1.   Introduction

As the understanding of Chinese medicine deepens worldwide, its ability to prevent and cure diseases has received increasing attention. The role of traditional Chinese medicine (TCM) in treating various diseases has been acknowledged by the world health organization ^[1]. Due to its special mode of knowledge transfer, the development and application of TCM has been relatively slow, and far from realizing its due value. Chinese Westernization, the outflow of talent, rarity, the lack of qualified successors, the technical and price difference of TCMs, the imbalance of Chinese medical technology development, and limited medical staff are factors restricting the development of TCM. The pace of TCM development cannot meet the current demands of society for TCM.

The Chinese Medicine Institute has issued a "Chinese medicine constitution classification and determination standard". According to this, many scholars have concluded that the clinical research of the current constitution types can be divided into two aspects: one is to investigate the distribution of the constitution types in disease, the other is to study the correlation between constitution types and disease, in order to study and prevent the development of the disease ^[2].

The application of the back-propagation neural network (BPNN) in the field of the Chinese medicine constitution is to bridge the gap between the field of TCM constitution identification and neural networks, the use of information technology to promote the rapid development of Chinese medicine and a TCM constitution identification system, to serve the benefit of a wider population ^[3].

2.   The Construction of the Correlation Between TCM Constitution and Physical Examination Index Using BPNN-based Algorithm

2.1   Characteristics of the constitution and physical examination indexes in TCM

To ensure that the neural network input and output can accurately represent the constitution and physical examination indicators, and to improve the convergence speed of the neural network, it is necessary to quantify the physical type and physical examination indicators of TCM ^[4].

2.1.1   Input information index of the neural network

The examination type is the input of the neural network, with a digital representation of peace quality, Qi deficiency, Yang deficiency, Yin deficiency, phlegm dampness, Blood stasis, Qi stagnation and special qualitiy to represent the nine types of the TCM constitution. Belonging to a physical type is denoted by 1, whereas 0 represents the constitution that the type does not belong to. Therefore, there are nine different TCM constitutions with digital representation as shown in Table 1.

Table  1.  Digital representation of TCM constitutional types

ON	Constitutionaltype	Digital representation
1	Peace quality	100000000
2	Qi deficiency	010000000
3	Yang deficiency	001000000
4	Yin deficiency	000100000
5	Phlegm dampness	000010000
6	Damp-heat	000001000
7	Blood stasis syndrome	000000100
8	Qi stagnation	000000010
9	Special quality	000000001
1	Peace quality	100000000
Note: ON represents order number.

下载: 导出CSV 

| 显示表格

2.1.2   Output information index of the neural network

According to the conventional physical examination, the examination indexes are divided into basic information (age, sex, patient height, weight, body mass index, systolic blood pressure, diastolic blood pressure), routine blood index (white blood cell count, neutrophil count, lymphocytes, monocytes, eosinophils, basophils, the percentage of neutrophils and lymphocytes, mononuclear cells, the ratio of eosinophils, basophil, erythrocyte, hemoglobin, hematocrit, mean corpuscular volume, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, red cell distribution width coefficient of variation, red cell distribution width, the degree of liver function, platelets, total protein, albumin, globulin, albumin/globulin, alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, glutamyl transferase, total bilirubin, direct bilirubin, and indirect bilirubin), renal index (urea nitrogen, creatinine, uric acid, glucose, and carbon dioxide combining power), blood lipids (cholesterol, triglycerides, high density lipoprotein cholesterol, and low density lipoprotein cholesterol).

The quantization index of each output parameter is represented by Ai, and the measured value is expressed as Mi. All the output quantization indexes are limited to the range of [0, 1], and the quantitative expression of the basic information of the patient is as follows:

Aage=Mage/100, Asex=0(man)|1(woman),

Aheight=Mheight/1000, Aweight=Mweight/100,

A_BMI=M_BMI/100,

A_{SystolicBloodPressure}=M_{SystolicBloodPressure}/1000A_{DiastolicBloodPressure}=M_{DiastolicBloodPressure}/100

2.2   Brief introduction of BPNN

Cellular neural network simulation is similar to the way that a human brain works but greatly simplified. The network is composed of many neural network layers, and each layer is composed of many elements. The first layer is the input layer, the last layer is called the output layer, while the rest being hidden layers in the BPNN. Information exchange is only allowed among the units in adjacent nerve layers.

2.2.1   BPNN overview

BPNN is a multilayer feed forward network, and it's also known as back-propagation neural network. It has a highly nonlinear mapping relationship between the input and output of the recognition model. To achieve a nonlinear classification and approximate arbitrary nonlinear function with arbitrary precision, we can transform the activation function or modify the threshold and the number of the layers of the BPNN. BPNN can learn through samples, and automatically adjust the weights in the network, so as to realize the non-inductive logic.

2.2.2   The principle of BPNN

BPNN is one of the most widely used neural network models. The BPNN can learn and store a lot of input-output model mapping to reveal the mathematical equations describing the mapping relation. It applies the method of steepest decent to adjust the weights and thresholds of the network through the BPNN in order to achieve the minimum error sum of squares.

