Information Studies: BIS Faculty of Science
Personalized visual analytics
in a health care environment
Author:
Roy Dolleman
roy.dolleman@student.uva.nl 10572759
Supervisor: dhr. prof. dr. Marcel Worring m.worring@uva.nl
Version: final April 6, 2016
Roy Dolleman
Faculty of Science, University of Amsterdam
Abstract
Because the amount of data is increasing, there is a necessity for managing data more effectively. Better insights can be obtained if this data is managed more effectively. The insights can be further improved with personalized visual analytics, which tailors a dash-board to a specific user. Research has been done in this field of study, but most papers are about personal contexts. The research in this thesis is done in a professional context (health care). A theory is formed in this thesis, using user-profiles from other papers. Users can be selected for these profiles and the dashboard will add specific functions ac-cordingly. This theory is then checked by interviewing professionals from a health care environment. Based on findings from these interviews conclusions can be drawn. The conclusion describes which parts of a visual analytics process can be personalized and also describes the user-profiles that can benefit from personalization.
Keywords. Health care, health care environment, analytics, IT, visual analytics, visual
analytics process, personalized visual analytics, information visualization, medical roles, personalized information, dashboard, personalized dashboard, up dashboards, mock-ups;
1
Introduction
The amount of data in the world is rapidly increasing. This trend is, among other things, made possible by the by the improvement of storage devices. Most of the data is, however, stored without filtering and refinement for later use. This causes organizations in virtually every branch of industry or business to create and store vast amounts of data without using it (D. Keim et al., 2008). The trend of digitizing hard-copy documents also helps creating more data. The estimated amount of data in 2012 was 2,837 exabytes, which is 2,97 billion terabytes, and predictions are that this will double every two years (McLellan, 2014). This trend of a growing amount of data also holds for health care organizations. Managing and analyzing this complex and huge amount of data (’big data’ (Murdoch & Detsky, 2013)) however, is an enormous challenge for organizations.
This challenging task can be done effectively by using visual analytics. Visual analyt-ics is described by D. Keim et al. (2008) as: “Visual analytanalyt-ics combines automated analysis
techniques with interactive visualizations for an effective understanding, reasoning and de-cision making on the basis of very large and complex data sets.”. Visual analytics has great
potential to create benefits for the (health care) organization by analyzing, filtering and illustrating data (Caban & Gotz, 2012), but can also lead to higher costs if it is not used correctly (Meniuc, 2014). When data is made usable it enables the health care organization
to obtain two kinds of insights: insights about patients (clinical analysis) or insights about the organization itself (support of the organization) (Raghupathi & Raghupathi, 2014). Both these kinds of insights have in common that they are generally used to get generic insights for multiple persons (Huang et al., 2014).
The personalization of visual analytics could possibly create more benefits for a (health care) organization. This will especially be the case in a medical setting because there are lives at stake. The scientific problem of personalized visual analytics, however, is that not much research has been done in this field of study (Huang et al., 2014). This makes it difficult for system designers (and others that can benefit from personalized vi-sual analytics) to get know-how and apply the knowledge about personalizing in practice. Consequently, it is important to investigate these problems and create a starting point for further research with personalization of visual analytics. With this problem in mind the main research question can be formulated.
• How can visual analytics be personalized to get better insights for medical profession-als in a health care environment?
To be able to answer the main research question, this question is divided in the following sub-questions:
1. Which parts are there in a visual analytics process?
2. Which parts of the visual analytics process can be personalized and how can this be achieved?
3. What types of users can benefit from personalization and in what way can they do so?
This research question could lead to conclusions that help developers in creating personalized dashboards that can be used in the process of visual analytics in a health care environment. The targeted conclusion of this paper is a best practice for developing personalized visual analytics in a medical setting.
2
Method
2.1 Structure
The thesis is structured in the following way. The thesis starts with a theoretical outline to define the context of the thesis and describes the terms upon which later chapters are based. In this theoretical outline the term personalized visual analytics will be defined. Based on this theoretical information, mock-ups of personalized dashboards are created. These (mock-up) dashboards will be used during the interviews with experts from the field of health care. This will eventually result in best practices for developing personalized visual analytics in a medical setting. The structure of this thesis can also be seen in Figure 1.
2.2 Method
As mentioned before, the main focus of this thesis will be personalized dashboards, which is only a part of the visual analytics process (see Figure 3.5). This is done to be able to
Figure 1 . Thesis structure
do qualitative research (interviewing experts from the field of study). Another reason why only a part of the process is chosen is due to time limitations. The personalization aspect is first defined in categories: user profiles. This option was chosen because this thesis is not created for a specific organization. It is therefore not possible to specify specific roles and personalization needs. These user profiles will then be linked to the different functionalities of dashboards. These relations will then be validated through a field research (interviews). This will result in a best practice for developing dashboards (with personalized functionalities).
2.3 Relationship with collaborative visual analytics
Multiple roles can use an analytics systems in a real-life medical setting. Daniel Tor will be researching the topic of collaboration between these roles with help from collaborative visual analytics. The same base dashboards will be used, but will be diverged when creating mock-ups for the research. Because of time limitations the researches will not be converged after they have been completed.
3
Theoretical outline and related work
This chapter aims to define the context of the thesis and describes the terms upon which later chapters are based. First health care, information needs in health care and visual analytics in health care are defined. Next, information visualization and visual analytics (including the process) are described. Lastly the medical roles and personalized visual analytics and its design are defined. This theoretical outline will allow the reader to get a general understanding before moving on to the next chapters.
3.1 Health care
Health care refers to diagnosis, treatment (of illness or injury), disease prevention, and other physical and mental problems of persons. Health care can be divided in primary, secondary, tertiary and quaternary care. Primary health care refers to health care professionals who act as first contact point for patients who need health care (Appendix: Definition of Terms, 2009). Secondary health care refers to health care professionals who are specialists and usually do not have the first contact with the patients (U.S. Medicare Health Insurance
Information, 2009). Tertiary health care refers to the health care professionals who are
even more specialized and usually do advanced medical investigation (U.S. Medicare Health
Insurance Information, 2009). Quaternary health care refers to highly specialized health
There are different types of health care. A distinction can be made between care and cure (Chaufan, Hollister, Nazareno, & Fox, 2012). Care is the long-term approach where the goal is to minimize the effects of the medical condition of the patient. Cure is the short-term approach where the goal is to resolve the medical condition of the patient.
