A novel classification method for aid decision of traditional Chinese patent medicines for stroke treatment

Abstract

Traditional Chinese patent medicines are widely used to treat stroke because it has good efficacy in the clinical environment. However, because of the lack of knowledge on traditional Chinese patent medicines, many Western physicians, who are accountable for the majority of clinical prescriptions for such medicine, are confused with the use of traditional Chinese patent medicines. Therefore, the aid-decision method is critical and necessary to help Western physicians rationally use traditional Chinese patent medicines. In this paper, Manifold Ranking is employed to develop the aid-decision model of traditional Chinese patent medicines for stroke treatment. First, 115 stroke patients from three hospitals are recruited in the cross-sectional survey. Simultaneously, traditional Chinese physicians determine the traditional Chinese patent medicines appropriate for each patient. Second, particular indicators are explored to characterize the population feature of traditional Chinese patent medicines for stroke treatment. Moreover, these particular indicators can be easily obtained by Western physicians and are feasible for widespread clinical application in the future. Third, the aid-decision model of traditional Chinese patent medicines for stroke treatment is constructed based on Manifold Ranking. Experimental results reveal that traditional Chinese patent medicines can be differentiated. Moreover, the proposed model can obtain high accuracy of aid decision.

Keywords

traditional Chinese patent medicines / stroke / aid decision / data mining / manifold ranking

Introduction

Western medicine treatment of stroke has made considerable progress in basic experimental research, but little progress in clinical practice. Stroke treatment based on Western medicine still lack effective methods [1,2]. In China, because clinical experts focus on the compatibility of traditional Chinese medicine (TCM) and Western medicine, the advantages of integrative medicine for the treatment of common diseases and chronic disease rehabilitation have been optimized. Traditional Chinese patent medicines are widely used to treat stroke in local general hospitals. Prescriptions of traditional Chinese patent medicines for stroke treatment are made by both traditional Chinese and Western physicians, with the majority of prescriptions coming from Western physicians. However, current prescriptions of traditional Chinese patent medicines are consistent with the situation referred to in Reference 3. Liu proposes that Western and traditional Chinese physicians account for 97% and 2.5% of the prescriptions of traditional Chinese patent medicines. Similarly, according to the incomplete statistics in Reference 4, 70% of the prescriptions of traditional Chinese patent medicines are made by Western physicians in general hospitals. Another survey found that 91% of the prescriptions of traditional Chinese patent medicines are made by Western physicians in Beijing general hospitals, which was comanaged by Beijing Administration of Traditional Chinese Medicine and Beijing Institute of Traditional Chinese Medicine in 2008. Another statistics [5] has shown that more than 75% of the prescription of traditional Chinese patent medicines is made by Western physicians. Apparently, majority of the prescriptions of traditional Chinese patent medicines are made by Western physicians. However, many Western physicians lack knowledge about TCM theory; therefore, they are unable to grasp key syndrome differentiation. Moreover, Western physicians have difficulty in relearning relevant knowledge of TCM. Thus, considering that the lack of knowledge of Western physicians in prescribing traditional Chinese patent medicines has been increasingly weighty, the irrational use of traditional Chinese patent medicines has become a critical factor threatening the safety and efficacy of TCM [6,7]. However, no research has attempted to study the problem of irrational usage of traditional Chinese patent medicines. In this study, by using advanced data mining technology, the researchers investigate the aid-decision model to resolve the serious misuse of traditional Chinese patent medicines for stroke treatment. With the collected information on symptoms and tongue and pulse diagnosis data, the aid-decision model is constructed to study the latent population feature of traditional Chinese patent medicines for stroke treatment. The aid-decision model provides reliable evidence on the clinical efficacy of traditional Chinese patent medicines for stroke treatment, thereby guiding Western physicians. The data mining technology is extensively applied to resolve several TCM problems [8–13]. Some aid-decision models have especially been proposed to promote clinical efficacy [14–19]. However, these aid-decision models have three disadvantages. First, the decision method belongs to the multiclass classification, i.e., the class is mutually exclusive. A patient is usually given several prescriptions of traditional Chinese patent medicines. Second, the above decision models are sensitive for small samples. In this paper, the stroke patients are relatively few because the methodology of the rational usage of traditional Chinese patent medicines is first discussed in the domain of TCM clinical research. Third, prior knowledge of clinical experts is neglected in these decision models. For the classification of traditional Chinese patent medicines, the knowledge of TCM experts is important to guide the clinical prescriptions. Thus, these decision models are unsuitable for the classification of traditional Chinese patent medicines for stroke treatment. Given the above disadvantages, Manifold Ranking, which is a semisupervised learning method, is applied to generate the aid-decision model of traditional Chinese patent medicines for stroke treatment.

