(这里M＝4)，

另外，c_m表示(i,j)的集合，且(i,j)满足

α为阈值，其取值范围为[0,1]；in,

Represents the optimization result of the i-th tester on the j-th card type, which is obtained by fusing multi-source decision-making information. W={w ₁ ,w ₂ ,...,w _M } is the weight distribution of M information sources (here M=4),

In addition, cm represents the set of (i, _j ), and (i, j) satisfies

α is the threshold, and its value range is [0,1];

设置：

(2) Initialize the weights. make

set up:

(3) Fix W, use the gradient descent method to solve Y ^* , the calculation formula is:

(4) Fix Y ^* , use the Lagrange multiplier method to solve W, and the calculation formula is:

(5) Steps (3) to (4) are repeated until the optimization objective converges, and the optimization result Y ^* of the tester's health state is returned.

4) Use the proposed method to analyze the tester's health status in the test data;

The proposed algorithm is compared with the prediction results of each information source, as shown in Table 2. It can be seen from Table 2 that the proposed algorithm can obtain the best results on most of the evaluation indicators by fusing the information sources A to D.

Table 2

Hamming loss 1 - Error rate coverage sorting loss average precision Information source A 0.0122±0.0010 0.1701±0.0330 0.3251±0.0435 0.0617±0.0081 0.6405±0.0265 Information source B 0.0180±0.0014 0.5624±0.0657 0.4090±0.0443 0.0978±0.0071 0.3536±0.0203 Information source C 0.0181±0.0015 0.6090±0.0470 0.4122±0.0412 0.0982±0.0058 0.3447±0.0160 information source D. 0.0181±0.0015 0.5968±0.0595 0.4075±0.0413 0.0968±0.0083 0.3516±0.0241 information fusion 0.0118±0.0011 0.1604±0.0272 0.3328±0.0634 0.0637±0.0108 0.6473±0.0191

The proposed algorithm is compared with other fusion algorithms, as shown in Table 3. Algorithm 1 corresponds to the verification result of the algorithm proposed in the present invention; Algorithm 2 corresponds to the average result predicted based on all information sources; Algorithm 3 corresponds to the voting result predicted based on all information sources, Algorithm 4 concatenates all information sources, and then Classification using SVM. It can be seen from Table 3 that the algorithm proposed in the present invention has the best results.

table 3

algorithm Hamming loss 1 - Error rate coverage sorting loss average precision 1 0.0118±0.0011 0.1604±0.0272 0.3328±0.0634 0.0637±0.0108 0.6473±0.0191 2 0.0161±0.0019 0.2386±0.0350 0.3716±0.0578 0.0778±0.0102 0.5222±0.0190 3 0.0177±0.0020 0.4073±0.0621 0.3715±0.0578 0.0803±0.0101 0.4422±0.0242 4 0.0123±0.0007 0.1427±0.0409 0.3473±0.0427 0.0673±0.0083 0.6343±0.0158

The present invention first preprocesses the information captured by the four-diagnosis acquisition instrument, then analyzes the prediction results of each information source separately to judge the health state of the tester, and finally fuses the prediction results of multiple feature representation information to maximize the consistency of state identification Therefore, it can provide an accurate and reliable reference for testers to formulate intervention plans.

Claims (3)

1. the four-diagnosis representation information fusion method used for TCM health state analysis, is characterized in that, comprises the following steps: 1) Collect the information of seeing, smelling, asking, and cutting the clinical patients, which is used to generate the multi-source information representation of the patients, and mark the syndrome type to which they belong; 2) Use the characteristic representation of each information source and its category information to analyze the tester's health state respectively, and obtain the auxiliary decision-making information for the tester from multiple information sources. The specific method is as follows: (1) Using SVM to predict the health status of the tester, the calculation formula is:

in,

represents the prediction result of the i-th tester on the j-th syndrome on the information source A,