3.2 Information needs in health care
Medical professionals need information to be able to do their job. Gorman (1995) has proposed the following categorization for the information a medical professional needs, using work that went before and his own research. The categories are:
• On particular patients
• Data on health and sickness within the local population • Medical knowledge
• Local information on doctors available for referral • Information on local social influences and expectations
• Information on scientific, political, legal, social, management, and ethical changes that will affect both how medicine is practiced in a society and how doctors will interact with individual patients
There are some problems with information in health care. Research has been done about the supply of information to medical professional and the problems that medical professionals experience with information (Williamson, German, Weiss, Skinner, & Bowes, 1989). The researchers used a survey and discovered that nearly two thirds of the medical professionals think that the current volume of scientific information was unmanageable. More than a third of the medical professionals stated that “most physicians find the effort to get information from the literature to be a major problem.”. Even though the research is conducted 20 years ago, it can be assumed that the results will be the same nowadays because the volume of information has only grown bigger.
Another research about information is conducted by Smith (1996). This research showed that medical professionals use up to two million pieces of information to manage patients and that textbooks, journals, and other existing information tools are not adequate (for different reasons) for use in a medical environment. The research showed that medical professionals need information sources that provides relevant, valid material that can be accessed quickly and with minimal effort.
3.3 Information visualization
Many people confuse information visualization and visual analytics because they are related and have some overlap between them, but there is an important difference between the two. Traditional information visualization does not use advanced data analysis algorithms and also does not contain the analytic task. That’s why visual analytics is more than information visualization and can be seen as an approach to make decisions, combine visualizations, human factors and data analysis. (D. Keim et al., 2008)
Wong and Thomas (2004) say that: “Visual analytics is an outgrowth of the fields
of scientific and information visualization but includes technologies from many other fields, including knowledge management, statistical analysis, cognitive science, decision science,
and many more.”. From this definition can be concluded that visual analytics includes
information visualization but adds analytical steps to it.
3.4 Visual analytics
One of the problems with visual analytics is that it can be described in different ways. Thomas and Cook (2005) describe visual analytics as: “Visual analytics is the science
of analytical reasoning facilitated by interactive visual interfaces.”. This means that visual analytics tools and techniques can be used to get insights from (massive, dynamic, ambiguous, and often conflicting) data. Visual analytics is also used to find expected and unexpected results. Furthermore, visual analytics can also be used to create defensible and understandable assessments which then can be communicated effectively for action.
Visual analytics has multiple focus areas, which makes it a multidisciplinary field (Thomas & Cook, 2005):
• Analytical reasoning techniques, which allows users to get insights (supporting assess-ment, planning, and decision making).
• Visual representations and interaction techniques, which allows the user to understand large data sets at once by making advantage of the human eye’s broad bandwidth pathway into the mind.
• Data representations and transformations, which supports analysis and visualizations by converting conflicting and dynamic data.
• Supporting techniques (production, presentation, and dissemination), to communicate the information from the analysis to a variety of audiences.
Many others have tried to describe visual analytics. D. Keim et al. (2008) describe visual analytics as: “Visual analytics combines automated analysis techniques with
inter-active visualizations for an effective understanding, reasoning and decision making on the basis of very large and complex data sets.”.
Visual analytics supports the user to (D. Keim et al., 2008):
• Get insights from (massive, dynamic, ambiguous, and often conflicting) data. • Detect expected and unexpected results.
• Provide timely, defensible, and understandable assessments. • Communicate these assessments to effectively take action.
Because data sets are growing, visual analytics solutions are also growing in complex-ity and size. This effects the design decisions at the beginning of a project of creating a visual analytics solution (Meniuc, 2014).
From these descriptions it can be concluded that visual analytics is a multidisciplinary approach and that visual analytics aims at processing high-volume data, discovering pat-terns, deriving insights from large and complex data sets and communicate the findings. Visual analytics uses and takes advantage of various related research areas (visualization, data mining, data management, data fusion, statistics, and cognition science) (D. Keim & Zhang, 2011), (Meniuc, 2014).
3.5 Visual analytic process models
Before the data can be analyzed, the data needs to be processed. This process of getting from data to insights (visual analytics) is described with different models. One of these models is developed by Pirolli and Card (2005) (see Figure 2). Two major activities are performed in this model:
• Foraging loop (External Data Sources, Shoebox and Evidence File): this loop is con-cerned with data foraging and structuring, the information space is organized and processed (this includes the steps of collecting data, selecting the relevant information and organizing the information to make it understandable).
• Sensemaking loop (Schema, Hypothesis and Presentation): this loop is concerned with analyzing, the data is interpreted and findings are communicated (this includes the steps of schematizing the information, formulation of hypotheses and selecting the proper presentation form).
The important activities of this model for this thesis are in the sensemaking loop. One of these activities is schematize, in this activity the (personalized) schema’s are developed to get insights from the data. Schema is not the only important activity because the activity presentation is also a form of visualization. The sensemaking loop can be seen in Figure 2.
Figure 2 . Visual analytics process of Pirolli and Card (2005)
Another model is developed by D. Keim et al. (2008) and D. A. Keim, Kohlhammer, Ellis, and Mansmann (2010) (see Figure 3). This model of a visual analytics process consists of the following elements:
• Data: this first step (sometimes referred to as sources) pre-processes and transforms the data to derive different representations for further exploration. Other tasks in this step are data cleaning normalization, grouping, or integration of heterogeneous data sources. After this step the analyst can choose to either apply automated analysis
(through data mining) to create models or hypotheses; or map the data to reach visual representations.
• Visualizations: in this step the visualization is generated. The visualization can be created from the data or from the hypotheses.
• Hypotheses: in this step the hypothesis is generated. Hypotheses can be generated from data or from visualizations.
• Insights: in this step insights are gained from the visualization and/or hypotheses steps. These insights always have five characteristics: complex, deep, qualitative, unexpected and relevant (North, 2006).
Figure 3 . Visual analytics process of Keim (D. Keim et al., 2008)
These models have some similarities, but are not completely the same. The models both have a data collection step, the (shoebox step) foraging loop in the model of Pirolli and Card corresponds with the data step in Keim’s model. Both models have a step where visualizations/schema’s and hypotheses are created. Another similarity is that the models assume that the analyst can use visualization at any step in the process (Meniuc, 2014).
