Study on intelligent syndrome differentiation neural network model of stomachache in traditional Chinese medicine based on the real world

Abstract

Stomachache is not only disease name of Traditional Chinese medicine (TCM) but also the clinical symptom. It is a common and multiple diseases. TCM has its particular advantage in clinical treatment of stomachache. Syndrome differentiation is an important concept in TCM practice. The therapeutic process is virtually a nonlinear mapping process from clinical symptom to syndrome diagnosis with processing and seeking rules from mass data. Artificial neutral network has strong learning ability for nonlinear relationship. Artificial neutral network has been widely used to TCM area where the multiple factors, multilevel, nonlinear problem accompanied by a large number of optimization exist.

We present an original experimental method to apply the improved third-order convergence LM algorithm to intelligent syndrome differentiation for the first time, and compare the predicted ability of Levenberg-Marquardt (LM) algorithm and the improved third-order convergence LM algorithm in syndrome differentiation.

In this study, 2436 cases of stomachache electronic medical data from hospital information system, and then the real world data were normalized and standardized. Afterwards, LM algorithm and the improved third-order convergence LM algorithm were used to build the Back Propagation (BP) neural network model for intelligent syndrome differentiation of stomachache on Matlab, respectively. Finally, the differentiation performance of the 2 models was tested and analyzed.

The testing results showed that the improved third-order convergence LM algorithm model has better average prediction and diagnosis accuracy, especially in predicting “liver-stomach disharmony” and “stomach yang deficiency”, is above 95%.

By effectively using the self-learning and auto-update ability of the BP neural network, the intelligent syndrome differentiation model of TCM can fully approach the real side of syndrome differentiation, and shows excellent predicted ability of syndrome differentiation.

1 Introduction

Traditional Chinese Medicine (TCM) stems from traditional Chinese culture, which studies human of human physiology, pathology, and the diagnosis, prevention and treatment of diseases. Currently TCM is an essential part of the healthcare system in many Asian countries, and is considered a complementary or alternative medical system in most western countries.^[1] According to an international science journal, the World Health Organization has included TCM in its globally influential medical compendium in 2019 for the first time, which will accelerate the development of TCM.

Research methods of TCM take the holistic concept as the dominant thought, the physiology and pathology of organs and meridians as the basis, and the syndrome differentiation as evidence of diagnosis treatment. Treatment based on syndrome differentiation^[2] including 2 processes: syndrome differentiation and treatment. TCM syndrome differentiation is a human thinking process, which is very complicated and without strict rules to follow. Through comprehensive analysis of clinical information gained by the four main diagnostic TCM procedures: observation, listening, questioning, and pulse analysis, the nature of the disease is revealed and the syndrome type is classified. All the data and information have the very complicated nonlinear relationship. In the course of modernization of TCM, some problems, such as there are many impact factors and the proportion of impact factors is difficult to be accurately quantified, need to be solved. In recent years, with the rapid development of computer information processing technology, especially for the increasing mature of data mining and artificial intelligence, these provide key technology for modernization of TCM.^[3]

Researchers have been studied intellectualized system of TCM at home and abroad since 1970s.^[4] In 1990s, artificial intelligence technology has been introduced into intelligent medical system of TCM, such as constructing clinical syndrome differentiation system by fuzzy discriminant model and neural network model,^[5–7] etc. The studies on intelligent medical system of TCM are similar domestic and abroad, such as simulating clinical experience based on the fuzzy pattern of discrimination, using data mining technology and decision tree method or neural network technology.^[8,9]

BP neural network has powerful nonlinear fitting ability and it can approach any complicated nonlinear mapping relationship sufficiently.^[10,11] By adjust the weight of each data constantly, the output values were close to the expected values. This process is very similar to the syndrome differentiation process of TCM. BP neural network was early used to syndrome differentiation and became the main modeling approach in nonlinear system discrimination. It is still a research hotspot in intelligent differentiation of TCM at present. At home and abroad, BP neural network has been successfully applied to auxiliary diagnosis and prognosis of some diseases by some researchers.^[12,13] They include pulse diagnosis expert system, cardiovascular and cerebrovascular diseases, identification of TCM constitution, diagnosis and prediction of tumor, etc.^[14–18]

