Exploring the combination and modular characteristics of herbs for alopecia treatment in traditional Chinese medicine: an association rule mining and network analysis study

We included all kinds of studies regardless of the study design, and we did not restrict the type of alopecia or herbal medicine used. Thus, all kinds of herbal medicines were considered—extracts, decoctions, pills, and even external application. There were no restrictions on sex, age, disease duration, or disease severity. The outcome of clinical studies was not considered.

We conducted an electronic search of the Chinese, English, and Korean databases from their inception to March 2017. We searched one Chinese database: the China National Knowledge Infrastructure (CNKI) database, three English databases: Embase, Medline (via PubMed), and the Central Register of Controlled Trials (CENTRAL), and one Korean database: the Oriental Medicine Advanced Searching Integrated System (OASIS), which specializes in traditional Korean Medicine research articles. [9] The following search terms for alopecia were included: “Alopecia”, “Alopecia areata”, “Diffuse alopecia”, “Androgenic alopecia,” and “Female pattern hair loss.” An additional file describes details of the search terms and search strategies used in each database to identify alopecia (see Additional file 1).

We extracted the name of each herbal formula, the medicinal herbs that comprised it, its origin (name of article or ancient literature), author, publication year, and internal/external application. The names of the herbs followed the Chinese Pharmacopoeia 2015 edition, [8, 10] which can be found on the OASIS and KIOM Herbarium website (http://​boncho.​kiom.​re.​kr/​herbarium/​codex.​php). [11] The categorization of each herb followed Phytology [12] and Chinese Pharmacy. [13] The names of the 20 classification categories of medicinal herbs are as follows: (1) Exterior-releasing medicinal (解表藥), (2) Heat-clearing medicinal (淸熱藥), (3) Purgative medicinal (瀉下藥), (4) Wind-dampness dispelling medicinal (祛風濕藥), (5) dampness-resolving medicinal (化濕藥), (6) Dampness-draining diuretic medicinal (利水滲濕藥), (7) Interior-warming medicinal (溫裏藥), (8) Qi-regulating medicinal (理氣藥), (9) Digestant medicinal (消食藥), (10) worm-expelling medicinal (驅蟲藥), (11) hemostatic (medicinal) (止血藥), (12) Blood-activating and stasis-dispelling medicinal (活血祛瘀藥), (13) Cough-suppressing and panting-calming medicinal (止咳平喘藥), (14) Tranquillizing medicinal (安神藥), (15) Liver-pacifying medicinal (平肝藥), (16) Orifice-opening medicinal (開竅藥), (17) Tonifying and replenishing medicinal (補益藥), (18) Astringent medicinal (收澁藥), (19) Emetic medicinal (湧吐藥), and (20) External application medicinal (外用藥).

First, we compared the meridian tropism and classification category of each medicinal herb between the internal and external applications. The criteria for meridian tropism followed the classification of the Phytology [12] and Chinese Pharmacy. [13] Next, we conducted a data mining analysis using ARM and network analysis. Because the mechanism of action differs depending on the route of administration, we only carried out this analysis on the internal application formulas only, not the external application formulas. Using ARM, we identified the most frequently used two-herb combination and three-herb combination. Using network analysis, we categorized the herbs used in alopecia treatment into several modules.

Using the list of prescriptions used to treat hair loss, we searched for combinations of herbs repeatedly used over several prescriptions. For this purpose, we applied ARM, which uncovers interesting relationships in large datasets, to our data. [14] Because ARM is generally used in business to analyze customers’ purchase data, the terms “item” and “transaction” are widely used. In our analysis, the herbs were defined as items, and the prescriptions were defined as transactions recording co-occurrences of items. We let H = {h1, h2, …hd} be the set of all herbs in the bald prescription data, and P = {p1, p2, …, pn} be the set of all prescriptions. In ARM, a collection of zero or more items is termed an itemset. An association rule is an expression of the form X → Y, where X and Y are disjoint itemsets. The expression represents the relationship between the occurrences of itemset X and itemset Y. The strength of the association rule can be measured in terms of its support, confidence, and lift. Support determines how often a rule is applicable to a given data set, while confidence determines how frequently items in Y appear in transactions that contain X. Support indicates how frequently the rule can be applied to a given set of data, and confidence indicates how often Y appears in transactions containing X. Lift is the ratio of observed support to expected support when X and Y are independent. Support is a measure of whether an association between X and Y happens by chance, and confidence represents the reliability of the association. Lift values larger than 1 indicate that the occurrences the two itemsets are dependent on each other. These measures suggest a strong co-occurrence relationship between itemsets X and Y. In the present study, ARM for combinations of two herbs and three herbs was applied using the a priori package of R (R Core Team (2013). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria), and minimum thresholds on support and confidence were set at 0 and 1, respectively.