According to Meniuc (2014), the main difference between the two models, is the presentation step, which is not completely missing in Keim’s model but it’s not described explicitly as in the model of Pirolli and Card. Another difference is that the model of Pirolli and Card assume that the analysis process is organized and straightforward, in contrast to Keim’s model.
For this thesis Keim’s model will be used because this model explicitly names the user interactions in the visualization step. This interaction with the user is necessary for
the mock-ups (see section 4.2).
3.6 Visual analytics in health care
History proves that large quantities of data are generated in the health care industry (driven by record keeping, compliance and regulatory requirements, and patient care) (Raghupathi & Raghupathi, 2014). This data shows great potential to be used for visual analytics in practice.
There are however three major problems with visual analytics in health care described by Caban and Gotz (2012): "(a) physicians and clinical practitioners are faced with the
chal-lenging task of analyzing large amount of unstructured, multi-modal, and longitudinal data to effectively diagnose and monitor the progression of a particular disease; (b) patients are confronted with the difficult task of understanding the correlations between many clinical values relevant to their health; and (c) healthcare organizations are faced with the prob-lem of improving the overall operational efficiency and performance of the institution while maintaining the quality of patient care and safety.". This paper concerns the challenging
task for the practitioner of analyzing the data. This can be done by personalizing the dash-board and the mock-ups of the dashdash-boards will therefore be personalized for the medical professional.
3.7 Medical roles
There are different medical roles in health care. Some general roles found are (Jones, Bellomo, DeVita, et al., 2009) , (McCloskey & Johnston, 1990):
• Doctor: medical professional with a general role, almost always the team leader • Nurse: medical professional with a (general) supporting role, rarely the team leader • Therapist: medical professional with a specialist role, rarely the team leader
The main difference between these roles is the generality, from a general role until an expert role. Wagner (2000) defined the roles in a medical team more specifically, but also didn’t mention the roles in the supporting processes (Visser, 2010). In addition to these roles there are also the roles in the supporting processes of a health care environment. These roles can for example be managing the human, financial or material resources and also can benefit from visual analytics (Dutrée, de Man, & Jebbink, 2002), (Amelsvoort & Jacobs, 2005).
3.8 Personalized visual analytics
Personalized visual analytics is the science of analytical reasoning facilitated by visual rep-resentations made for a specific user (Huang et al., 2014). This personalizing of the vi-sualization can be user driven, which allows the user to directly invoke and support the visualization (Sáinz Ibáñez, 2002). There is also automatic visualization, which means that the system identifies the input (user) and tailors the visualization to this user (Sáinz Ibáñez, 2002). Personalized visualizations can also be used in medical environments. This means that each analyst (medical professional) sees a different visualization which can be used for analysing data. This form of visual analytics can be seen as a next step of regular visual
analytics because they have the same use, but in an extended way (Huang et al., 2014). Huang et al. (2014) has described some challenges when visual analytics is personalized. These challenges are:
1. A personalized visual analytic tool has to fit in personal routines and physical envi-ronments
2. The appropriate context for interpreting the data may not be in the form of data that is easily accessible
3. Defining appropriate baselines for comparisons, to gain insights from the data 4. Sharing and privacy of the data, when the data is displayed
5. Most designs are created by system designers who decide which information is pre-sented and how this should be displayed, without considering the unique perspectives of individuals.
6. Adding computer assisted pattern recognition, to find patterns in large amounts of data
7. Evaluation of the personalized visual analytic tool, define the conceptual metric to evaluate
Some of these challenges do not only apply for personalized visual analytics, but also for ’normal’ visual analytics (for example: 6, the computer assisted pattern recognition).
Other interesting papers have been written about personalized visual analytics, but most of these papers are written about personal contexts. A personal context refers to non-professional situations where people have different goals, priorities, role expectations, environments or time and resource budgets than in professional aspects of their lives. The use of personalized visual analytics in a personal setting are getting insights about the as-pects of daily life (for example: residential energy consumption, fitness, personal health, social networks, politics, residential environment, life logging, personal finance, and recy-cling). The difference is that personalized visual analytics in a personal setting is mostly used to satisfy their curiosity, to reminisce about experiences, or to share with others and not to gain deep insights. Another difference is that personalized visual analytics in a per-sonal setting is that the information is about the user, which is generally not the case in the professional setting. (Carpendale, Tory, & Tang, 2014)
There are, however, also some similarities with personalized visual analytics in a personal setting which are interesting for this research. Both are tailored to a specific user and specific information for this user. They also have in common that information is always available.
3.9 Designing (personalized) visual analytics
Designing a dashboard in a visual analytics process is different from designing a personalized dashboard. These dashboards are tailored to the different users. All of the design rules however, can be used when designing these personalized dashboards. Few (2006) described the following 13 design rules:
1. Use 1 main screen to show all information. It is confusing for the user to get insights by scrolling or switching to different screens.
2. Provide adequate context for the data. Without the right context the user cannot get insights from the dashboard.
3. Choose appropriate level of detail and precision. If the information on the dashboard is in too much detail, the user will not use this information to get insights.
4. Express measures directly. It is confusing for the user to get insights from dashboards where the measures are expressed indirectly.
5. Choose appropriate display mechanisms. The user will only get proper insights if the right graph type is selected to show the information in.
6. Use consistent visualization mechanisms. It is confusing for the user to get insights if there is a meaningless variety in graph types.
7. Properly design the display mechanisms. If the dashboard uses poorly designed display mechanisms (many non-data pixels), the user is shown unnecessary information. 8. Encode data accurately. It is misleading and confusing for the user if the data is
inaccurately encoded and this does not help the user to gain insights.
9. Arrange the data in a proper way. If the data is not arranged properly (dashboards must show a a lot of information in a limited amount of space), the user will have difficulty getting the right insights.
10. Use effective highlighting of important data. It is not possible for the user to see (at a glance) which data is important without highlighting this data.
11. Use decoration sparsely. It is confusing for the user to get insights if the screen is filled with useless decoration (the purpose of the screen is not to entertain).
12. Appropriate use of color. If the choices of the colors used are not made thoughtfully, the user can interpret the significance of color difference wrong (which possibly results in a wrong insight).
13. Use appealing visual displays. If the dashboard is unattractive, the user will be inclined to avert the dashboard and this will result in a user that is distracted from getting the right insights.
