Optimization of GERD Therapeutic Regimen Based on ANN and Realization of MATLAB

1.   Introduction

The "rough set" modeling of the "syndrome" of gastro-esophageal reflux disease (GERD) was conducted in our earlier study ^[1]. The model was effective, and the accuracy of differentiation exceeded 90%. However, "treatment determination" following "syndrome differentiation" needs to be studied further.

Treatment determination is a problem of "one condition and multiple decisions." A corresponding treatment regimen (a collection of herbs) should be proposed based on the confirmed type of syndrome, which cannot be performed through the rough set. "Cluster analysis, " which is a classification method for observing individuals and variables, was not considered an optimal choice based on previous discussion. Herbs targeting a certain syndrome can eventually be divided into several categories according to corresponding rules, but the cluster cannot be used to demonstrate which category is more valuable to that syndrome. Additionally, the cluster completely neglects the interaction among Chinese medicinal materials as well as the discrepancy of "monarch, ministerial, adjuvant and conductant herbs" in traditional Chinese medicine (TCM) prescriptions, thus having difficulty in guiding clinical prescriptions.

The diagnosis and treatment for individuals differ from the optimization of decision schemes of large-sample-size evidence-based medicine (EBM). Therefore, a machine was considered for the simulation of the development process of excellent TCM doctors. "learning and training" has been conducted by famous, experienced medical practitioners since ancient times, and relevant rules were determined to consequently achieve optimized results. In this study, artificial neural networks (ANNs) were used for the optimization of therapeutic regimens and realized by MATLAB programming.

2.   Feasibility of ANN to Simulate TCM Syndrome Differentiation

An ANN is an information processing system for simulating the structure and functions of the human brain.

Such simulation is critically important in the TCM field. Firstly, TCM is a highly experiential science by virtue of imaginal thinking or inspirational thinking, which is different from logical thinking. Therefore, it is difficult for traditional methods to simulate this thinking process; however, nonlinear parallel distribution-processed ANNs open up a wider world for the redevelopment of the TCM expert system. Secondly, the experience and knowledge of TCM can be acquired through "learning and training." There are numerous reports on TCM clinical experience for all ages, including prescriptions, medical cases, and treatises, which can better reflect the actual diagnosis and treatment rules of TCM than large-sample-size RCT reports. Learning from mass prescription and medical cases and treatises enables ANNs to grasp TCM differentiation and treatment rules and further realize the artificial intelligence of treatment. Moreover, the fuzziness and uncertainty of TCM theory are difficult to be solved by general mathematical methods, but ANNs may serve as an effective attempt.

The neural network toolbox of MATLAB is the most functional selection to realize ANN. Therefore, the ANN toolbox of MATLAB was adopted as the major study method to perform machine learning and explore the artificial intelligence algorithm.

3.   Realization of ANN through MATLAB Programming

3.1   Thinking of "treatment determination"

The process and pattern of "treatment determination" in the TCM syndrome differentiation and treatment determination need to be understood before the simulation of MATLAB ANN during treatment determination. The process of TCM treatment determination is complicated. After a doctor comprehensively understands a patient's clinical manifestations, he/she has to judge the patient's condition by using his/her knowledge and experience and put forward the corresponding therapeutic regimen; here, judging the patient's condition is syndrome differentiation and putting forward the therapeutic regimen is treatment determination. Certainly, the "syndrome differentiation" is generalized and should contain "differentiation of disease" and/or "differentiation of symptom." Particularly, in specialties such as acupuncture and massage therapy, "differentiation of disease" or "differentiation of symptom" may prevail. There is a mass of treatment of "disease" or "symptom" in acupuncture classics such as Jade Dragon Song (Yu Long Ge, 《玉龙歌》) and Ode for Acupuncture in Treating Various Symptoms (Bai Zheng Fu, 《百症赋》). In addition, a large number of TCM prescriptions for medical decoction only target at "disease" rather than syndrome differentiation. Therefore, "disease diagnosis and treatment determination" may be a more effective expression. This problem was also discussed by the author of another published paper ^[2].