The paper is organized into the following sections. Section “Materials and methods” introduces the data type and the data acquisition method of traditional Chinese traditional patent medicines for stroke treatment. The proposed aid- decision model based on Manifold Ranking is detailed in Section “Construction of the aid-decision model.” In Section “Results, ” the performance of the novel approach is evaluated on the experimental data set. Lastly, the conclusions and recommendations for future work is discussed.

Materials and methods

Data acquisition of traditional Chinese patent medicines for stroke treatment is composed of two parts, i.e., the particular indicators and the types of Chinese patent medicine. The entire data are constructed to form the standard data set of traditional Chinese patent medicines for stroke treatment. The basic flowchart is illustrated in Fig. 1.

Acquisition of patient information

A total of 115 stroke patients are included from the first class of three TCM hospitals, namely, Beijing Dongzhimen Hospital of Traditional Chinese Medicine, Guang’anmen Traditional Chinese Medicine Hospital, and Beijing Tongzhou Hospital of Traditional Chinese Medicine. The training set, which is used to study the aid-decision model based on Manifold Ranking, is composed of 80 patients. Meanwhile, the testing set, which is used to verify the performance of the aid-decision model, is composed of 35 patients.

From the view of stroke patients, four aspects of patient information are collected to mine the particular indicators for traditional Chinese patent medicine for stroke treatment. First, the symptom scale is developed through literature research and expert interview. Second, disease information, e.g., the course and history of disease, is obtained from electronic medical records. Third, tongue and pulse diagnostic information are acquired through four diagnostic instruments from Dawson Medical Devices Company. Finally, the physical and chemical indicator is captured using the hospital information system (HIS) and laboratory information system (LIS). The summary is illustrated in Table 1.

Selection of particular indicators

For the entire patient information, more than 300 indicators are acquired and optimized to represent the population characterization of the stroke patients. The patient information is extensive and numerous for the whole sample. Therefore, the particular indicators are exploited to effectively represent the population feature of stroke patients. Aiming at the characteristics of patient information, the researchers applied data mining methods, e.g., the statistical analysis method, the feature selection method, and the dimensionality reduction method, to exploit the particular indicators. The patient information characteristics are given as follows: (1) the symptom scale is conformed to the binary distribution; (2) the disease information is conveyed by the txt data; (3) the tongue and pulse diagnosis information is obtained by the four diagnostic instruments so that the data are more objective and conformed to the normal distribution; (4) the physical and chemical information is obtained through LIS or HIS to improve objectivity and conformed to the normal distribution. Lastly, 36 particular indicators, i.e., tongue information, are used to establish the aid-decision model of traditional Chinese patent medicines for stroke treatment.

Classification of traditional Chinese patent medicines

A total of 21 common traditional Chinese patent medicines, i.e., Ginkgo Leaf Tablets, Breviscapine tablet, Panax notoginseng Saponins tablet, Xinnaoshutong capsule, SanQi Tongshu capsule, Naodesheng pills, Yindan Xinnaotong capsule, Xiaoshuan Tongluo granules, Naoxuekang tablet, Huatuo Zaizao pill, Zhongfeng Huichun pill, Naoan granules, NaoXinTong pill, Xiaoshuan granules, Tongsaimai tablets, Erigeron shengmai capsule, Day Dan know luo capsule, Xingnao Zaizao pill, bezoar sedative pill, Angong Niuhuang Wan, and Peiyuan Tongnao capsule, were selected for prescription for each patient. Here, 21 common Chinese patent medicines were determined according to the survey from the first class of three TCM hospitals. Subsequently, 16 established TCM clinical experts were recommended to make a prescription for all the stroke patients. First, each stroke patient is given a prescription by more than three clinical experts. Second, for each stroke patient, the prescription is scored according to the rank of traditional Chinese patent medicine. Third, three maximum scores of traditional Chinese patent medicines are taken as the final medicine classification of the patient. In addition, the correlation among traditional Chinese patent medicines is described using the obtained medicine data set, which are provided by these TCM clinical experts. In addition, the classifications of traditional Chinese patent medicines are discretized to binary distribution, but the weight probability of medicines is operated and conformed to the normal distribution.

Construction of the aid-decision model

With the obtained standard data set of traditional Chinese patent medicines, the aid-decision model is constructed based on Manifold Ranking. First, the spatial manifold of patients is constructed by propagating the similarity between the patient information and the traditional Chinese patent medicines. Second, the spatial manifold of medicines is established by propagating the similarity among Chinese patent medicines. Finally, the aid-decision model is characterized by fusing the results of two manifolds. The detailed process of the proposed aid-decision model is illustrated in Fig. 2.