3.10 Advantages and disadvantages of personalized dashboards
Personalized dashboards have a couple of advantages over ’normal’ dashboards. The ad-vantages and disadad-vantages are written from both the perspective of the company and the perspective of the user. These advantages are:
• Personalization makes the dashboard seem more relevant and meaningful to the user (Hawkins, Kreuter, Resnicow, Fishbein, & Dijkstra, 2008)
• Personalization of the dashboard increases the attention, involvement and motivation of the user (Webb, Simmons, & Brandon, 2005)
• Personalization ensures that the user is able to identify with the dashboard and allows the user to analyze their data based on their preferred style of working (Ritterband, Thorndike, Cox, Kovatchev, & Gonder-Frederick, 2009)
• Personalization allows better communication and service for the user (no unnecessary information on the dashboard), thus it will take less time (cheaper) to analyze the data (Vesanen, 2007), (Lustria, Cortese, Noar, & Glueckauf, 2009)
There are also disadvantages of personalized dashboards. These disadvantages are:
• Personalization requires investments in technology because not every system is capable of personalization (Vesanen, 2007)
• Personalization requires the use of user profiles which consists of saving and using privacy-sensitive data (Deconinck, 2005), (Poelhekke, 2012)
• Personalization is more difficult and thus takes more time to develop ’normal visual analytics’ (which will initially cost the company more) because there is more devel-opment needed than when there is only one default dashboard available (Brusilovsky, Kobsa, & Nejdl, 2007)
• Personalization causes the need for an extra eduction for the users which also will initially cost the company more (Vesanen, 2007)
• Personalization will make the system used more complex because of the extra func-tionalities (Brusilovsky et al., 2007)
4
Personalized visual analytics designs
This chapter aims to define the design choices and selection of design hypotheses upon which the dashboards and experiments are based. First, the user profiles and functionalities are described. Next, the correlations of these user profiles with these functions are explained. Finally, the design choices and selection of design hypotheses are defined.
4.1 User-personalization profiles
User profiles can be used to personalize a dashboard. These user profiles contain the user’s interests. Such user profiles are also used by most websites as recommendation system and are used to recommend a product based on the information in the profile (Brusilovsky et al., 2007). The profile can be created by the user or automatic (see section 3.8). These profiles are grouped in this section to keep it organized.
There are different reasons to personalize a dashboard (Claassen & Van Beurden, 2008), (Mienes, 2005), (Bra, 1998), (Mayhew, 1991), (Ahn, Brusilovsky, Grady, He, & Syn, 2007), (Brusilovsky et al., 2007). These reasons can be:
• Limited visibility, this can be visually impaired or colorblind. When a person is visually impaired the screen is personalized by enlarging everything on the screen and when a person is colorblind some (differences in) colors are visible for that person.
• Location, the location of the person. The location will personalize the screen with regard to the timezone the person is in or the currency of the location of that person. • Time, the time the user visits. The time can be important when personalizing to know what must be shown. This can also be different every hour or for example what part of the day it is (night, morning, afternoon or evening).
• Pattern, pattern of the user. It is possible that the user often has the same pattern. The screen can be personalized on the basis of this pattern. The pattern can be found in logs (or statistics) made during usage of the user.
• Requirements, this can be requirements chosen by the user. This could be important if the user doesn’t want anything to be automatically adjusted on the screen. This can for example be the language chosen (on the basis of the location of the user) on the screen while the user wants another language.
• Role/function, this can be the function of the user (see section 3.7). Every function in a company is interested in different information. This will result in different dashboards for every role in the company.
• Shown data, this can be superfluous data. This can be important for the user if the user thinks there is too much data on the screen. This will simplify the dashboard. • Visualization type, this can be the graph type the user wants to see. This can be
important for the user if the ’normal’ graph doesn’t show something important
4.2 Functionalities and interactions of dashboards
Dashboards have various functionalities. These functionalities enable users to zoom in on interesting parts and details, to customize the content and graphical form, and also to ex-plore large amounts of data. The ultimate dashboard even facilitates a playful experience to engage the user more than when info-graphics or videos are used. Seven key functionalities are described in this paragraph (Yi, ah Kang, Stasko, & Jacko, 2007).
• Select. Select allows the user to mark an interesting item to keep track of it. This can be done when it is difficult to follow an item if the representation is changed or when there are too many data items showed to the user.
• Explore. Explore allows the user to investigate a different subset of data cases. Be-cause of limitations (of the user and/or system) only a limited number of data items are shown at a time and explore allows the user to let new data items enter the view as others are removed.
• Reconfigure. Reconfigure allows the user to change the spatial arrangement of rep-resentations to provide different perspectives. By changing the perspective, the way data items are arranged or the alignment of data items the view reveals hidden char-acteristics of data and the relationships within the data.
• Encode. Encode allows the user to alter the fundamental visual representation of the data. This includes the visual appearance (shape, color, or size) of the data elements. These visual elements are not only important affect pre-attentive cognition but they are also important for users to understand the relationships and distributions of the data items.
• Abstract/Elaborate. Abstract/Elaborate allows the user to adjust the level of abstrac-tion of a data representaabstrac-tion. This means that the user can change the representaabstrac-tion from an overview to more the details (with levels in-between) of data cases.
• Filter. Filter allows the user to change (based on some specific conditions) the pre-sented data items. Only the data items which meet the criteria specified by the user (range or condition) are shown. Data items that do not meet the criteria are hidden or showed differently, but these items are not changed. The hidden or differently shown data items can be recovered when the user resets the criteria because they are unchanged.
• Connect. Connect allows the user to highlight relations between data items and show hidden data items that are relevant to a specified item.
• Other interaction techniques. There are some other interaction techniques that are not unique to information visualization but still have value as useful interactive capa-bilities.
– Undo/redo. Undo/redo allow the user to go backward or forward to pre-existing
system states. This is not only undo and redo, but also reset and history.
– Change configuration. Change configuration allows the user to change the various
configurations and settings of a system.
4.3 User profiles and functionalities
Table 1 shows the relation between the user-personalization profiles and functionali-ties of dashboards. By showing these relations important functionalifunctionali-ties for each user-personalization profile will become clear. Based on this information, design choices for the mock-ups can be made. There are four values possible in the table (+, -, P or N). The sign + indicates a positive correlation (ie. more of characteristic X means more necessity of element Y) and the - sign indicates a negative correlation (ie. more of characteristic X means less necessity of element Y). P is used if there possibly is a relationship but it’s not quantifiable. N is used if there is no relationship.