However, BP neural network also has some disadvantages. For example, the small amount of training data will affect the accuracy of neural network prediction, and the long training time is easy to get into local minimization. Therefore, researchers proposed many improved algorithm of BP neural network, such as gradient descent method, Newton method and Levenberg-Marquardt (LM) algorithm, etc. LM algorithm stems from gradient descent method and Newton method and combines their respective merits. When training BP neural network, scale factor μ is used to control the value of weights, and accelerate the learning speed of network effectively. At the same time, the prediction accuracy is enhanced. But, in the course of training LM algorithm, it costs too much time for the calculation of large matrix, which affects convergence speed and prediction accuracy. To solve these problems, many improved high-order convergence LM algorithms were proposed, such as third-order convergence LM algorithm,^[19] forth-order convergence LM algorithm^[20] and line search-based LM algorithm.^[21,22] These improved LM algorithm can well solve convergence speed and prediction problems and has been used to wireless communication, automatic control, medical areas by engineering experts domestic and abroad.

The real world research was first proposed by Williamson and Barrett initially in 1966.^[23] However, it was not until 1993 that Professor Kaplan proposed it formally in his paper Research on treatment of hypertension with ramipril that RWR received increasing attention.^[24] The clinical research normal form of TCM in the real world is a human-centered, data-oriented, problem-driven, alternation between medical practice and scientific calculation normal form with integration of clinical research.^[25–27] The TCM clinical research paradigm in the real world is shown in Fig. 1. This mass data-based in real world ideas and methods, using modern information network and mathematical statistics, data mining, embody the special advantages of individual diagnosis and treatment in syndrome differentiation process of TCM and make the syndrome differentiation results more objective.

Figure 1: Schematic diagram of TCM clinical research paradigm in the real world.

This study mainly discusses the theory of TCM syndrome differentiation and intelligent treatment in the real world. Based on the real electronic clinical medical data sampled from hospital information system, combining with the theoretical method of machine learning, LM algorithm and the improved third-order convergence LM algorithm were used to build the BP neural network model for intelligent syndrome differentiation of stomachache, respectively. Finally, the differentiation performance of the two models was tested and analyzed. By effectively using the self-learning and auto-update ability of the BP neural network, the intelligent syndrome differentiation model of TCM can fully approach the real side of syndrome differentiation, and shows excellent predicted ability of syndrome differentiation.

2 Methods

The fundamental form of LM algorithm is given by 

The error function of BP neural network is assumed as 

where, y_lk representative the expected output, representative the actual output after calculation. The performance index of neural network F(x) can be expressed as 

Thus, the j_th gradient component is 

The matrix of gradient is given by 

where, J(x) is the Jacobian matrix of E(x).

The second gradient is given by 

where, , it is very small and can be ignored. Thus, can be expressed approximatively as 

The third-order convergence LM algorithm is the improvement for LM algorithm. To reduce the calculated amount of Jacobian matrix, a LM step is calculated by ; another approximateLM step is calculated by , and then the approximate iterative step can be obtained. By replacing , we do not need to calculate J(y_k) but F(y_k). d_k is descent direction, so the algorithm is global convergence which is calculatedby line search or trust region. However, is not always descent direction, the global convergence of algorithm is testified by trust region direction. is nonnegative, but is not always nonnegative. Thus, the actual reduction value of function is defined as . The estimated declining quantity of model function is defined as . It can be proved that the improved LM algorithm is third-order convergence.

The steps of the improved LM algorithm with third-order convergence are given by

Step 1: Selecting initial point x₁, parameter

Step 2: If , then the end; Otherwise, calculate , calculate can be obtained;

Step 3: Calculating and let ;

Step 4: can be obtained;

Step 5: Let k := k + 1 and return step 2.