To clearly represent the aid-decision model based on Manifold Ranking, the definition of the symbols is shown first. Let the patient set be P=\{{{\displaystyle p}}_{ki},\hspace{0.17em}k=\mathrm{1},\hspace{0.17em}...,\hspace{0.17em}K\mathrm{;}i=\mathrm{1},\hspace{0.17em}...,\hspace{0.17em}M\}, where K is the total number of patients and M is the total number of the chosen particular indicators. Let the classes of traditional Chinese patent medicines be C=\{{{\displaystyle c}}_j,\hspace{0.17em}j=\mathrm{1},\hspace{0.17em}...,\hspace{0.17em}N\}, where N is the total number of medicine classes. Let the weight probability of the decided medicine be W=\{{{\displaystyle w}}_{ij},\hspace{0.17em}i=\mathrm{1},\hspace{0.17em}...,\hspace{0.17em}M\mathrm{;}\hspace{0.17em}j=\mathrm{1},\hspace{0.17em}...,\hspace{0.17em}N\}, which denotes the probability of the patient {{\displaystyle p}}_i assigned with the medicine class {{\displaystyle c}}_j. The summary is shown in Table 2.

Manifold based on patients

All the patients, including the patients with and without medicine classes, are combined to compute the similarity based on the distance function of the particular indicators. The similarity measurement function is designed according to the feature of the various data, which is illustrated in Table 3.

Then, the propagation of medicine classes is achieved by minimizing the cost function, which is shown in Eq. (1).

(1)Q(r)={\displaystyle \frac{\mathrm{1}}{\mathrm{2}}}\left({\displaystyle {\sum }_{i,j=\mathrm{1}}^N{{\displaystyle S}}_{ij}{\Vert {\displaystyle \frac{{{\displaystyle r}}_i}{\sqrt{{{\displaystyle D}}_{ii}}}}-{\displaystyle \frac{{{\displaystyle r}}_j}{\sqrt{{{\displaystyle D}}_{jj}}}}\Vert }^{\mathrm{2}}}+\mu {\displaystyle {\sum }_{i,j=\mathrm{1}}^N{\Vert {{\displaystyle r}}_i-{{\displaystyle y}}_i\Vert }^{\mathrm{2}}}\right),

where {{\displaystyle r}}_i denotes the medicine class of the patient {{\displaystyle p}}_i. The smoothness measurement is obtained on all the medicine set and is represented by the first part of Eq. (1). Meanwhile, the measurement of the standard deviation between the final medicine class and the compounding true medicine class is represented by the second part of Eq. (1). Furthermore, two measurements are weighted and combined by the factor \mu. Given the factor \mu, the final optimization results, i.e., the propagating results of medicine class, can be achieved at the iteration process. The iteration process is illustrated in Table 4.

Particularly, the characteristic root of the similarity matrix {{\displaystyle M}}_s is conformed to -\mathrm{1}\le {{\displaystyle e}}_k\le \mathrm{1}; hence, the converged solution is obtained and is represented by Eq. (4).

(4){{\displaystyle M}}_w^{*}=(\mathrm{1}-\alpha ){(\mathrm{1}-\alpha {{\displaystyle M}}_{norm})}^{-\mathrm{1}}{{\displaystyle M}}_w(\mathrm{0}).

Manifold based on medicines

The medicines of all the patients are combined to calculate the similarity of medicines by TF-IDF (term frequency–inverse document frequency), which is represented by Eq. (5).

(5){{\displaystyle K}}_{medicine}({{\displaystyle c}}_i,{{\displaystyle c}}_j)={{\displaystyle K}}_c({{\displaystyle c}}_i,{{\displaystyle c}}_j)\times \log \left({\displaystyle \frac{{{\displaystyle N}}_k}{{{\displaystyle n}}_i}}\right),

where {{\displaystyle c}}_i and {{\displaystyle c}}_j denote traditional Chinese patent medicines, {{\displaystyle K}}_c({{\displaystyle c}}_i,{{\displaystyle c}}_j) is the co-occurrence frequency of {{\displaystyle c}}_iand {{\displaystyle c}}_j in the given medicines of all the patients, {{\displaystyle n}}_i is the number of patients with given medicine {{\displaystyle c}}_i, and {{\displaystyle N}}_k is the total number of patients. Here, {{\displaystyle K}}_c({{\displaystyle c}}_i,{{\displaystyle c}}_j) is taken as the similarity matrix of medicines. Similar to the propagating process in Table 4, the manifold based on medicines is constructed by iterating Eq. (6).

