Classifications of Chinese herbal medicines are multifaceted and multileveled [2]. Siqi (四气), Wuwei (五味), and Guijing (归经) are the basic attributes for herb classification and have been an important part of TCM research. Siqi represents the properties of Chinese herbal medicines according to their functions on the human body: cold (寒), hot (热), warm (温), and cool (凉). In addition, herbs with gentle properties are namely calm (平). Wuwei means flavors: pungent (辛), sweet (甘), sour (酸), bitter (苦), salty (咸), tasteless (淡), and astringent (涩). It is believed that these factors are associated with body heat production processes or metabolic activities and may also play a role in the digestive system, nervous system, and cardiovascular system [8]. Guijing regards the orientation of Chinese herbal medicines, which is to closely connect the functions of herbs with the organs and meridians (脏腑经络) of the human body.

Another important concept for herbs in the formula is Jun-Chen-Zuo-Shi (君臣佐使). Jun-Chen-Zuo-Shi is the principle of the compatibility of TCM formulas. Junyao (君药), namely, principal herbs as used hereafter, plays a major role against the main disease or syndrome. It is the primary herb used in the formulas. Footnote c in Table 1 indicates Junyao in the corresponding formulas.

In this work, the medicine formulas data were extracted from the key medicine formulas of the textbook Chinese Herbal Formulas (Tenth Edition) [7], as shown in Table 1. Multidimensional herb attribute data were retrieved from the SymMap database [9]. Siqi has 5 dimensions: cold, hot, warm, cool, and calm. Wuwei has 7 dimensions: pungent, sweet, sour, bitter, salty, tasteless, and astringent. Guijing has 11 orientations: liver meridian, heart meridian, spleen meridian, lung meridian, kidney meridian, bladder meridian, large intestine meridian, small intestine meridian, stomach meridian , gallbladder meridian, and pericardium meridian. These properties were combined and represented as a 23-dimensional vector for each herb.

This study did not involve human subjects research. The data used in this study were obtained from a publicly available database and a textbook.

Our goal was to devise a joint visualization method of medicine formulas and the attributes of corresponding herbs. The visual design should support the comparison of formulas and facilitate the classification of herbs based on their properties (Siqi, Wuwei, and Guijing). Visualization and TCM experts worked closely together to analyze the requirements of the visual analysis method for medicine formulas. The requirements are summarized as follows:

Our method is the result of an iterative development process using quick prototypes. Prototypes were realized based on the requirements and proposed to the TCM expert (SP, one of the authors), and improvements were made based on the feedback of the TCM expert.

The workflow of our method is shown in Figure 1: the medicine formulas information and the multidimensional medicine attribute data are prepared as the input; medicine attribute data are projected to the low-dimensional space (2D) and pair-wise distances are calculated; medicine formulas data are arranged by our similarity-based layout algorithm and visualized as an icicle plot; shared herbs of each pair of formulas are calculated and visualized as a matrix; and next, colors are designed for herbs using our perceptual-guided, data-driven color-encoding method.

The attributes of an herbal medicine can be written as an M-dimensional (M=23) vector P of binary valued elements:

The M-dimensional space is then dimensionality reduced to 2D with a vector p of real values:

Uniform manifold approximation and projection for dimension reduction (UMAP) [10] is used for its structure preservation ability and computational efficiency. The distance between the herbs is the basis of our subsequent similarity-based layout computation and visualization. We defined the distance d (u, v) between 2 herbs u and v as the L2-norm, that is, Euclidean distance, between their corresponding 2D vectors p_u and p_v, respectively:

The distance between P_u and P_v in the original M-dimensional space is also considered. However, our experiment shows that the difficulty of discriminating herbs based on the distance with P is higher than that with the projected vectors p, and the resulting visualization based on P is more difficult to compare and comes with more visual clutter.

A co-occurrence matrix view of formulas is included to complement the icicle plot for comparing formulas that are far apart, for example, having different principal herbs. The benefit of using a matrix view is that all formulas’ complete pair-wise intersection information can be effectively represented and easily identified.

As shown in Figure 4, the matrix contains formulas as rows and columns and the number of shared herbs as the element value. With a sequential color map, this view allows the user to quickly examine the overlapping information of each formula against all others by focusing on a row or column. In addition, the color encoding effectively draws the attention of the user to formulas with the highest number of shared herbs: in this case, Zuoguiwan (左归丸) and Youguiwan (右归丸) as highlighted in red in Figure 4.