4.4 Design choices
This section describes the choices made in Table 1. An explanation will be given for every value between a user profile and a functionality.
• Limited visibility is positively correlated with select, encode, abstract, connect and other. A user with limited visibility has more necessity for select (because it will be easier to follow an item), encode (because it’s easier to see when the representation
1. Select 2. Explore 3. Reconfigure 4. Enco de 5. Abstract 6. Filter 7. Connect 8. Other A. Lim. visibility + P P + + - + + B. Location N P - N P + P + C. Time N P - N P + P P D. Pattern + P + + + + P + E. Requirements + P P + P + - + F. Role + + P N + + N + G. Shown Data + + N N + + N -H. Vis. type + - + + P - - N Table 1
Correlations between user profiles and functionalities
of an item can be changed), abstract (because changing the representation can make it easier to see different items), connect (because the connection makes it easier to see which item is related to others), and other (because it may be useful to change a setting of the system such as size of the screen). Limited visibility is possibly positively (explore and reconfigure) and negatively correlated (filter) with the other functions. • Location is positively correlated with filter and other. When the location of a person
is known, there is more necessity for filter (because items can be filtered according to the location), and other (because it may be useful to change a setting of the system such as timezone on the screen). Location is possibly positively (explore, abstract and connect) and negatively correlated (reconfigure) with the other functions.
• Time is positively correlated with filter. When the usage time of a person is known, there is more necessity for filter (because items can be filtered according to the time). Time is possibly positively (explore, abstract, connect and other) and negatively cor-related (reconfigure) with the other functions.
• Pattern is positively correlated with select, reconfigure, encode, abstract, filter and other. When the pattern of a user is known, there is more necessity for select (because items can be selected according to the known pattern of the user), reconfigure (because items can be reconfigured according to the known pattern of the user), encode (because the visual representation can be configured according to the known pattern of the user), abstract (because the level of abstraction can be configured according to the known pattern of the user), filter (because filters can be recommended according to the known pattern of the user), and other (because it may be useful to change a setting of the system according to the known pattern of the user). Pattern is possibly positively (explore and connect) and not negatively correlated with the other functions.
• Requirements is positively correlated with select, encode, filter and other. When the requirements are set by the user, there is more necessity for select (because items can be selected according to the requirement of the user), encode (because the visual representation can be configured according to the requirement of the user), filter (because filters can be recommended according to the requirement of the user) and other (because it may be useful to change a setting of the system according to the requirement of the user). Requirements is possibly positively (explore, reconfigure and abstract) and negatively correlated (connect) with the other functions.
• Role is positively correlated with select, explore, abstract, filter and other. When the role of the user is known, there is more necessity for select (because items can be selected according to the role of the user), explore (because subsets of data cases can be selected according to the role of the user), abstract (because the level of abstraction can be configured according to the role of the user), filter (because filters can be recommended according to the role of the user) and other (because it may be useful to change a setting of the system according to the role of the user). Role is possibly positively (reconfigure) and not negatively correlated with the other functions. • Shown data is positively correlated with select, explore, abstract and filter. When it is
known of the user which data is superfluous, there is more necessity for select (because items can be selected according to the preference of the user), explore (because subsets of data cases can be selected according to the preference of the user), abstract (because the level of abstraction can be configured according to the preference of the user) and filter (because filters can be recommended according to the preference of the user). Shown data is not possibly positively correlated to other functions but is negatively correlated (other) with the other functions.
• Visualization type is positively correlated with select, reconfigure, encode and ab-stract. When it is known of the user which visualization type is preferred, there is more necessity for select (because visualization type can be selected according to the preference of the user), reconfigure (because the visualization type can be reconfig-ured according to the preference of the user) and encode (because the visualization type can be configured according to the preference of the user). Visualization type is possibly positively (abstract) and negatively correlated (explore, filter and connect) with the other functions.
4.5 Selection of testable design hypotheses
Not every design choice of section 4.4 can be tested in a real life setting. This is due to time and tool limitations (the tool does not include everything to build the mock-ups as proposed). The following selection of design choices has been made (these are made gray in Table 1):
• Limited visibility • Location
These design choices have been chosen because these were the most suitable choices for this research. This is due to the fact that for these choices not all detailed information about the existing dashboards was needed. Role could also be an interesting choice for this research, however there was more detailed information necessary to create mock-ups.
5
Experiments
This section describes the experiments. The different mock-ups with the added functions that will be used are shown. This section also describes how the interviews will be conducted and the background of the interviewees will be described.
5.1 Prototype mock-ups
The user-personalization profiles must be visualized in prototype mock-ups so that they can be tested. Tableau 1 software is chosen to create the personalized dashboards. To add components to the visualization that Tableau doesn’t include Microsoft Paint is used. There will be a mock-up created for all design choices (see section 4.5). The original dashboard is shown first to be able to see the difference the personalization makes (see Figure 4).
Figure 4 . Original dashboard: Overview employees
Figure 5, Figure 6 and Figure 7 show the mock-ups that were used during the inter-views. The added components in the dashboard can be found by looking at the numbers. The numbers in the mock-ups correspond with the numbers in Table 1. Positive correla-tions can be identified in the mock-ups by the ’+’ sign. The other correlacorrela-tions (negative, no or possible) are also shown on these mock-ups. These other correlations are shown as ’O’ because these correlations were not interesting for this research.
1
Figure 5 . Dashboard mock-up: limited visibility
Figure 6 . Dashboard mock-up: location
5.2 Interviews
The mock-ups will be discussed with a professional from the field during an interview. This (semi-structured) interview will be done in a workshop-like setting. The goal of the interview is to get opinions and arguments from medical professionals about the personalized dashboards. With this information the correlations in Table 1 can be tested.
Figure 7 . Dashboard mock-up: visualization type
type of research allows the interviewees to describe their thoughts (McKernan & McKer-nan, 2013). This also allows the interviewer to better understand the answers and ask extra information when this is not the case (Sharpe, 1999). Another advantage is that the interviewees are not restricted to the answers provided. Furthermore, the reason for choosing this type of research is the limited time to find enough professionals and this type of research needs a smaller sample size (Boxill, Chambers, & Wint, 1997).