(6){{\displaystyle M}}_c^{*}(t+\mathrm{1})=\beta {{\displaystyle M}}_c(t){{\displaystyle K}}_c+(\mathrm{1}-\beta ){{\displaystyle M}}_c(\mathrm{0}),

where \mathrm{0}\le \beta \le \mathrm{1}. In particular, {{\displaystyle M}}_c(\mathrm{0}) is set to be the convergence results {{\displaystyle M}}_w^{*} in Eq. (3). The final weight probability of medicines is achieved at the convergence of Eq. (6), i.e., {{\displaystyle M}}_c^{*}.

Aid-decision process

According to the final weight probability of medicines {{\displaystyle M}}_c^{*}, several kinds of decision methods for choosing the final medicines were found. In this study, each patient without medicine classes is chosen by selecting three medicines with the maximum weight probability.

Results

Performance of aid-decision model

The aid-decision model based on Manifold Ranking is studied in 80 training patients, and then the precision and the recall of the model is evaluated in 35 testing patients. The performance is evaluated from two views. First, from the view of medicine evaluation, two evaluated criterions are defined by Recall={\displaystyle \frac{B}{C}} and Precision={\displaystyle \frac{B}{A}}. Here, A represents the number of patients automatically decided by a given medicine in the top three returned medicine list, B is the number of patients correctly decided with that medicine in the top three returned medicine list, and C is the number of patients having that medicine in the ground truth medicine. Second, from the view of patient evaluation, the average precision is defined by Precision={\displaystyle \frac{P}{N}}. Here, P is represented as the number of patients with more than two accurate medicines in the top three returned medicine list, and N is the number of tested patients. The experimental results are illustrated in Table 5.

Effectiveness of particular indicators

The training data are difficult to obtain in the current clinical practice; hence, the particular indicators are necessary to be mined to improve the performance of the aid-decision model. The dispersion of the patient information is analyzed for each medicine class using the statistical and feature selection methods. Finally, the tongue information is salient for the classification of traditional Chinese patent medicines. Three statistics based on different indicators are illustrated in Fig. 3.

Discussion

Based on the experimental results, traditional Chinese patent medicines for stroke treatment can be classified with proper data mining method. Thus, the better aid decision guidance helps Western physicians improve the irrational usage of traditional Chinese patent medicines for stroke treatment. However, from the experimental results, we find that some aspects are significant for the aid-decision model to influence the precision and recall performance.

First, from the results in Table 4, the average precision of 21 medicines is achieved at 85.79%, and the average precision of patients reached 62.69%. The main reason for the difference is that the accuracy of one medicine is considered by the average precision of medicine. In addition, average recall of medicine is achieved at 61.44%. The result shows that the training data of some medicines is less to reduce the performance of medicine classification. Therefore, the samples of traditional Chinese patent medicines are enlarged to ensure better performance of the aid-decision model based on Manifold Ranking.

Second, from the results in Fig. 3, the optimal performance for three evaluated criterions, i.e., average precision of medicine, average recall of medicine, and average precision of patients, is achieved for the aid-decision model based on tongue information. The result is identical to the conclusion, i.e., the tongue information is salient for the classification of traditional Chinese patent medicines based on the dispersion statistics. Conversely, the worst performance is achieved for the aid-decision model based on scale. The possible reason is that the representation of scale is sparse to convey less quality of differentiation information. Another reason is that tongue information is the significant prescription criterion of stroke patients in clinical practice. In addition, the performance of integrated patient information is normally worse than that of single patient information, e.g., pulse and tongue. The possible reason is that 115 patients are relatively less than the high dimension of integrated information. Therefore, more effective indicators are necessary to be mined to characterize the population feature of traditional Chinese patent medicines for stroke treatment. Moreover, effective indicators are helpful in the improvement of the performance of the aid-decision model based on Manifold Ranking.

Finally, some patient information, e.g., the quality improvement of symptom scale and the method of medicine selection, must be enhanced to improve the performance of the aid-decision model.

Conclusions

In this study, the aid-decision model based on Manifold Ranking is proposed to automatically classify traditional Chinese patent medicines for stroke treatment. The particular indicators are selected to characterize the population feature of traditional Chinese patent medicines for stroke treatment and the differentiation property of medicines. This is helpful to improve the performance of the aid-decision model. Furthermore, the Manifold Ranking process can naturally make full use of both the relationships among all the patients and the relationships between the given medicine and the patients. The weight probability is determined for each patient in the Manifold Ranking process to denote the biased information richness of the patients. With the studied aid-decision model, the accuracy of prescription is improved for Western physicians so that the efficacy of traditional Chinese patent medicines for stroke treatment is largely promoted.

