This study succeeds a mapping study that clustered the AMR field into 88 topics (Luz et al., 2021). The map was generated by assessing the entire body of AMR literature available on PubMed between 1999 and 2018 (152,780 articles). Using a machine learning algorithm (STM), topics were identified based on the title and abstract text (Roberts et al., 2019). The present study used all articles of three of the identified topics: stewardship (n = 3,383 articles), infection control (n = 1,687 articles), and institutional surveillance (n = 2,176 articles). These three topics reflect the core components of an integrated, comprehensive stewardship concept in institutional healthcare (Dik et al., 2016).

For each topic, a sample of 60 articles that contained at least one data visualization was randomly drawn. Data visualization was defined as the graphical representation of quantitative data. Geographical maps and flowcharts were excluded, as geographical data have distinct visual characteristics and challenges beyond the scope of this study (see e.g., Wang et al., 2017; Singleton and Arribas-Bel, 2021). From the sampled articles, one visualization per article was randomly sampled resulting in 180 data visualizations. The study design is shown in Figure 2. To analyze inter-rater reliability, 10 randomly picked data visualizations per topic were analyzed in duplicate, and the joint probability of agreement was calculated by dividing the number of agreements per categorical assessment form question (i.e., visual characteristics described in section “Data Visualization Analysis”) by the total number of assessments (O’Connor and Joffe, 2020).

The resulting 150 data visualizations (Supplementary Material S1) were analyzed using the nomenclature and categorization by Munzner adapted for this study (Munzner, 2014). This approach dissected data visualizations into visual characteristics:

In addition, data visualizations were labeled with the visualization type used (e.g., bar chart, line chart, scatter plot, etc.) and the use of faceting (multiple linked visualizations in a design grid). Each visualization was assessed upon its interpretability without additional text (yes, if interpretable without additional information; partially, if a description was given in a caption; not at all, if a description was absent or only available in the article text).

Visualization quality was captured by rating the first and last impression during the analysis process on a scale form 1 (poor) to 5 (good). The choice of the visualization type given the underlying data was rated on a scale from 1 (poor) to 5 (good). In addition, free, written text was recorded to capture comments and remarks about the data visualization.

A structured assessment form (Supplementary Material S2) was developed comprising all the above-mentioned elements. The form was discussed within a multidisciplinary team of data-visualization and AMR experts. The assessment form was applied to each data visualization in a two reviewer (JK and CL) process. First, the assessment form was used for training the analysis process with 10 data visualizations not part of the final study data. Next, each reviewer analyzed 50% of the study data visualizations followed by a re-review through the other researcher. Consensus was reached through discussion if the first assessment differed.

Descriptive statistics were calculated for visualization type, number of attributes, faceting, rating, and visualization type choice. Attributes were analyzed for pairwise co-occurrence and presented if a combination occurred more than twice in total.

The visual dictionary was created based on the visual vocabulary (stewardship-related content) and visual design space (characteristics used to design the visualization). The vocabulary was analyzed by identifying attributes and grouping the attribute names using inductive coding. Next, quantitative analyses of visual characteristics (channel, marks, etc.) were performed stratified per attribute, thereby adding the visual design space to the vocabulary. Linking the vocabulary and design space enabled the creation of a visual dictionary to help identify attributes (e.g., resistance) with associated channels (e.g., points and lines on a common scale).

Comments about the visualizations were coded in Microsoft Excel by two researchers (CL and JK). An open coding round was followed by axial coding to discover related concepts in the sub-codes. Differences were discussed until consensus was reached, which increased the internal validity (Patton, 1999). Next to improvements, CL and JK coded remarks about chart junk (i.e., the unnecessary and/or redundant use of visualization embellishments) (Tufte, 2001).

In total, 48 different attributes were coded. The 10 most used attributes were time (n = 69, 46.0%), setting (n = 43, 28.7%), antimicrobial consumption (n = 32, 21.3%), resistance (n = 31, 20.1%), antimicrobials (n = 27, 18.0%), percentage (n = 26, 17.3%), count (n = 24, 16.0%), incidence (n = 24, 16.0%), numeric value (n = 20, 13.3%), and bacteria (n = 12, 8.0%). Attributes could be grouped into objects (e.g., bacteria) and measurements (e.g., percentage). However, the following analysis focuses on attribute combinations and attributes are thus kept ungrouped.