The "disease diagnosis" of TCM must consider various factors, which broadly concern the combination of "heaven, earth, and people." Treatment determination is proposed on the premise of "disease diagnosis." The knowledge and experience of doctors are vital. It is known that a doctor chooses herbs A, C and D for a disease (or syndrome) rather than B, E or F because he/she knows A, C and D are more suitable for that disease and more effective after reading the literature. Therefore, one scenario is that the selection of a certain type of herb is directly proportional to the doctor's memory related to the herb. Hence, the frequency of occurrence of the herb in the literature can serve as one of the bases for a (virtual) doctor to select herbs; another scenario is that the frequency of occurrence of that herb in the literature may be rare, but it was reported once for its specialty, or there was actual application experience by the doctor who recognized the effectiveness of the herb on a particular disease, and who therefore would initiatively choose this herb when treating the same disease. In fact, it is an offset neuron model. It is easy to correspond the former to wp and the latter to b; the output end a represents a numerical matrix permuted according to different selection frequencies of herbs; herbs can be selected based on their frequencies, from high to low. For a specific disease, the prescription with 10 herbs or 20 herbs is determined based on the doctor's preference and experience; i.e., the machine only provides reference, but does not make the final decision. The advantage is that the doctor remains the leading role in "treatment determination" and the machine acts as a significant reference tool. Then, the following questions arise: Why is this process unable to be solved by simple statistical methods? Why is the frequency of occurrence of herbs corresponding to the symptom (and/or symptom group), though able to be calculated, unable to guide the usage of herbs? Because every disease contains a symptom set, and the frequency weight of one herb may be opposite for two symptoms, that is, the weight used for one symptom is high whereas the weight for the other is low, and therefore, it is difficult to make a choice. However, ANN and machine learning can carry out a sufficiently pertinent simulation of samples, thus being able to commendably adapt to and solve this problem.

In other words, "treatment determination" is a process of "fuzzy judgment." Based on his/her knowledge and experience, the doctor organizes the prescription of a disease (or a syndrome). There are two reasons that the doctor chooses one prescription instead of another: one is that his/her knowledge reminds him/her of the effectiveness of the former and ineffectiveness of the latter; the other is that his/her experience informs him/her of the foregoing. The doctor believes the fact of the effectiveness of the two prescriptions owing to certain fuzzy rules in his/her mind. When these fuzzy rules are translated into the weight in the neural network, they can be amended and optimized by using the self-learning and self-adaption abilities of neural network to attain the goal of optimizing the effect of fuzzy logic control.

If "disease diagnosis" is the input end of neural network, the number of messages on experience related to treatment determination is P, and correlation between each message and the curative effect can be expressed as a range of [m, n]. The judgment may be made after i assumptions, and these i assumptions form i layers in the neural network. Finally, after repeated weighting, the prescription is produced, which is the output a. Therefore, it is complementary to the above rough set model. In the rough set model, "syndrome differentiation" has been explained, and in this section, the problem of "treatment determination" is dealt with. That is, so long as information about "disease" and "symptom" is fully extracted, the neural network model is needed to determine the scheme for solving the problem.

3.2   Process and pattern of "treatment determination"

3.2.1   Correspondence of "symptom" and "herb"

In ancient medical books, there is a diagnosis and treatment mode: "D1, D2… Dn are principal for S1, S2…Sm." Si represents symptom and Dj represents herb. Therefore, Si and Dj can constitute an m×n cell array, as shown in Table 1. Pursuant to eight classical medical books Golden Cabinet Stored Recipes (Jin Kui Yao Lue, 《金匮要略》), Thousands of Golden Prescriptions (Qian Jin Fang, 《千金方》), Theory of Spleen and Stomach (Pi Wei Lun, 《脾胃论》), Danxi's Experiential Therapy (Dan Xi Xin Fa, 《丹溪心法》), Wide Aid Formulas (Bo Ji Fang, 《博济方》), Compilation of Jingyue's Works (Jing Yue Quan Shu, 《景岳全书》), Compendium of Materia Medica (Ben Cao Gang Mu, 《本草纲目》) and Practice Records of Chinese Medicine with Reference to Western Medicine (Yi Xue Zhong Zhong Can Xi Lu, 《医学衷中参西录》), the corresponding frequencies of herbs and symptoms are collected, which are standardized as numbers in the interval of [0, 1] to approximately match with doctors' engram; furthermore, they are converted to the output Dj through a certain memory function. Owing to the intermediary syndrome differentiation and disease diagnosis process, the number of hidden layers is set to 2 with 5 nodes in the first hidden layer and 12 nodes in the second. The transfer function of tansig serves as the memory function. The specific syndrome differentiation process and disease diagnosis process are operation processes of functions that can be automatically commanded by the machine through learning.

Table  1.  Induction of correspondence of symptoms and herbs

	D1	D2	…	Dn
S1
S1
…
Sm

 | Show Table

DownLoad: CSV

Summarized correspondence of symptoms and herbs is shown in Figure 1. Blanks represent no relevant experience records in ancient medical books (marked as 0 in MATLAB). Note: expressions of the same symptom in ancient medical books may vary vastly, such as aforementioned "acid regurgitation" for "acid reflux" or "vinegar regurgitation"; "belching" for "hiccup." For avoidance of confusion, the expressions are unified and endowed with related English words; similar processing of herbs, such as "ormosia (Hong Dou, 红豆)" is included in "vigna umbellate (Chi Xiao Dou, 赤小豆)".