Before the construction of BP neural network model for TCM intelligent syndrome differentiation of stomachache, the data source of the model must be confirmed. To increase the accuracy of differentiation model, the true clinical case data are used in our research. The data source is clinical electronic medical record data on “digital diagnosis and treatment platform of TCM” of Sichuan administration of TCM. In this platform, electronic medical record data was collected from many TCM hospitals in 19 cities and administered counties, such as Chengdu City, Mianyang City, Deyang City from January 2016 to January 2019. All these uploaded electronic medical records are more than 2 million. These data are the most direct source of data for RWR came from 33 TCM hospitals, with the advantages of large sample size, multi-dimensional richness and authenticity. We select 2436 electronic medical record data of stomachache. These record data is 5 stars, that is, their integrity is highest and curative effect is the best. The database structure needed be normalized before the data extraction and integration. Two Excel format files were derived from the “Digital Diagnosis and Treatment Platform of TCM”. One file includes patient's basic information, disease category, disease location, nature of disease, syndrome type, etc, and the other includes patient's symptom information. The study was approved by the ethics committee of Sichuan administration of TCM.

Even if the integrity of electronic medical record data we selected is the highest in the platform, some problems such as entering error, missing, repetition, lack of standardization may exist because as these data is recorded manually by clinical TCM doctors. These problems result in low quality of data, and it is difficult to make intelligent differentiation analysis. Thus, these data need to be preprocessed before simulation modeling.

For patient's symptom and syndrome type are not in one file, data integration are needed at first. We connect patient's symptoms and syndrome types and then save them in the same file. To guarantee the accuracy of data integration, we compile the program for data integration. The integrated data file retains the patient number, treatmentdate, symptom, syndrome type, disease entity, diseaselocation, nature of disease, therapy and prescription. And then, the data file is cleaned by deleting repetitive patient number and treatmentdate.

On the basis of data structure normalization, further standardization of data processing is carried out. Data standardization in this study included: standardized the symptoms of stomachache on the basis of “GB/T20348-2006 Basic theory terms of TCM” and standardized the syndrome of TCM on the basis of “GB/T16751.2-1997 Clinical diagnosis and treatment terms of TCM-syndrome portion” issued by State Administration of Traditional Medicine of China.

Based on the statistical analysis of the symptoms and syndrome types before normalization, there are 317 different symptoms and 49 different syndrome types. On the basis of China's national TCM clinical diagnosis and treatment terms, a database corresponding to pre-standard and post-standard symptoms and syndromes is established for the symptoms and syndromes in the data file. And then we compiled a program for data replacement. According to the established comparison database of TCM terms, we replace, split, or merge the symptoms and syndrome types in data file. The number of symptoms and syndrome types were reduced to 156 and 37 respectively after normalization.

The symptoms in data file need to be classified and be quantified before used for differentiation model. Based on the characteristics, these data are classified to 16 different types, namely

• 1. stomach feeling,
• 2. stomach pain,
• 3. aggravation or remission factors of stomach pain,
• 4. abdomen,
• 5. face color,
• 6. mouth and pharynx feeling,
• 7. diet,
• 8. urine and poop,
• 9. mental condition,
• 10. head feeling,
• 11. waist and chest and rib,
• 12. cold and hot,
• 13. other feelings,
• 14. pulse,
• 15. tongue, and
• 16. coating on the tongue.

Each kind of symptom is a neuronal node of input layer. The symptom data are is quantified and conversed into coded numerical value used in neural network input. As there are many kinds of symptoms, only the coded numerical value of stomach feeling, pain nature, urine and poop are listed in Table 1. The number of input node decrease because of classification of syndromes. Thus, the calculated amount for learning procedure of differentiation model reduces, which increases the calculated speed.

Table 1: The corresponding coding value of classification symptom.

The symptom of each patient corresponds to 1 one syndrome type, that is, differentiation result of patient is a single syndrome type. Thirty seven kinds of syndrome types need to be divided into 1 kind, and then there is only 1 neuronal node. Syndrome types are converted to the coded numerical value shown in Table 2.

Table 2: The corresponding coding value of syndrome.