Attributes showed different co-occurrence patterns (Figure 3). The 10 most frequent combinations were time and antimicrobial consumption (n = 21), time and incidence (n = 18), antimicrobial consumption and antimicrobials (n = 12), antimicrobials and resistance (n = 12), time and resistance (n = 12), time and antimicrobials (n = 11), antimicrobial consumption and setting (n = 10), resistance and setting (n = 9), time and setting (n = 9), and percentage and setting (n = 8).

To illustrate problems in data visualization, we designed a visualization that exhibits several of the identified problems based on simulated data (Figure 4). Figure 5 proposes an alternative to Figure 4 where the identified problems were avoided. Of note, data such as the simulated data in these figures can be visualized in many different ways, depending on the underlying research questions.

Figure 6 summarizes the results of this study and presents the visual dictionary of stewardship. In addition, it provides a set of recommendations to avoid the most common problems in data visualizations as identified in this study.

With the systematic analysis of the visual domain context of stewardship we revealed common practices and identified problems with current implemented visualizations. In general, the use of data visualizations for communicating data is highly encouraged. It greatly supports the interpretation, memorization, and communication of insights and knowledge gained from data. In this study, we identified 14 different visualization types used in the visual domain context of the field. However, more than 80% of all visualizations used classical (stacked) bar or line charts; quite homogenous design choices. We argue that the visualization type choice is based on tradition and habits as a systematic approach to data visualization in the field was missing until now (Salinas et al., 2020). A lack of awareness and knowledge about data visualization design and design alternatives might lead to suboptimal data visualizations. Examples from our findings were the use of less effective visual channels, suboptimal plot types for the presented data, or mismatches in color choices for different data types. Similar visualization pitfalls were identified in studies focusing on common visualization pitfalls in multidisciplinary research related to visual representations and for environmental data, emphasizing instances where data visualization creators require more support in visualization design choices (Kelleher and Wagener, 2011; Bresciani and Eppler, 2015). Now that we revealed common data visualization practices in the visual dictionary for stewardship by linking often used attributes (i.e., content) and associated design choices (e.g., visualization type or marks), visualization creators in the field can match their visualizations with the audience’s visual expectations and habits.

However, given the wide variety of data in the field and the increased complexity that big data will add (in terms of volume, velocity, variety, veracity, validity, volatility, and value), more “visual variability” might be expected and even needed in the future (Khan et al., 2014; Gotz and Borland, 2016; Galletta et al., 2019). A first step toward visual variability is informing and teaching visualization creators and users about data visualization design alternatives. We see a clear role here for data visualization experts and software developers to cocreate open-source/access tools that support visualization creators in their visualization choices (e.g., reminders for adding labels and legends, suggestions for optimal color schemes, warnings in case of chart junk). Our results and findings from similar studies in other fields can support them in doing so by providing an overview of common data visualization practices in the field, including dos and don’ts (Crisan et al., 2019; Nusrat et al., 2019).

Of note, academic journals play an important part in this process by providing the platform for data visualizations and should be encouraged to promote high quality data visualization practices. Furthermore, it could be worth considering standardizing data visualization for often used data types and contents in the field, given the prominent patterns in the visual dictionary (e.g., time series were part of 43.3% of all studied visualizations) within the large variety of content (48 different attributes such as antimicrobials, bacteria, or time) observed in this study. For time series specifically, an overview of data visualization methods exist (Fang et al., 2020), and similar standardizing initiatives can be found in the AMR field [e.g., European Committee of Antimicrobial Susceptibility Testing (EUCAST; Brown et al., 2015)] and other fields [e.g., the Intergovernmental Panel on Climate Change (IPCC) and standardized medical data visualization based on the ISO13606 data model (Kopanitsa et al., 2015; Gomis and Pidcock, 2018)]. This could help ensure high quality data visualizations for reliable insights in AMR/stewardship related data.