Our visualization method supports interactive exploration within the formula view, the matrix view, and the herb view. Brushing and linking enables connections between these 3 views (requirement 4). In the formula view, the names of all formulas are shown whenever the mouse hovers over an herb, as shown in Figure 8A. The matrix view highlights the corresponding formulas when the mouse hovers over an element. In the herb view, a lasso tool allows users to flexibly select the herbs of interest. All potential formulas are shown as text in the scatterplot of the herb view (Figure 8C). Representative herbs can be assigned and updated through selection boxes on top of the herb view (Figure 9). These user interactions are easy to use and intuitive for users who are not familiar with interactive visualization. Therefore, requirement 6 is satisfied.

Brushing and linking enables visual connections between the formula view and the herb view interactively. All herbs are highlighted in the herb view with enlarged size (Figure 8B) if any formula is selected in the formula view (Figure 8A). Conversely, whenever any herb is selected in the herb view (Figure 8C), the formula view is updated, as shown in Figure 8D. Here, all selected formulas are highlighted with blue solid lines, and formulas containing the selected herbs are highlighted with red dashed lines. As a result, brushing and linking helps enhance the understanding of users regarding the composition of herbs in formulas (requirement 5).

The proposed method was implemented as a web-based visual analysis tool, as shown in Figure 9. Data processing procedures were performed in Python aided by the “umap” package for dimensionality reduction, the “scipy” package for RBF interpolation, and the “color” package for color space transformations. Visualization and user interactions were realized in JavaScript aided by the “D3” package, and the communication between Python and JavaScript components is achieved using the “eel” package.

The evaluation of our method was performed as a usability study with the analysis of 2 representative use cases—tonic and heat-clearing formulas—by 2 TCM experts (SP and XH). They were asked to analyze the formulas using the web-based tool with think-aloud protocol analysis and provide feedback after the session. Both experts were systematically trained in TCM and obtained clinical degrees and certificates in TCM. One has obtained a doctoral degree in TCM (SP), whereas the other has been working in clinical for over 9 years (XH). Both experts have ≥14 years of expertise in TCM.

After introducing our method to the participants, they were asked to explore the medicine formulas data using our visualization tool, whereas the observer observed and talked to the participants. Afterward, they were asked to provide further feedback on the method. Visualizations of the 2 use cases presented to the TCM experts, as in the web-based tool, are shown in Figure 9.

The tonic formulas (Figure 9A) contained 20 formulas and 58 herbs with 17 principal herbs and a median of 1 principal herb per formula. The median number of herbs per formula was 7.5, with a minimum of 2 and maximum of 15. The average number of shared herbs in a pair of formulas was 1.09 (SD 1.22).

The heat-clearing formulas (Figure 9B) contained 26 formulas and 73 herbs with 25 principal herbs with a median of 1 principal herb per formula. The median number of herbs per formula was 6.5, with a minimum of 2 and maximum of 17. The average number of shared herbs between a pair of formulas was 0.98 (SD 1.24).

Expert PS started the analysis by looking at the overall distribution of herbs and used her knowledge to assign representative herbs for each herb category listed in Figure 6. The resulting continuous 2D colormaps show that the center of the attribute space of tonic formulas is red (Figure 7A), whereas heat-clearing formulas have the center of their space as green and black (Figure 7B). These results indicate the different properties of tonic and heat-clearing formulas and are in line with related TCM concepts.

In the icicle plot of tonic formulas (Figure 10, right), it is easily seen that 2 adjacent columns are similar: the Bazhentang (八珍汤) contains the Sijunzitang (四君子汤) as highlighted in the yellow box. The TCM expert then analyzed the differences between these 2 formulas. She used the lasso tool in the herb view to select 4 other herbs in Bazhentang, as shown in Figure 10 (left). The text below the scatterplot shows that formulas containing these herbs are Bazhentang and Siwutang (四物汤). These 2 formulas were selected with red dashed lines, and the selected herbs are highlighted with solid blue lines in the formula view (Figure 10, right). A close examination showed that the lasso-selected herbs form Siwutang. Moreover, it can be seen that Bazhentang is the combination of Sijunzitang and Siwutang.

In the matrix view (Figure 9A, right), most formulas have overlapping herbal herbs with Sijunzitang (四君子汤) and Bazhentang (八珍汤), suggesting that tonic formulas are built on the herb composition of these 2 formulas.

It is known that the main role of Sijunzitang or Bazhentang is “invigorating Qi and blood.” The understanding of Qi and blood in TCM is the basic substance of the human body, which can reflect the importance of all supplements to Qi and blood in the matrix view. Yin and Yang are 2 interdependent, opposite, complementary, and exchangeable aspects of nature. Qi is Yang (阳, positive), blood is Yin (阴, negative), and Qi and blood are dependent. TCM physicians usually prescribe for diseases in which Qi and blood deviate from balance. The expert considered that this visualization is suitable for beginners to pay attention to the “Qi and blood” supplement for tonic formulas.

