As an algebraic model of information retrieval, LSA was suggested by Susan and other investigators working at Bell Telephone laboratories in 1988 (16–18). It is a calculation theory and method that has been used for knowledge acquisition and representation. With 20 years of development, LSA, which has advantages including strong computability and a decreased requirement of patient involvement, surpasses the disadvantage of the vector space model (VSM) analytical method. In the present study, an LSA-based semantic classification model of syndrome differentiation was used (Fig. 1).

The latent semantic space constructed by SVD was the core of the model. As the basic semantic meaning of syndromes and organs was described in the clinical manifestation collection, the semantic meaning of syndromes and organs is labeled with clinical manifestation collection. Subsequently, classification occurs in the latent semantic space and the correlative degrees with space vectors of syndromes and organs are computed and sorted, and the corresponding organ whose correlative degree is highest as the belonging class is selected.

This model is relatively easy to extend and can be used to classify any aspect on the condition that the classes and objects to be classified have the same description collection, such as the 5-classes classification (the 5 elements) in the present study.

The definition of SVD is as follows:

i)

where U is a mxm dimensional orthogonal matrix whose column vectors are left singular vectors of matrix A. V is an nxn dimensional orthogonal matrix whose row vectors are right singular vectors of matrix A. Σ is an mxn dimensional diagonal matrix whose elements, σ1≥σ2≥…≥σr [r≤ min (m, n)], are singular values of matrix A. Decomposition such as this can be applicable to any matrix. In addition, the rank of matrix A can be the total numbers of non-zero singular values. The definition is as follows:

ii)

where the top ^{γ Α} columns of matrix U is based on the column vectors of matrix A, and the top ^{γ Α} rows of matrix V are based on the row vectors of matrix A. To obtain the similar matrix of matrix A, Α_k (^{k ≤ γ Α}), singular values (in addition to the k highest ones) are altered to zeros. As is shown in the theory of SVD by Brain (19), the distance between matrix A and its similar matrix is determined by minimizing similar matrix Α_k. This is indicated as follows:

iii)

where Α_k=U_k Σ_kV^T_k, U_k is a tx k dimensional matrix whose columns are the top k columns of matrix U, and ^V_k is a dxk dimensional matrix whose rows are the top k rows of matrix V. Σ is a kxk dimensional diagonal matrix whose diagonal elements are the highest k singular values of matrix A. In the case that k is known, it is possible to identify the optimal similar matrix Α_k by using SVD (19).

Major features from samples of each organ are extracted, which represents a certain organ by degrees of vertices. The more degrees the vertex is, the higher the possibility the vertex is to be selected. As shown in Table II, palpitation has a high degree and is likely to be selected. The frequency matrix of syndromes and organs are then constructed based on Table I, as shown in Table III. Subsequently, the frequency matrix in the first step with weighting function was processed, in order to calculate the final mxn dimensional matrix of syndromes and organs ^X = [X_ij]:

iv)

where X_ij is the weight of the clinical manifestation i in syndromes/organs j; row vector X_i = [X_i1, X_i2, …, X_in], i = {1, 2, … m} is the weight of clinical manifestation i in each syndrome/organ corresponding to one row of matrix X; column vector X_j = [x_1j, x_2j, …, x_mj]^T, J = {1, 2, …, N} is the syndromes/organs vector corresponding to one column of matrix X (20,21).

The weight in traditional vector space is obtained using the method term frequency/inverse document frequency (TF/IDF) from statistical computing on the marked frequency of clinical manifestation in syndromes/organs. However, the simple structure TF/IDF cannot effectively provide the expression that indicates the importance and distribution of clinical manifestation. Therefore, it is inappropriate to continue using TF/IDF in the LSA-based semantic annotation. Thus, the improved computing method shown earlier (22) was employed to compute the weight, which was divided into a) the global weight of clinical manifestation, and b) the global weight of syndromes/organs.

Global weight of clinical manifestation, marked T_w (i), is defined as:

v)

where |R| is the quantity of syndromes/organs as the total frequencies of clinical manifestation Ti in the whole syndromes/organs collection Rj. H(R|Ti) can be zero. Adding ‘+1’ created a positive number.

The global weight of syndromes/organs, marked-R_w (j), was:

vi)

where |T| is the quantity of clinical manifestation; Σ^|γ|_i=1 P(i, j) is the total frequency of clinical manifestation Ti in syndromes/organs Rj. Adding ‘+1’ in the formula R_w (j) created a positive number.

The weighting function is composed by equations (v) and (vi) and was defined as:

vii)

The advantage of the weighting function is that it considers the TCM organ classification in its entirety: a) Each syndrome/organ is regarded as a point in the space that uses clinical manifestation as the dimension, and b) each clinical manifestation is regarded as a point in the space that uses syndrome/organ as the dimension.