References

[1] Li S, Ji J. Rational use of Chinese traditional patent medicine. China Mod Med (Zhongguo Dang Dai Yi Yao) 2011; 18(12): 102–104 (in Chinese)

[2] Wu J, Hu M, Song M, Wu P, Jiang Y. Catalog proprietary varieties of essential drugs listed by the production distribution. China Pharm (Zhongguo Yao Fang) 2010; 21(24): 2214–2218 (in Chinese)

[3] Yang L. Adverse reaction reports and analysis of 416 cases in a region proprietary. Med Hyg (Yi Yao Wei Sheng) 2015; 1(8): 159–159 (in Chinese)

[4] Lu L. Chinese traditional patent medicine comparable to antibiotic abuse. Nanfang Daily (Nanfang Ri Bao), August 23, 2011; B01 (in Chinese)

[5] Guo C, Huang L. Rational application of proprietary Chinese medicines in the treatment of coronary heart disease. Chin J Integr Med Cardio Cerebrovasc Dis (Zhong Xi Yi Jie He Xin Nao Xue Guan Bing Za Zhi) 2016; 14(18): 2131–2133 (in Chinese)

[6] Wu S. Reason of abuse for Chinese Traditional Patent Medicine. Xinhua Daily Teleg (Xin Hua Mei Ri Dian Xun), August 24, 2011; 01 (in Chinese)

[7] Wan K. Analysis and unreasonable application of preventive measures in the prescription medicine. Medicine (Yi Yao) 2015; 4(1): 251 (in Chinese)

[8] Guan Y, Zhou C, Liu J. Problem analysis of Chinese traditional patent medicine made by Western medicine doctors. Chin J Clin Ration Drug Use (Lin Chuang He Li Yong Yao Za Zhi) 2010; 3(13): 61 (in Chinese)

[9] Li G, Yan S, You M, Sun S, Ou A. Intelligent ZHENG classification of hypertension depending on ML-kNN and information fusion. Evid Based Complement Alternat Med 2012; 2012: 837245 

[10] Zhang X. Application of topic model in the TCM clinic. Doctoral dissertation. Beijing Jiaotong University, 2011 (in Chinese)

[11] Zhang R, Zhou X, Yao N. Herb combination relation of liver tune certification based on association rule. Chin J Info Tradit Chin Med (Zhongguo Zhong Yi Yao Xin Xi Za Zhi) 2010; 17(2): 97–99 (in Chinese)

[12] Zhou X, Chen S, Liu B, Zhang R, Wang Y, Li P, Guo Y, Zhang H, Gao Z, Yan X. Development of traditional Chinese medicine clinical data warehouse for medical knowledge discovery and decision support. Artif Intell Med 2010; 48(2-3): 139–152 

[13] Zhao YF, He LY, Liu BY, Li J, Li FY, Huo RL, Jing XH. Syndrome classification based on manifold ranking for viral hepatitis. Chin J Integr Med 2014; 20(5): 394–399

[14] Feng Q, Zhou X, Huang H, Yu J, Zhang Y, Tong X, Zhang R, Liu B. A MDP solution for traditional Chinese medicine treatment planning. International Conference of Biomedical Engineering and InformaticsOctober 16–18, 2010, Yantai, China. 2250–2254

[15] Feng Q. POMDP approximate solution and in Chinese medicine clinic program optimization.  Doctoral dissertation. Beijing Jiaotong University, 2011 (in Chinese)

[16] Soni J, Ansari U, Sharma D, Soni S. Predictive data mining for medical diagnosis: an overview of heart disease prediction. Int J Comput Appl 2014; 8(17): 43–48

[17] Wu D, Lin J, Wang P, Wang Y, Lu F. Analysis on composition principles of prescriptions for erectile dysfunction by using traditional Chinese medicine inheritance system. Chin Tradit Patent Med (Zhong Cheng Yao) 2016; 38(4): 755–759 (in Chinese)

[18] Li Y, Zheng G, Liu L. Treatment rules of Sinomenium acutum by text mining. World Chin Med (Shi Jie Zhong Yi Yao) 2015; 10(6): 823–827 (in Chinese)

[19] Peng J, Hu J, Liu B. Clinical aid-decision support system developed based on the network information technology. International Conference of Traditional Chinese Medicine EngineeringDecember 6–8, 2006, Shanghai, China. 134–137 (in Chinese)

RIGHTS & PERMISSIONS

Higher Education Press and Springer-Verlag Berlin Heidelberg