2.2.3   BPNN advantage analysis

The BPNN maps from input to output function, the existing mathematical theory has proved that it can achieve any nonlinear mapping ability for complex nonlinear problems. The neural network can be regarded as a "black box", to facilitate the internal mechanisms ^[5]. Moreover, the subsystems are relatively independent and need no decoupling. It has a high tolerance to errors and self-learning ability.

2.2.4   BPNN algorithm

In the ij hidden layer input vector, jk is the hidden layer output vector, wij is the weight of the connection between the input layer and the middle layers, θj is the hidden layer neurons and $\sum\limits_{i=1}^{n}{{{w}_{ij}}{{x}_{i}}}$ is the hidden layer threshold, the input vector of nerve cell I is xi (i=1, 2, ..., n).

The output function of hidden layer uses the f(h_j) excitation function

In the formula: j=1, 2, ..., m; θ_j=w(n+1)jxn+1, xn+1=-1, yk is the threshold of each neuron in the output layer. The input of the output layer K node is k, and the output layer uses f(h_k) excitation function:

In the formula: k=1, 2, the BP algorithm is a supervised learning algorithm, input the learning sample data for p(x1, x2, ..., x3), and the corresponding monitoring object is t1, t2, ..., t_p, the learning algorithm will be the actual outputy1, y2, ..., y_p and t1, t2, ..., t_p error to modify its weight and threshold, so that yp and tp are as close as possible. The w_jk is the connection weight value of the hidden layer and the input layer, and η is the step size. Set n0=5000 for the number of iterations, using the output of the $\sigma _{jk}^{{{\text{p}}_{i}}}$ and hidden layer neurons of each neuron of the output layer to correct the connection weight, and the correction formula is as follows:

In the formula: y_pis the actual output value, t_p is the error value, o_j is the threshold of each neuron of the hidden layer, and to use the output $\sigma _{jk}^{{{\text{p}}_{i}}}$ neurons of the hidden layer and input layer of each neuron to correct the connection weight, the correction formula is as follows:

In the formula, t_kis the desired output vector and y_k is the actual output vector. The BP algorithm uses a gradient descent algorithm to update the weights in the network. If a batch update algorithm is used, the batch size is set to p, a square error and calculation formula are used, and then the global error of the batch is:

3.   Testing and verification

3.1   The process of the neural network algorithm for comparing the constitution and physical examination index of TCM

According to the demand of the correlation model between the TCM constitution and physical examination index, and the characteristics of BPNN, the algorithm flow of the network model shown in Figure 1 is established.

Figure  1.  TCM constitution and physical examination index BPNN algorithm flow

下载: 全尺寸图片 幻灯片

The constitution of the TCM and Medical Association index of the learning algorithm of NN process is as follows ^[6].

1) The structure of BPNN is constructed. The network model includes an input layer (physical examination index), hidden layers and output layer (physical type).

2) The weights and thresholds of the BPNN are initialized, and the steps and accuracy of learning are determined.

3) A group of Chinese medicine and physical indicators of learning samples (such as the TCM constitution and blood lipid index) is entered for each sample (input vector and the desired output) for learning.

4) Connection weights and thresholds input vector calculate the value of activation of neurons in the hidden layer. The hidden layer activation function calculates the output value of each unit.

5) Connect weights and thresholds with Zi, calculate the activation of the output layer, then use activation function, calculate the output Yj of output layer.

6) The calculation of expected output error, hidden layer output and the actual output from the output layer, BP, modify the weights and thresholds layer by layer.

7) Learning the sample data until the sample data is finished, and judging whether the global error is within the given range of accuracy.

8) Save the threshold, complete the construction of the model.

3.1.1   Collect system input data

The system input data mainly includes blood, blood lipid, liver function and kidney function of the four modules of the data acquisition. The study included 500 subjects from the Affiliated Hospital of Chengdu University of TCM. The TCM constitution types and examination indicators were obtained to assess the corresponding relationship between the samples. With basic personal information hidden, the data were input into an Excel sheet, then the data were entered into an Access database.

3.1.2   Quantitative result data

The output value of the BPNN model is nine types of physique, through testing multiple sets of data and recording the test results to determine the accuracy.

3.2   Applying the model to identify the TCM constitution types

3.2.1   Input generation

The blood urea nitrogen, creatinine, uric acid, glucose, carbon dioxide binding force of several indicators were collected test the physique, the sample data of 253 groups were selected to determine the types of constitution.

3.2.2   Results analyses and documentation

The results generated from TCM diagnosis usually look different from that of Western medicine, but the actual difference between them is small. Using the algorithm, the accuracy of the results was 60%, which still needs optimisation. After the test, the results are recorded.

4.   Discussion

According to the existing constitution types, the identification, application and development status quo and problems in the process of Chinese informatization, the BPNN algorithm was used to construct a correlation between constitution identification and examination index. After the system construction, we collected historical data to train, optimize and collect data to test and analyze the results. Through the blood lipid data detection, the current accuracy was 60%, which is still below the expectation. There is a need to continually improve and perfect the code to obtain a higher accuracy. It is hoped that the application of this system can improve the scope of the TCM constitution identification service and application, and further advance the informatization of TCM.

Acknowledgements

We thank for the funding support from the Young Talents for research on Traditional Chinese Medicine Science and Technology in Sichuan (No.2016Q065).

Competing Interests

The authors declare no conflict of interest.