Finally, symptoms and syndrome type in normalized data file were converted to coded numerical value. According to the number of all syndrome types, it shows that the case number of syndrome types of “liver-stomach disharmony” and “stomach yang deficiency” is 1154, and 760 respectively and the number of the other types are all less than 50. The data amount of training and testing in the neural network will affect the prediction accuracy. In normal situation, if the data amount is less than 100, the neural network prediction value may have a large error. Thus, we only select the case data ranked in the top 10 syndrome types, shown in Table 3. And then, we used the top 10 syndrome type cases for the training and testing of BP neural network. Here, the data file amounts to 2148 cases after preprocessing, shown in Table 4. For there were too many data rows and columns, only part of data were shown in Table 4. The data file corresponded by Table 4 has 19 columns, in which the 3 to 18 column are 16 kinds of classified syndrome code and the last column is syndrome types code.

Table 3: Syndrome case statistics.

Table 4: Two thousand forty eight sample data.

Based on preprocessed data, BP neuralnetwork model is built. According to the characteristic of TCM syndrome differentiation for stomachache, we design 2 BP neural network models of TCM syndrome intelligent differentiation for stomachache based on LM algorithm and improved third-order convergence LM algorithm. These 2 models all include 3 portions. The first 1 is the input layer with 16 neuronal nodes which are represented by classified syndrome. The second are the double hidden layers to enhance the processing capacity and increase the prediction accuracy of neural network. The node number of hidden layer has a great influence on the performance of the model. If node number of hidden layer is too few, BP neuralnetwork cannot establish complex mappingrelation, leading to a bigger prediction error of network. If node number of hidden layer is too many, the training time of network will be too long. Thus, it is very necessary to select appropriate node number of hidden layer. Here we adopt empirical formula to calculate node number of the hidden layer. Where, “m” and “n” are neuronal node number of input and output layer respectively, “a” is a constant between 0 and 10. By calculations and experiment, when the first node number of hidden layer is 7 and the second one is 8, the best prediction effect can be obtained in LM algorithm network model. When the first node number of hidden layer is 12 and the second one is 10, the best prediction effect can be obtained in improved LM algorithm network model. The third part is the output layer, which has only 1 neuron node, i.e., the syndrome type node.

3 Results and discussion

We program the BP neural network with LM algorithm and its improved third-order convergence algorithm, respectively, on Matlab to realize the above-mentioned intelligent syndrome differentiation model. In the program, the function “getdata” is used to read the preprocessed Excel file into Matlab. The function “divide” is used to divide every case corresponded by every syndrome. For every syndrome, 75% of cases are used as training data and 25% of cases are used as testing data. The function “main_lm” is main function based on LM algorithm. This function gets the data matrix by calling “getdata”. And then it divides the data matrix into training data matrix and testing data matrix by calling “divide”. Having created BP neural network with double hidden layers, we test the network, which is trained, by using the toolbox function “trainlm”. The function “main_3order_lm” is the main function of improved third-order convergence LM algorithm, which is also call “getdata” and “divide” firstly. For there is no improved LM algorithm in the toolbox of Matlab, we program this algorithm based on the above-mentioned steps. In this program, we create the BP neural network with double hidden layers, normalize training data and testing data, iterate it and judge whether it is convergent. Finally, we adjust the weights according to the error and test the trained network.

Definition 1: SA = PCN/ECN × 100%. Where, SA is the syndrome accuracy; PCN is the matching case number of predictive output syndrome; ECN is the case number of expected output. If the predictive output syndrome and expected output syndrome coincide, we consider it is of syndrome accuracy.

Definition 2: DA = PACN/EACN × 100%. Where, DA is the diagnosis accuracy; PACN is the all-syndrome matching case number of predictive output; EACN is the all-syndrome case number of expected output. If the all-syndrome predictive output syndrome and expected output syndrome coincide, we consider it is of diagnosis accuracy.

The data for training network is selected randomly and the network testing accuracy would fluctuate within a certain error range. We adopt the method of training and testing on the network repeatedly to make the accuracy approach the true value.