The analysis of heat-clearing formulas is shown in Figure 11. TCM expert XH was interested in Sanhuang (3 yellow herbs, 三黄): Huanglian (rhizome of Chinese Goldthread, 黄连), Huangqin (root of Membranous Milkvetch, 黄芩), and Huangbo (Phellodendron bark, 黄柏), which is a commonly used herb combination for clearing heat and detoxification in TCM. The 3 herbs were relatively close in the herb view (Figure 11, left), and the expert used a lasso to select them. Both Huanglian-jiedutang (黄连解毒汤) and Danggui-liuhuangtang (当归六黄汤) contain Sanhuang as suggested by the following text. The expert further examined the formula view (Figure 11, right), where these 2 formulas were highlighted. According to the herb attributes, the function of Huanglian-jiedutang is to clear heat and detoxify. Although the composition of Danggui-liuhuangtang contains tonic herbs, meaning that in addition to clearing heat and detoxification, it also has the effect of nourishing Yin (滋阴). Unlike the tonic formulas, not many overlaps are seen in the matrix view (Figure 9B, right). Most formulas have overlapping herbal medicines with Qingwenbaiduyin (清瘟败毒饮), which have the function of clearing heat and detoxification. This can be a reminder for beginners to pay attention to the relationship between this formula and other formulas in the heat-clearing formulas.

Overall, both experts believe that our method can clearly disassemble complex formulas and assist in the memorization of their functionalities. The interactive visual analysis process is new to them and is helpful in enhancing their understanding of formula composition theories by making and testing their own hypotheses. They believe that the color encoding of herbs allows TCM students and beginners to understand the effect of herbs more intuitively and facilitate memorization. Beginners have difficulty understanding the similarities and differences between multiple similar formulas. With the lasso tool, beginners can test multiple herb combinations to better understand the similarities and differences between formulas and, therefore, better understand an actual prescription. In addition, they consider brushing and linking to be a beginner-friendly way to understand the relationships between herbs and formulas. Both experts made positive comments on the coloring of herbs. For example, Danggui (root of Chinese Angelica, 当归) is a blood tonic herb and corresponds to red. On the other hand, Shigao (Gypsum, 石膏) works on the lungs and is colored white.

The experts suggest that in addition to assisting the learning of TCM formulas for beginners, the method can be extended to facilitate the learning of actual treatment plans for TCM physicians. The TCM theory system includes the process of “theory, method, formula, and herb,” and a treatment plan with prescriptions is performed to assess the effectiveness of formulas. The experts suggest supporting multiple lassos as future work to facilitate the building-up of a prescription by adding herbs from an initial known set of herbs to learn actual treatment plans.

Our new visualization method could effectively reveal the compositional principle of medicine formulas and assist in the learning of TCM formula composition theories. The proposed method can effectively visualize complex TCM formulas and multidimensional herb attribute information. The joint analysis of medicine formulas and corresponding herbs is possible with user interactions and brushing and linking between multiple views within our web-based tool.

Our method does not directly support the visualization of overlaps of ≥2 medicine formulas, that is, intersections of ≥2 sets. However, such information can be implicitly gained by visual searching in the medicine formula view and by interactively selecting herbs of interest that would highlight all formulas containing shared herbs.

Another limitation is that the dimensional reduction view does not explicitly show multidimensional properties but rather the relative distances between herbs. This could be addressed using additional multidimensional visualization techniques, such as parallel coordinates.

In the future, we would like to further enhance the comparison capability of our method. For example, we could support comparing multiple formulas that are not adjacent and apply set visualization techniques to show the correspondence of medicines and formulas directly in the herb view.

Moreover, we would like to apply their method to analyze more groups of formulas and TCM prescriptions in a clinical setting to assist TCM students and doctors to enhance their understanding of formula composition theories and improve their practice.

We introduced a visualization method for TCM formulas. The requirements and design choices of our method are made through a close collaboration between visualization and TCM experts in an iterative, quick-prototyping fashion. Our method supports interactive visualization of medicine formulas with a similarity-based layout complemented by a matrix view of shared herbs by formulas, and multidimensional attribute data of herbs are visualized using a dimensionality-reduction method. The colors of visual elements are assigned with a perceptual-guided, data-driven color-encoding method that achieves perceptual uniformity and reflects TCM concepts. The web-based tool that implements our method supports the interactive analysis and comparison of medicine formulas and corresponding herbs with brushing and linking between different views. The usability study of our method with TCM experts demonstrated the effectiveness of our method for joint TCM formula composition and herb property analysis. Further feedback from experts suggests that our method has potential for educating TCM formula composition theories and modernizing TCM inheritance methods.