In the light of growing complexities and increasing data volumes, genomic data and their visualization play a special role in the field. Although genomic data visualizations were included in this study, most visualizations were simple dendograms and phylogenetic trees. As with the a priori excluded geographical data, these complex data require dedicated research and visualization techniques which are provided in great detail by others (see e.g., Crisan et al., 2019; Nusrat et al., 2019). An additional important aspect for high quality data visualization in the stewardship and AMR data field is the visualization of uncertainties. The visualization of uncertainties was not within the scope of this study and further research into the optimal display of uncertainty is highly encourage. For more information readers are referred to the work of others (Wilke, 2019; Korporaal et al., 2020; Padilla et al., 2020).

Studying the visual domain context is as important as studying data visualizations themselves. The importance of assessing visual habits and perceptions in data visualization has been demonstrated before in other medical fields revealing that personal preferences and visualization habits might not always match with novel data visualization approaches and recommendations (Backonja et al., 2018; Aung et al., 2019b). Aung et al. (2019b) published an exemplary study in the field of reproductive, maternal, new-born and child health, focusing on data visualization interpretation capacity and preferences in their target audience by combining interviews on interpretability and card-sorting of preferred visualizations. Thus, for data visualization in general we strongly believe that incorporating best practices is essential, but advocate that these should be carefully balanced with visual habits and expectations in the field and the message to be conveyed. Additionally, research is needed to better understand how data visualizations in general impact the viewers in terms of changes in opinions or attitudes that direct decision-making or behavior changes (Pandey et al., 2014).

In future research special attention should be paid to matching the visual dictionary and the context in which the visualization will be used in terms of users, their tasks and current practices (e.g., studying questions like “How do current visualizations support to do current tasks?” and “What visualizations would the target audience like to see?”) (Lam et al., 2012). This also includes color-blindness considerations, as extensively studied by others (Ahmed et al., 2020; Crameri et al., 2020; Burggraaff et al., 2021). We see a clear parallel with user-centered eHealth design that emphasizes the need for a holistic understanding of the interrelations between technology, people, and their context (van Gemert-Pijnen et al., 2018). Both qualitative (e.g., interviews) and quantitative (e.g., eye-tracking in current data visualizations) study designs can contribute to such a holistic understanding, which in turn can inform or improve the design of visualizations (or eHealth) in terms of required content, functionalities, and usability (Keizer et al., 2020). Therefore, complementing research on data visualizations, as the current study and many other studies do, with research that primarily focuses on the interaction between people, their context and how data visualizations can support them, is needed (Lam et al., 2012).

This study has several limitations. Despite sampling from a comprehensive set of articles that cover the stewardship field, only a limited number of data visualizations were included. Moreover, only data visualizations from scientific publications and not from other sources relevant to stewardship data visualization creators [e.g., data systems used in practice (Huber et al., 2018; Sedrakyan et al., 2019) and AMR policy reports (Anderson et al., 2019; World Health Organization, 2020a)] were included. Therefore, we missed data visualizations for other stewardship content, making our findings potentially more applicable to stewardship researchers than healthcare professionals. However, the observed homogeneity of data visualization types suggests saturation regarding the visual design space for stewardship. Subsequent research into the visual domain context of stewardship should include these additional sources to ensure a more comprehensive picture for healthcare professionals. Even though the extracted data visualizations were systematically analyzed using a structured assessment form based on existing data visualization nomenclature and categorization (Munzner, 2014), the analyses relied on the subjective interpretation and rating by the coding researchers. Several measures were taken to validate our findings, including discussing the assessment form and results within a multidisciplinary team of data-visualization and AMR experts, analyzing the interrater-reliability, and comparing our findings to other data visualization studies. Our study is one of the first empirical studies that explores the use of data visualization in stewardship, thereby adding to the few review studies providing primers for data visualization recommendations and best practices in the stewardship field (Carroll et al., 2014; Salinas et al., 2020). Furthermore, our structured assessment approach can be applied in future studies in the broader One Health field to unravel the visual dictionary of the fields of human, animal, and ecosystem health, considering interdisciplinary differences in data and data visualizations and their integration and interpretation (Lapinski et al., 2015; European Centre for Disease Prevention and Control [ECDC], European Food Safety Authority [EFSA], and European Medicines Agency [EMA], 2017; EFSA, 2021).