Firstly, based on the BP neural network with improved third-order convergence LM algorithm, we conducted training and testing for 10 times. The test result is showed in Table 5. Where, the corresponding syndrome of SA1, SA2,…, SA10 are “liver-stomach disharmony”, “stomach yang deficiency”, “stomach ying deficiency”, “cold pathogen invading stomach”, “heat stagnation in liver and stomach”, “deficient cold of spleen and stomach”, “intermingled heat and cold”, “deficiency of both heart and spleen”, “liver depression with spleen deficiency” and “food stagnation in stomach and intestine”, respectively. The last row is the average value of SA in 10 times test of 10 syndromes. The value at the intersection of the last row and the last column is the average value of DA in 10 times test.

Table 5: SA and DA tested by improved third-order convergence LM algorithm.

It can be observed in Table 5 that the average values of SA1 (“liver-stomach disharmony”) and SA2 (“stomach yang deficiency”) are 98.30% and 97.90%, respectively. The others are lower than 80% and they are all lower than the syndrome accuracy for clinical diagnosis. It also can be observed that in the 10 times test, except the first test result in which the DA is 83.76%, the DAs and their average value in the other 9 times test are lower than 80% and lower than the DA for clinical diagnosis as well. It can be seen in Table 3 that the case numbers of syndromes “liver-stomach disharmony” and “stomach yang deficiency” are 1154 and 760, respectively. The case numbers of other syndromes are all lower than 50. The case number of syndrome “food stagnation in stomach and intestine” is even only 12, which is far from the number for training and testing network and result in the reduction of SA average values of the later 8 syndromes and the reduction of DA and the average values of DA in 9 times test. It can be seen than except “liver-stomach disharmony” and “stomach yang deficiency”, the case number of other syndromes is not enough and can not be used in clinical differentiation. Thus, we just keep the syndrome “liver-stomach disharmony” and “stomach yang deficiency” which are meaningful for clinical differentiation. The test results are showed in Table 6.

Table 6: SA and DA tested through improved third-order convergence LM algorithm of “liver-stomach disharmony” and “stomach yang deficiency”.

It can be seen in Table 6 that, by keeping only syndromes “liver-stomach disharmony” and “stomach yang deficiency”, the highest value, lowest value and average value of DA are 99.13%, 97.04% and 98.10%, respectively. The highest value, lowest value and average value of SA of “liver-stomach disharmony” are 99.31%, 96.19% and 98.30%, respectively. Those of “stomach yang deficiency” are 99.47%, 96.32% and 97.90%, respectively.

The highest DA, i.e. the first test, was selected and predictive results of network are showed in Figure 2. It can be seen in Figure 2 that there are only two cases whose predictive outputs are unmatched with expected outputs as to “liver-stomach disharmony” and there only 3 cases whose predictive outputs are unmatched with expected outputs as to “stomach yang deficiency”. The cases whose predictive outputs do not correspond with expected outputs in other syndrome types are many.

Figure 2: Predictive output and expected output of BP neuralnetwork with third-order convergence LM algorithm.

Secondly, based on the BP neural network with LM algorithm, we conducted the same training and testing for 10 times. We found the same question. Except the first 2 SAs whose values are higher than 80%, the other values of SAs are lower than 80%. The values of DA and the average value of DAs in 10 times test are all lower than 70%. Thus, we just also keep the syndrome “liver-stomach disharmony” and “stomach yang deficiency” which are meaningful for clinical differentiation. The test results are showed in Table 7.

Table 7: SA and DA tested through LM algorithm of “liver-stomach disharmony” and “stomach yang deficiency”.