ARM cannot inspect the overall pattern of how herbs are used together, because it assesses the association between limited numbers of itemsets. Therefore, we constructed a network that connected the herbs used together in alopecia prescriptions. We also examined the modularity analysis to identify patterns and group herbs into specific modules. The network between the herbs was configured using Python’s networkx® package (https://​networkx.​github.​io/). [15] The nodes of the network were defined as all the herbs that appeared in alopecia prescriptions. The herbs that appeared together in at least one prescription were defined as having a linkage between each other, and the network was constructed as a graph that weighted connections based on the number of co-occurrences in different prescriptions. The dose of the herb in each prescription was not considered in determining weight or linkage. Modularity analysis and network visualization were performed using Gephi. [16] Modularity analysis was performed using the Louvaine method, with a resolution value of 1.0 [17]. Visualization was performed using a circular layout in which modules were classified into categories.

To observe differences among the modules identified modularity analysis, the meridian tropism of the herbs composing each module was examined. Meridian tropism is the notion that a herb predominantly exerts a therapeutic effect on a specific organ or meridian in the human body [18]. The meridian tropism (引經) of each herb is recorded in the classic book, and it reveals the characteristics of each herb from the perspective of Korean medicine. Based on the meridian tropism of each herb listed in herbal textbooks and the Korean Intellectual Property Office database, we investigated the ratios of therapeutic preferences for each meridian of the herbs in each module. The permutation test was then applied to find statistically significant meridian preferences. Briefly, a list of module labels of herbs was randomly permutated, the meridian preference ratio per module was calculated, and the process was repeated 10,000 times to obtain a null distribution of meridian preference ratios. A p-value was then calculated based on the location of the observations within the simulated null distribution. We tested 12 meridian preferences separately for each module, with a correction for multiple testing using the false discovery rate.

A total of 585 articles were screened: 286 in the English databases, 73 in CNKI, and 226 in OASIS. Ultimately, 56 articles were included after screening of the full text: 12 from the English databases, [19‐30], 13 from CNKI, [6, 31‐42] and 31 from OASIS. [43‐73] Details of the screening process are shown in the PRISMA flow chart (Fig. 1).

From the 56 articles, 489 herbal medicines (312 internal and 177 external) and 374 medicinal herbs were identified. Among the 312 internal medicines, 258 medicinal herbs were identified. Among the 177 external medicines, 257 medicinal herbs were identified. Many medicinal herbs were used in both internal medicine and external applications. Table 1 lists the frequently described medicinal herbs.

The frequency trend of the 258 herbs in the 312 internal alopecia prescriptions were analyzed using the a priori ARM method to elucidate whether certain herbs are used more frequently in combinations of two or three. The top 10 association rules between two herbs are described in Table 2. The association of Polygonum multiflorum Thunb. (何首烏) and Angelica sinensis (Oliv.) Dlels (當歸) had the highest support, with 38.5%. The six possible pairs from a group of 4 herbs—Polygonum multiflorum Thunb. (何首烏), Angelica sinensis (Oliv.) Dlels (當歸), Ligusticum chuanxiong Hort. (川芎), and Rehmannia glutinosa Libosch. (Prepared) (熟地黃)—were listed as the top 6 association rules, indicating that these four herbs are the most commonly prescribed herbs, and that they are frequently used together. Other rules in the top 10 list contained Ligustrum lucidum Ait. (女貞子), Lycium barbarum L. (枸杞子), and Eclipta prostrata L. (旱蓮草) as additional combination herbs.

The top 10 association rules for three-herb combinations are described in Table 3. The result again showed the importance of four main herbs. Of the four possible combinations of three herbs from among the four main herbs—Polygonum multiflorum Thunb. (何首烏), Angelica sinensis (Oliv.) Dlels (當歸), Ligusticum chuanxiong Hort. (川芎), and Rehmannia glutinosa Libosch. (Prepared) (熟地黃)—three combinations were listed in the top 4 association rules, with support of 23.4, 21.5, and 21.2%, respectively. One other three-herb combination of the four main herbs, excluding Angelica sinensis (Oliv.) Dlels (當歸), was listed as the No. 8 association rule.

Paeonia lactiflora Pall. (白芍藥) did not appear in the top 10 association rules of two-herb combinations, but it was listed as a member of three-herb combinations in association rule Nos. 5 and 9. Cuscuta chinensis Lam. (菟絲子) was also not shown in the top 10 association rules of two-herb combinations, but it did appear among the top 10 association rules of three-herbs combinations.

Based on the frequency of co-occurrences of herb pairs, we constructed a weighted unidirectional network, assigning “frequency of co-occurrence” as a weight value of the connection. We then performed a modularity analysis on the weighted graph using the Louvain method. The results showed that the network could be divided into three modules, with a modularity value of 0.141. The number of herbs comprising each module were 58, 86, and 111, respectively (Fig. 2). Unconnected single herbs were excluded from the modularity analysis. The top 20 most frequent herbs in each module are described in Table 4.

The frequency trend for meridian tropism in each module is described in Fig. 3. Significantly high and low meridian tropism frequency ratios were found in each module using the permutation test. Module 1 had significantly more herbs that preferred “Liver,” while Module 3 had significantly more herbs that preferred “Stomach.”