Figure  1.  Correspondence of symptoms and herbs

DownLoad: Full-Size Img PowerPoint

3.2.2   Numerical normalization

Based on the occurrence frequencies of symptoms and herbs, the 12×12 numerical matrix of the most common symptoms and herbs is shown in Figure 2. Zero in this table corresponds to blanks in the original table. Normalizing its frequencies into memory weights within [0, 1] by using standard method, the equation is: (x-min)/(max-min). In this case, the maximum value is 22 and minimum value is 0; therefore, normalized values can be obtained by dividing each value by 22 (4 decimal places). Based on the normalized results (taking 4 decimal places), a matrix is built as follows:

Figure  2.  Optimization of symptoms and herbs

DownLoad: Full-Size Img PowerPoint

This matrix is a memory weight matrix, where values represent the possibility of choosing one herb for one disease. It serves as the weight of 12×1-dimension input end in the selection of one herb. SPSS11.5 is used to conduct normality test on this matrix, and the result is compliant with the normal distribution. Eighty percent of unilateral lower limit of normal distribution is taken as the threshold value of one symptom selecting corresponding herbs, i.e., when memory weight is lower than this value, the herb will not be selected; when memory weight is higher than this value, the herb will be selected. Consequently, lower limits of herb selection of each symptom are as follows:

S1:0.3636;S2:0.2273;S3:0.2182;S4:0.5272;S5:0.0909;S6:0.0455;S7:0.4545;S8:0.2273;S9:0.3636;S10: 0.4909;S11:0.0455;S12:0.0818。

3.2.3   Establishment of dataset

To simplify this problem, occurrence of 12 symptoms is considered as groups. There are several numbers of these groups; for example, groups of 5 arbitrary symptoms are $c_{12}^{5}=792$, 4 arbitrary symptoms are 495, and 3 arbitrary symptoms are 220. Therefore, the mapping relation can be built according to the memory weight between sets established based on symptom groups and sets established based on herbs groups. The matrix (n×12) consisting of several symptom sets is the input end and the matrix (m×12) consisting of several herbs sets is the output end.

Given symptoms: s1=acid regurgitation, s2=gastric discomfort, s3=regurgitation, s4=emesis, s5=chest discomfort, s6=globus sensation, s7=chest stuffiness, s8=retching counterflow, s9=stuffiness of stomach, s10=poor appetite, s11=hypochondriac pain, s12=belching. Given herbs: d1= pericarpium citri reticulatae (Chen Pi, 陈皮), d2= Glycyrrhiizae Radix (Gan Cao, 甘草), d3=bighead atractylodes rhizome (Bai Zhu, 白术), d4= Pinellia ternate (Ban Xia, 半夏), d5=rhizomaatractylodis (Cang Zhu, 苍术), d6=poriacocos(Fu Ling, 茯苓), d7=Ginseng Radix (Ren Shen, 人参), d8=ginger (Sheng Jiang, 生姜), d9=medicated leaven (Shen Qu, 神曲), d10=fructus evodiae (Wu Yu, 吴萸), d11=fructus amomi (Sha Ren, 砂仁), d12=Coptis chinensis (Huang Lian, 黄连).

In fact, symptom groups are arbitrary, and there are groups of 2 to 12 symptoms. Four hundred and ninety-five groups consisting of 4 arbitrary symptoms are selected as objects to determine the treatment rules of corresponding herbs. However, these 495 groups have to be figured out first to obtain more accurate data.

MATLAB: combntns ([1,2,3,4,5,6,7,8,9,10,11,12], 4) are used to acquire symptom groups, as shown in Figure 3. The first row is the set consisting of S1, S2, S3, and S4, and corresponding herbs can be obtained from the above weight matrix and lower limits of herb selection; therefore, the first row input into matrix is [1,1,1,1,0,0,0,0,0,0,0,0] and the first row output from matrix is [1,1,1,1,0,0,0,1,0,0,0,1]. For output matrix, 0 represents that the herb is not selected and 1 represents that the herb is selected. Basis of selection is higher than the mentioned lower limits of herb selection, and if 2 or several symptoms establish mapping with one herb, it results in such herb being selected.

Figure  3.  Symptom combination condition

DownLoad: Full-Size Img PowerPoint

3.2.4   Formation of BP network

The newff function is applied to generate a BP network (limited to the length, only 3 data are presented here).

S1= [1 1 1 1 0 0 0 0 0 0 0 0];

S2= [1 1 1 0 0 1 0 0 0 0 0 0];

S3= [1 1 1 0 0 0 0 1 0 0 0 0];

D1= [1 1 1 1 1 1 0 1 0 1 0 1];

D2= [1 1 1 1 1 1 0 1 0 1 0 1];

D3= [1 1 1 1 1 1 0 1 0 1 0 1];

S= [S1; S2; S3]';

D=[D1;D2;D3]';

net=newff([0, 1;0, 1;0, 1;0, 1;0, 1;0, 1;0, 1;0, 1;0, 1;0, 1; 0, 1;0, 1], [5,3,12], {'tansig', 'tansig', 'tansig'});

With this, ANN is generated.