Correlation calculation is considered based on the formula of similarity calculation. Commonly used similarity calculating formulas are the inner product formula, Pearson formula, Dice coefficient method formula, Jaccard coefficient method formula and cosine formula (22). As the information of the syndromes and organs was expressed as vectors, the cosine formula was used to calculate the degree of correlation. The syndromes and organ vectors D processed by LSA, were divided into part of a) organ, and b) syndrome vectors:

viii)

Therefore, the degree of correlation was calculated in the k-dimension semantic space with D_d and D_q. The formula used to calculate C_q was:

ix)

where sim (D_q, D_dj) was the angle cosine value of the syndrome q vector and organ vector dj. It was determined that the bigger the value, the greater the degree of correlation.

Given a document collection, LDA expresses each document as a theme set, each topic is a multinomial distribution and is used for capturing the relevant information between words (23). In the LDA, the themes are shared by all the documents and embodied by the specific vocabulary in the text (24). Therefore, the implicit theme may be considered as the probability distribution of the vocabulary, and a single document as the mixture of the implicit theme in specific proportion.

The LDA is a modeling method of the text theme information using probability (25). As shown in Fig. 2, it contains the words, topics and document of three institutions. The (α, β) is the parameters of the document collection layer, which determines the LDA model. In the document collection, α is used to describe the relative strength between the themes, β is used to describe the probability distribution of the implicit theme, and θ constitutes a document layer parameter, with the component of θ indicating the weight of each implicit theme of the target. The (z, w) constitutes a word layer parameter, z is the share of implicit theme each word accounts for, and w denotes the word vector of the target document.

In the theory framework of the theme model and the background of the application of the TCM (24), an SHTDT model, inspired by the hidden structure method was established (Fig. 3). The four dark circles are the significant variables, w is the sample symptoms, m is the corresponding TCM for a collection of symptoms, t denotes the corresponding treatment under the theme, t denotes the corresponding diagnosis set under the theme, open circles are the hidden variables, the outermost rectangles is the number of patients, the internal rectangular box indicates the sample of N types of symptoms and their corresponding themes and drugs with regard to the patient, the sample of the treatment methods and the diagnosis in the corresponding theme concerning the patient.

Given the symptoms of the first n (w_i=n) and the drug vector m of the current patient, the Gibbs method was used to estimate the probability of current symptoms assigned to the topic Z_i=k according to the distribution of the theme and the drugs of symptoms except w_i, based on the sampling form shown in formula i. At the same time, the Gibbs method was used to estimate the probability of each drug of the m (x_i=j) to the current symptoms with the corresponding theme. Subsequently, according to the treatment vector u and theme distribution Z_i=k, based on the form shown in formula ii, the Gibbs method was used to estimate the probability of each treatment (u_i=1) of u to the theme Z_i=k. Furthermore, according to the diagnosis vector r and theme distribution Z_i=k, based on the form shown in formula iii, the Gibbs method was used to estimate the probability of each diagnosis (r_i=s) of r to the theme Z_i=k.

x

xi

xii

According to the Gibbs sampling process, the process was iterated to obtain the distribution of φ, θ, η, ε, as indicated:

The theme vector was identified according to the distribution of the theme (from β matrix) in the V-dimensional word space. The similarity between the theme vector was measured by the standard vector cosine distance based on the formula shown in (iv).

xiii

xiv

where the smaller the corre (z_i, z_j), the smaller the correlation of the theme, and the stability of the structure of the theme according to the average similarity between all subjects based on the formula shown in (v).

Improvements to the SHTDT model were made based on the weight, initially when stacking for each set, and adding the weight_i, instead of not adding 1. The weight was derived from the distribution based on the Gaussian function. Since for the diagnosis and treatment data, each symptom in each patient sample generally appeared only once, TF=1. However, if weighted using TF-IDF, it would lead to an increase in the weight of the low-frequency words, and a decrease in the weights of the high-frequency words, which was not a viable result. In addition, during the SHTDT model's initialization, the random assignment of the variables was provided from a wider range of variables. However, the assignment accuracy was not high, which affected the subsequent circulation sampling link. Thus, use of the statistical principle leads to yielding statistics of the drug-symptom, treatment-symptom, and diagnosis-symptom correlations of the patients' record, resulting in the sorting of each combination. For example, to allocate drugs for certain symptoms, ininitally x_i=rand.next (0, the patient of the corresponding drugs). Based on the correlation of the statistics (drug-symptoms), in the corresponding drug episodes of the patients, the most frequent drug was employed. If several drugs were equally frequent, one drug was randomly selected. Of note is that the smaller the selected range, the higher the assignment accuracy. The algorithm of SHTDT model based on weight is shown in Table IV.