Before starting the interview, the professional will be asked some descriptive questions (such as age and job) about themselves. During the interview the professional is told about the specific user profile and a mock-up is shown (with the original to show the difference) and the professional is then asked: "Do you think these personalization aspects would be useful in real-life?". This will lead to follow-up questions which will reveal the arguments given by the professional. These arguments will be used to describe the results of the experiments. The answers of the professionals will be (audio, if the interviewee gives permission) recorded and interpreted in a qualitative way. The results of the interviews are described in section 6 Results.
5.3 Interviewees
This paragraph describes the background of the interviewees. For the experiments six professionals have been interviewed. The total of six is chosen because similar researches also interviewed six professionals (Nobarany, Haraty, & Fisher, 2012), (Tor, 2015). These interviewees are health care professionals which were chosen from health care organizations in the Netherlands. The characteristics of the interviewees can be described as follows:
• All interviewees are Dutch (and raised in the Netherlands)
• All interviewees work in a health care setting
• Age of the interviewees varied between 24 and 53 (avarage age was: 41) • Gender of the interviewees: 4 males and 2 females
• Level of education of the interviewees differed, 4 interviewees were at least univer-sity educated (a Bachelor’s degree or more) in their field of work and the other 2 interviewees didn’t have a university degree but were educated in their field of work • Two interviewees were business intelligence developers and the other four interviewees
were users
Table 2 shows detailed information about the interviewees who have direct and exten-sive knowledge of using dashboards in a medical environment. This table shows the gender, age, job and (dashboard) usage of the interviewees.
Gender Age Role Use
Interviewee 1 Female 50 Care coordination Sometimes
Interviewee 2 Male 24 BI developer Development daily
Interviewee 3 Female 24 Elderly caregiver Rarely
Interviewee 4 Male 42 Caregiver handicapped people Avg. 3 times a week
Interviewee 5 Male 53 Business analyst Daily
Interviewee 6 Male 53 BI developer Development daily Table 2
Descriptive information about the interviewees
6
Results
This chapter describes the results coming from the interviews. First the general results will be described after which the results of each user profile will be described.
6.1 General results
The interviews were used to check if the interviewee did or did not agree to the statements. The interviewees who answered with ’no, because ..’ were not specifically asked if there was no correlation, a possible correlation or a negative correlation. Most answers of the interviewees were similar to Table 1, but not one interviewee completely agreed or disagreed with this original table.
6.2 Limited visibility
In general most interviewees agreed with the statements made in this thesis, only two inter-viewees disagreed with some statements made. Four of them agreed with all the statements made. These are results per function and interviewee (see also Table 3):
• Select function. All interviewees agreed that this function has a positive correlation with this user profile. Most interviewees recommended to use a contrasting color for the selection. One interviewee (a developer) also mentioned that this would cost some time to implement but potentially could save the user time.
1. Select 2. Explore 3. Reconfigure 4. Enco de 5. Abstract 6. Filter 7. Connect 8. Other Original + P P + + - + + Interviewee 1 + O O + + + + + Interviewee 2 + O O + + O O + Interviewee 3 + O O + + O + + Interviewee 4 + O + + + O + + Interviewee 5 + O + + + O O + Interviewee 6 + O + + + O O + Table 3
Results from interviews - limited visibility
• Explore function. All interviewees agreed that this function does not have a positive correlation with this user profile. Most of them argued that this function does not clarify the screen and only changed the data shown on the screen. One interviewee mentioned that if such a function should be added, buttons were preferred over a filter function.
• Reconfigure function. Three interviewees argued that this function does not have a positive correlation with this user profile because this function does not clarify the screen for this user profile. The other three interviewees argued that this function does have a positive correlation. Their argumentation for this positive correlation is that reconfiguring the representation could help this user profile because some representations could make the data on the screen clearer for the user.
• Encode function. All interviewees agreed that this function has a positive correlation with this user profile. Some interviewees also mentioned that there are different types of visual representations of the data (for example: patterns or different colors). Two of them preferred using a pattern, two others preferred using different colors and the last two interviewees did not have a preference. One interviewee (a developer) also mentioned that this would cost some time to implement but potentially could save the user time.
• Abstract function. All interviewees agreed that this function has a positive correlation with this user profile. Most (4 out of the 6) interviewees thought that this function also could be useful for themselves. One interviewee said that it is important that it always has to be clear what exactly a function does. Another interviewee (a developer) also again mentioned that this would cost some time to implement but potentially could save the user time.
corre-lation with this user profile because this function does not clarify the screen for this user profile. Two of them mentioned that this is a useful function but not specifically for this user profile. One interviewee however said that this function could be useful because this function allows the user to remove (by using a filter) data on the screen which means there is less data on the screen and clarifies it for the user. This inter-viewee also mentioned that this function could be useful for elderly people because they are often faced with problems with their sight.
• Connect function. Three of the interviewees argued that this function has a positive correlation with this user profile. They agreed that this would clarify the data for the user and allow the user to get a quick overview of the screen. The other three intervie-wees argued that that this function does not have a positive correlation with this user profile because this function will only create confusion for the user. One interviewee (a developer) mentioned that this specific function is difficult to develop because the axes of the graphs must have the same ratio. Another interviewee suggested that a function where the graphs slide into one another would be more useful.
• Other functions. All interviewees agreed that other functions do have a positive correlation with this user profile. Three of the interviewees said that they would also find such functions useful for themselves. One interviewee (a developer) mentioned that it is important with this specific function that the user does not have to scroll to be able to see the data when he or she is using the zoom function.
6.3 Location
In general, this user profile divided the interviewees in two equal groups. Three interviewees disagreed with some of the statements made in this thesis and the other three interviewees agreed with all the statements made. These are results per function and interviewee (see also Table 4): 1. Select 2. Explore 3. Reconfigure 4. Enco de 5. Abstract 6. Filter 7. Connect 8. Other Original O O O O O + O + Interviewee 1 O O O O O + O + Interviewee 2 O O O O O + O + Interviewee 3 O O O O O + O + Interviewee 4 + + O O O + + + Interviewee 5 + + O O + + O + Interviewee 6 O O O O + + O + Table 4
• Select function. Four of the interviewees argued that this function does not have a positive correlation with this user profile. They all stated that this specific function was not useful for this user profile because this function has nothing to do with the location of the user. Two of the interviewees thought that this function does have a positive correlation with this user profile because data items relevant to the location can be selected although this selection is not made by the user.