3.2.5   Training and simulation of BP network

Train function is adopted for training whereas sim function is used for simulation.

net.trainParam.epochs=5000;

net=train (net, S, D)

The training process is shown in Figure 4. It is learned that the approximate error is 0 after 3090 epochs.

Figure  4.  Network training process

DownLoad: Full-Size Img PowerPoint

Then, what forms the basis for the generalization effect of this network? One case is input for verification, assuming that the patient has symptoms such as sour regurgitation, bloating, poor appetite and belching; then, these symptoms should be input as:

S4= [1 0 0 0 0 0 1 0 1 0 0 1]';

D4=sim (net, S4)

Figure 5 Network simulation result lists the output effect. This result suggests that medicine that must be selected for the patient includes pericarpium citri reticulata (Chen Pi, 陈皮), Pinellia ternata (Ban Xia, 半夏), rhizoma atractylodis (Cang Zhu, 苍术)and Coptis chinensis (Huang Lian, 黄连), which roughly constitute the major ingredients of Cang Lian Pill (refer to Chapter 5 of Ancient and Modern Medical Guide, 《古今医鉴》) ^[3]. Cang Lian Pill is composed of rhizoma atractylodis (Cang Zhu, 苍术), pericarpium citri reticulatae (Chen Pi, 陈皮), Pinellia ternate(Ban Xia, 半夏), Coptis chinensis (Huang Lian, 黄连), Poria cocos (Fu Ling, 茯苓) and fructus evodiae (Wu Yu, 吴萸), for treating stasis and acid regurgitation. In addition to sour regurgitation, patients always have chronic atrophic gastritis, poor appetite, and belching, which is exactly the presentation of Stasis. Therefore, the patient accords with the index for having Cang Lian Pill.

Figure  5.  Network simulation result

DownLoad: Full-Size Img PowerPoint

Let us consider another example. If the patient has epigastric upset, obstruction of qi in the chest, hypochondriac pain and belching, the symptoms should be input as:

S5= [0 1 0 0 1 0 0 0 0 0 1 1]';

D5=sim (net, S5)

According to the output result, the medicine that can be selected includes Pericarpium citri reticulatae (Chen Pi, 陈皮), Glycyrrhiizae Radix (Gan Cao, 甘草), Rhizoma Atractylodis Macrocephalae (Bai Zhu, 白术), Pinellia ternata (Ban Xia, 半夏), rhizom aatractylodis (Cang Zhu, 苍术), Poria cocos (Fu Ling, 茯苓), fresh ginger (Sheng Jiang, 生姜), Tetradiumdaniellii (Wu Yu, 吴萸), and Coptis chinensis (Huang Lian, 黄连), basically covering the formulas for treatment including Er Chen Wan (二陈丸) and Zuo Jin Wan (左金丸), which well reflects the principal of treatment based on syndrome differentiation of TCM.

The output layer of the aforementioned process can also adopt sigmoid function "hardlim." Then, the result would be output in the form of "0" or "1", the former suggesting this medicine is not selected while the latter meaning this medicine is selected.

MATLAB 6.5 and above is equipped with a neural network toolbox, which can be opened by inputting "nntool" in command window. The computational results are same as aforementioned.

Given limited time and insufficient application experience, it is not feasible to list the research progress and result in detail. Undoubtedly, ANNs open up a new path for research on the ancient prescriptions. However, there is much to be studied in this aspect.

4.   Programmable Disease Diagnosis and Treatment Process of TCM

Figure 6 is excerpted from the literature ^[4], which well reflects the programmable disease diagnosis and treatment process of TCM.

Figure  6.  Structure of rough neural expert system

DownLoad: Full-Size Img PowerPoint

For a disease, research on a large number of samples is first required to conclude the major symptoms and type of syndrome. Then, the attribute set and attribute value are simplified based on the rough set analysis to output the decision rule, which is the treatment process based on syndrome differentiation. Based on this, several ancient and modern literature works are extensively studied to build a database and learn relevant rules of "diagnosis and treatment" by the aforementioned "memory weight" matrix. Through neural network training and generalization, the ANN model is established, and then, the "diagnosis and treatment" process is completed. After programming, automatic treatment based on syndrome differentiation by machine is accomplished via human–computer interaction interface, which is an ideal pattern of the TCM expert system.

5.   Preliminary Achievement

After years of research and exploration, we have built the first edition of "Wiki TCM Artificial Intelligence Expert System" and obtained a software copyright (2017SR696973). Development patterns in the future are worthy of further discussion.

Competing Interests

The authors declare no conflict of interest.