From Figure 3, it showed that the BP neural network with improved third-order convergence LM algorithm and the BP neural network with only LM algorithm have well diagnosis accuracy. However, the diagnosis accuracy of the former is higher and its predictive capability of differentiation is better. Because the calculated amount of Jacobian matrix based on improved third-order convergence LM algorithm are reduced greatly, the SA and DA predicted by neural network increase by 10% or so. Therefore, the BP neural network with improved third-order convergence LM algorithm can well simulate the nonlinear mapping relationship between TCM symptom and syndrome type of TCM. In the network model, the symptoms of TCM are the input nodes and the syndrome types are the output nodes. The symptoms and the syndrome types are summarized as abstract mathematic mapping relation of input and output. BP neural network converts the nonlinearity between symptom and syndrome type of TCM to a nonlinear optimization problem by mapping of input and output. The network is trained based on training data, which include symptoms input value and syndrome types output value. It makes output value as close as possible to the desired value by inputting data and adjusting node weight of the hidden layer. After a lot of training, the weights of neural network are adjusted continually to reach the best values. The weights get sampling knowledge in the form of data and reflect their characteristics and relationships. This is an effective way to reflect inherent rules and characteristics of TCM differentiation by building a nonlinear model of TCM differentiation without opening the “black box” of the human body.

Figure 3: SA and DA of third-order convergence LM algorithm and LM algorithm.

However, there are also some problems in BP neural network used for TCM differentiation. For example, the SA of “liver-stomach disharmony” and “stomach yang deficiency” are very high and the SA of other syndrome types are very low as is shown in Table 5. It shows that if cases are not enough when training and testing the BP neural network, the accuracy of TCM intelligent syndrome differentiation will be affected seriously. Hence, it is indispensable to increase the number of training and testing of the network, so that the accuracy of TCM intelligent syndrome differentiation can be greatly improved.

4 Conclusion

We applied LM algorithm and improved third-order convergence LM algorithm and constructed BP neural network model of intelligent differentiation of TCM for stomachache. The differentiation performances of 2 models were tested and analyzed. Experimental results showed that while the corresponding cases of syndrome “liver-stomach disharmony” and “stomach yang deficiency” were sufficient, the predictive SA and DA of improved algorithm model are all very high, above 95%. It showed that this intelligent syndrome differentiation model can fully approach the real aspects of syndrome differentiation of TCM and showed perfect predictive ability of syndrome differentiation.

Experimental results also showed that when the cases were too few, the predictive accuracy of network would be affected severely. It suggests that if we want to use the BP neural network to realize the TCM syndrome differentiation, there must be enough data to ensure the accuracy of TCM syndrome differentiation. The data in this study were directly recorded by the hospital according to the actual condition of patients. These mass data are from reliable source and reflect the clinical cases of TCM stomachache in real world. The analyzed conclusions can be the reference for clinical diagnosis.

The data in real world comes clinical practice. By transferring relatively vague, nonlinear theory of syndrome differentiation of TCM into verbal sigh understood by public through systematic science, information science, digital technology, and syndrome differentiation of TCM will have not only its theoretical system but also practical computer operation platform. It can not only solve practical problems, but also feedback the found problem to clinic practice to make up for the deficiency of clinic. It provides powerful tools for the popularization and promotion of TCM. Our case data comes from the “digital diagnosis and treatment platform of TCM” of Sichuan administration of TCM, China, and new reliable clinical data of TCM are uploaded every day. With the increase amount of data, more and more clinical case data can be used to provide data assurance for realizing TCM intelligent differentiation based on BP neural network. Our research team will combine more advanced technology to establish an intelligent syndrome differentiation model of TCM for different diseases, and hope to make clinical auxiliary diagnosis system of TCM become clinical intelligent diagnosis system and promote the modernization of TCM.

Author contributions

All the authors have read and approved the final manuscript. YH and GY contributed to the collection, sorting, diagnostics of inquiry information of Stomachache and the article writing. GY is a co-first author of the paper. ZL and LW conceived and revised the paper, designed algorithms and experiments. WCB implemented the algorithms and performed the calculation. LW and WCB are co-corresponding authors. ZY and CY contributed to the Data pre-processing and programming.

Conceptualization: Liang Zhao.

Investigation: Ye Zhang, Yue Cao.

Methodology: Yuan Gao, Chuanbiao Wen.

Supervision: Liang Zhao.

Validation: Ye Zhang, Yue Cao.

Writing – original draft: Hua Ye.

Writing – review & editing: Li Wen.