• Explore function. The same four interviewees as the ones before thought that this function does not have a positive correlation with this user profile. They again stated that this specific function was not useful for this user profile because this function has nothing to do with the location of the user. The two other interviewees thought that this function has a positive correlation with this user profile because a new set of data items can be explored based on the location.
• Reconfigure function. All interviewees agreed that this function does not have a positive correlation with this user profile. All of them thought that this specific function was not useful for this user profile because this function has nothing to do with the location of the user. One interviewee mentioned that this function maybe could be useful to select the right the semantic arrangement of the data to comply to the culture of the user, but that this is very unlikely.
• Encode function. All interviewees agreed that this function does not have a positive correlation with this user profile. They again all thought that this specific function was not useful for this user profile because this function has nothing to do with the location of the user. One interviewee mentioned that this function maybe could be useful to select a specific semantic shape of the data items to comply to the culture of the user, but that this is very unlikely. Another interviewee also mentioned that he would rather use function 1 (select) than this function.
• Abstract function. Four of the interviewees argued that this function does not have a positive correlation with this user profile. They all stated that this specific function was not useful for this user profile because this function has nothing to do with the location of the user. The two other interviewees thought that this function does have a positive correlation with this user profile because the level of abstraction can be adjusted based on the location (zoom-in on the data of the location) according to them.
• Filter function. All interviewees agreed that this function has a positive correlation with this user profile. They all agreed that this is a useful function because with this function the data can be filtered for the user based on the location ("which reduces
the actions the user has to do"). Two of the interviewees mentioned that some users
need to see other data, so it is important to be able to change the filter manually. One interviewee also mentioned that the language of the screen can be change based on the location, but he didn’t know if this belonged to this function.
• Connect function. Five interviewees agreed that this function does not have a positive correlation with this user profile. Two of them argued that this specific function
was not useful for this user profile because this function has nothing to do with the location of the user. The other three argued that this function only creates confusion for the user. One interviewee however thought that this function does have a positive correlation with this user profile because he thought that the useful connections for this user profile could be showed on the screen.
• Other functions. All interviewees agreed that other functions do have a positive correlation with this user profile. All of them stated that this would be useful because sometimes the user wants to see data from other locations (two of them mentioned that the user must have permission to do this). One of the interviewees also mentioned that this is an easy to use function and yet another interviewee said that it differs per user how much this function will be used.
6.4 Visualization type
In general most interviewees didn’t completely agree with the statements made in this thesis. From the three user profiles in the interview, these results differed the most from the original statements made. Only one interviewee completely agreed with the statements made. These are the results per function and interviewee:
1. Select 2. Explore 3. Reconfigure 4. Enco de 5. Abstract 6. Filter 7. Connect 8. Other Original + O + + O O O O Interviewee 1 + O + + O O O O Interviewee 2 + + + + O + + O Interviewee 3 + O + + + O + O Interviewee 4 + O + + O O O O Interviewee 5 O O O + + O O + Interviewee 6 + + + + O O O + Table 5
Results from interviews - visualization type
• Select function. Five interviewees agreed that this function has a positive correlation with this user profile. They agreed that this function is useful for this user profile because they thought that the selection of the user could be saved for when the user returns. One of them mentioned that this function will save the user some time. One interviewee thought that this function does not have a positive correlation with this user profile. He preferred to start at the original screen rather than saving the selection made previously.
• Explore function. Four interviewees agreed that this function does not have a positive correlation with this user profile. They all stated that this function only changes the data and does not change the visualization. Two interviewees argued that this function has a positive correlation with this user profile. One of them argued that this function could be a good check for the user when the visualization type has changed. The other argued that this function would be useful because he thought it would be useful to save the users progress for when the user returns.
• Reconfigure function. Five interviewees agreed that this function has a positive cor-relation with this user profile. They agreed that this function is the most important function for this user profile. One interviewee mentioned that this function adds flexi-bility for the user. Another interviewee mentioned that it is important to only let the user choose useful (selected by the developers) visualization types. One interviewee (a developer) argued that this function does not have a positive correlation with this user profile because the developer specifically chose this visualization type.
• Encode function. All interviewees agreed that this function has a positive correlation with this user profile. They all agreed that this could help the user. One of them mentioned that this adds flexibility for the user, while another thought this was useful but was not sure if he would use this function. Another interviewee mentioned that this would be useful but maybe only selected users could use it (to prevent that the majority has an added function for some specific people while they do not need it). • Abstract function. Four of the interviewees argued that this function does not have a
positive correlation with this user profile. Two interviewees stated that this function only changes the data and does not change the visualization. One interviewee did not think this function was useful and would prefer using a filter. Another interviewee mentioned that this function would create confusion and he preferred to start at the original screen rather than saving the abstraction level made previously. Two interviewees argued that this function has a positive correlation with this user profile. They both argued that it would be useful to save the abstraction level of the user could be saved for when the user returns.
• Filter function. Five interviewees argued that this function does not have a positive correlation with this user profile. Four of them stated that this function only changes the data and does not change the visualization. The other interviewee stated that this function would create confusion and he preferred to start at the original screen rather than saving the abstraction level made previously. One interviewee argued that this function does have a positive correlation with this user profile because this function ca be used as a check for the user (as long as this function will not become a burden). • Connect function. Four interviewees agreed that this function does not have a posi-tive correlation with this user profile. These four interviewees argued that this specific function was not useful for this user profile and that this function only creates confu-sion for the user. The two other interviewees argued that this function could be useful for this user profile because these connections can be saved for the user.
• Other functions. Four interviewees agreed that this function does not have a positive correlation with this user profile. These interviewees argued that this specific function was not useful for this user profile because this function only changes the data and does not change the visualization. Two interviewees argued that these other functions could be useful for this user profile because this allows the user to return to the original start screen after they changed (for example the visualization type) the screen.
7
Conclusion
This chapter describes the conclusions of this research. In this chapter the research question (and sub-questions) will be answered.
7.1 Research question
The main research question for this thesis is:
• How can visual analytics be personalized to get better insights for medical profession-als in a health care environment?