Represents the characteristic representation information of the i-th tester on the information source A; (2) Combine the feature representation and the corresponding prediction information to search the top-k neighbors of the tester in the training set. The selection of the neighbors is based on the similarity between the tester and the training samples. The calculation formula is:

in,

Contains the similarity between the tester and the training sample in the feature space, calculated by the cosine similarity method;

Obtained by the Jaccard similarity method, including the similarity between the tester and the training sample in the label space; β is the threshold, and its value range is [0,1]; (3) Use the similarity relationship sim ^A to model the correlation between the syndromes to reconstruct the tester's labeling space:

in,

Represents the state analysis result of the i-th tester on the j-th pattern on the information source A, and Y _zj represents the actual value of the i-th tester's z-th neighbor on the j-th pattern; (4) Repeat steps (1) to (3) to obtain state analysis results based on information sources B to D respectively; 3) Build an information fusion model to maximize the consistency of decision-making, which is used to return the optimized health state analysis results. The specific methods are: (1) Use multiple state results predicted by the four-diagnosis representation information of clinical patients to obtain the final result of the tester, and construct the following optimization objective function to solve:

in,

Represents the optimization result of the i-th tester on the j-th card type, which is obtained by fusing multi-source decision-making information;

Represents the state analysis result of the i-th tester on the j-th syndrome on the information source m; W={w ₁ ,w ₂ ,...,w _M } is the weight distribution of the M information sources, where M= 4,

In addition, cm represents the set of (i, _j ), and (i, j) satisfies

α is the threshold, and its value range is [0,1]; (2) Initialize the weights, let

set up:

(3) Fix W, use the gradient descent method to solve Y ^* , the calculation formula is:

(4) Fix Y ^* , use the Lagrange multiplier method to solve W, and the calculation formula is:

(5) Repeat steps (3) and (4) until the optimization objective converges, and return the optimization result Y ^* of the tester's health state; 4) The performance of the proposed algorithm is evaluated by comparing the tester's actual health status with the corresponding predicted results. 2. as claimed in claim 1, it is characterized in that, in step 1), in the four-diagnosis representation information fusion method that is used for Chinese medicine health state analysis, it is characterized in that, in step 1), described gathering the sight, smell, question, cut information of clinical patient, for The specific method for generating the multi-source information representation of the patient and marking the syndrome category to which it belongs is as follows: (1) Extract the four-diagnosis representation information of clinical patients from the electronic medical records to form information source A; obtain the patient's tongue image by using the inspection instrument, realize the tongue image segmentation based on the U-Net network model, and then use HSV, LAB and RGB to describe The operator obtains multiple feature representations of the tongue image, and forms information source B, information source C and information source D respectively; (2) The doctor marks the health status of the clinically treated patients as {l ₁ ,l ₂ ,...,l _q }, 1≤j≤q, where l _j is the jth syndrome of the clinically treated patient , q is the total number of category labels; (3) The algorithm is verified by the ten-fold cross-validation method: the processed standardized data is divided into training data and test data according to the ratio of 9:1. 3. the four-diagnosis representation information fusion method that is used for TCM health state analysis as claimed in claim 1, it is characterized in that, in step 4) in, described contrast tester's actual health state and corresponding prediction result to evaluate the proposed. The specific method for the performance of the algorithm is: The proposed algorithm is used to predict the category labels of testers in the test data, and the following five indicators are used to evaluate the performance of the proposed algorithm: (1) Hamming loss: used to examine the misclassification of the sample on a single marker, the smaller the evaluation index, the better; (2) 1-Error rate: It is used to investigate the situation in which the label at the front end of the sequence does not belong to the relevant label set in the class label sorting sequence of the sample. The smaller the evaluation index, the better; (3) Coverage rate: It is used to investigate the search depth required to cover all relevant tags in the category tag sorting sequence of the sample. The smaller the evaluation index, the better; (4) Sorting loss: It is used to investigate the case of sorting errors in the sorting sequence of the class labels of the samples. The smaller the evaluation index, the better; (5) Average precision: It is used to investigate the situation where the tags before the relevant tags are still relevant tags in the category tag sorting sequence of the samples. The larger the evaluation index, the better.

Images