To be able to answer the main research question, this question is divided in the following sub-questions:
1. SQ1 - Which parts are there in a visual analytics process?
2. SQ2 - Which parts of the visual analytics process can be personalized and how can this be achieved?
3. SQ3 - What types of users can benefit from personalization and in what way can they do so?
7.2 SQ1 - Visual analytics process
Two visual analytics process models have been chosen to discuss in this thesis because these are the most known. The first model is developed by Pirolli and Card (2005) (see Figure 2). The second model is developed by D. Keim et al. (2008) and D. A. Keim et al. (2010) (see Figure 3). For this thesis Keim’s model is used because this model explicitly names the user interactions in the visualization step. This interaction with the user is necessary for the mock-ups (see section 4.2). Keim’s visual analytics process model does contain the following elements:
• Data: this first step (sometimes referred to as sources) pre-processes and transforms the data to derive different representations for further exploration. Other tasks in this step are data cleaning normalization, grouping, or integration of heterogeneous data sources. After this step the analyst can choose to either apply automated analysis (through data mining) to create models or hypotheses; or map the data to create visual representations.
• Visualizations: in this step the visualization is generated (or updated). The visual-ization can be created from the data or from the hypotheses.
• Hypothesis: in this step the hypothesis is generated (or updated). Hypotheses can be generated from data or from visualizations.
• Insights: in this step insights are gained from the visualization and/or hypotheses steps. These insights always have five characteristics: complex, deep, qualitative, unexpected and relevant (North, 2006).
7.3 SQ2 - Personalizing visual analytics
The steps; visualizations, hypotheses and insight in the process of model of D. Keim et al. (2008) can be personalized. Personalized visualizations can be created from data or the hypotheses. Personalized hypotheses can also be created from the data but can also be created from visualizations. Data cannot be personalized because data transforms the data coming from (not personalized) data sources. The personalized visualizations and hypotheses will lead the user to personalized insights. Insights is thus more of a follow-up step of the previous personalization steps.
For this thesis research is done for the visualization and hypotheses steps. These steps were personalized in mock-ups and discussed with professionals. The results of these personalized steps of the visual analytics process are discussed in section 6.
7.4 SQ3 - Benefitting types of users
All user-profiles that were selected in this thesis for interviewing could benefit from person-alization. These selected user profiles were:
1. Limited visibility - personalize the dashboard for a visually impaired or colorblind user
2. Location - personalize the dashboard to the location of the user
3. Visualization type - personalize the dashboard to a graph type the user selected During the interviews all user-profiles had (one or more) positive correlation with the func-tions. Some only had two positive correlations while others had five positive correlations (out of eight). Nothing conclusive can be said about the other user-profiles, because of time limitations these were are not addressed during the interviews.
7.5 Conclusion
Visual analytics can provide better insights for medical professionals by personalizing the visualizations and hypotheses steps in the process of model of D. Keim et al. (2008). This will lead to personalized insights.
This thesis showed how to personalize these (visualizations and hypotheses) steps, which can be done by creating user-profiles in which the user is put into. A dashboard will then create a personalized version of itself based on this profile. The user will then have added functions to the dashboard which will help the user in getting personalized insights. With these personalized insights, the user can make better decisions from which the company will benefit. The functions the user-profiles are benefitting from, can be found in section 6.
While developing these functions, the designer has to keep the design rules (see sec-tion 3.9) in mind. The designer of a dashboard has follow these rules when creating the personalized dashboard for the user-profiles. For selections that the user can make it is important that the designer makes a pre-selection with approved options (for example rule six) because this can prevent the user from drawing incorrect conclusions. The designer also has to keep in mind that the dashboard will be on one screen and is properly designed to prevent the dashboard becoming unclear because of the added functions.
If there is a user-profile that does not have a positive correlation with a function then this profile does not have to exist because the original dashboard will be showed. If there is a user-profile that has positive correlations with all functions, this profile is probably not specific enough.
8
Discussion
This chapter describes the limitations of this research and possible future research based on this thesis.
8.1 Limitations
Although this research provides insights about personalized visual analytics, there are some limitations that have to be addressed. The biggest limitation of this research is its gen-eralizability. Although the interviews showed correlations between the user-profiles and functions, this cannot be proven statistically because of the limited amount of people in-terviewed. This research can maybe also be conducted statistically in a way similar to this research, because all interviewees said that the interview questions were easy to understand. A problem with the interviews is that some correlations required more explanation which could affect the opinion of the interviewee.
Furthermore, this research was conducted in the Netherlands with all interviewees coming from the Netherlands (all of them came from the Province of Noord-Holland). This means that the outcomes of the research maybe are not valid for other parts of the Netherlands or other countries. Further study with more interviewees could add these findings to this research.
Another limitation is that not all user profiles (in this thesis) are provided with the results of interviews. Only three have been selected because of time limitations. Another limitation with the user profiles is that there are many more profiles than the eight profiles that were selected for this thesis. Some user profiles that could be added are for example: age (adjust for elderly people), gender, living area, ethnicity, culture and personal interests. The last limitation is the objectivity of the interview questions. These are created by the author himself which may cause a certain degree of subjectivity. The questions would have been more objective if these were created by multiple people.
8.2 Future work
Future research could add results of the other user-profiles in this thesis after interviewing professionals from the field of study (time, pattern, requirements, role and shown data).
As mentioned before (in section 8.1) the research could also be extended by adding new user-profiles. These profiles could for example be:
• Age, adjust the dashboard based on the age of the user (this could for example be useful for elderly people)
• Gender, adjust the dashboard based on the gender of the user
• Living area, adjust the dashboard based on the living area of the user (where the focus should be on the area the user is coming from and not where the dashboard is used)
• Ethnicity/culture, adjust the dashboard according ethnicity or culture of the user (this could prevent a user being offend by something on the dashboard)
• Personal interests, adjust the dashboard according personal interests of the user (but this would probably not apply to a professional context)
Furthermore, future research could add results from other parts of the Netherlands (or other countries) to this research because this research was conducted in the Netherlands with all interviewees coming from the Netherlands (all of them came from the province of Noord-Holland). By increasing the number of professionals (quantitative research which adds statistic results) that participate this future research could also be proven statistically. The last recommendation for future research is splitting function eight (other inter-action techniques). This was a function which sometimes confused the interviewees because there can be multiple functions in this function.
Acknowledgments
I would like to thank my thesis supervisor Marcel Worring for the useful comments and support during the writing of this master thesis. Furthermore, I would like to thank the interviewees who shared their precious time with me by answering my questions. Lastly, I would like to thank my family and friends who supported me throughout the entire process. Writing this thesis would not have been possible without all these people.
