Background and Significance
Crohn's disease is an inflammatory bowel disease (IBD) with symptoms that can include
diarrhea, inflammation of both the gut and other parts of the body, fatigue, abdominal
pain, and weight loss, among others. Colitis refers to inflammation of the inner lining
of the colon, and commonly co-occurs with Crohn's disease. There is no known cure
for either condition, although certain therapies can help treat their symptoms, sometimes
bringing about long-term remission. Thus, treatment largely consists of disease management.
Given the varied ways in which these conditions can present themselves in different
patients, and their chronic nature that affects every facet of patients' lives (e.g.,
social interaction, family, work, diet, and sleep), researchers in the University
of North Carolina at Chapel Hill chapter of the Crohn's and Colitis Foundation—IBD
Partners (formerly Crohn's and Colitis Foundation of America [CCFA] Partners)—are
interested in engaging patients to aid them in disease management and to collect information
useful for researching potential treatments. To this end, they have created an interactive
website that provides a discussion forum for patients to talk about their experiences,
suggest and discuss new lines of research into their conditions, and vote on promising
research topics.[1]
Although such a forum can be invaluable for generating and prioritizing research questions
based on patient experiences, it can be time and labor intensive sifting through all
of the questions and comments on the discussion forum, trying to effectively interpret
such a large volume of text. IBD Partners is interested in developing more efficient
approaches for identifying common themes and determining which research questions
are most frequently discussed by patients. Interactive visualization offers a potential
solution to help clinicians and researchers explore the data and identify the salient
questions and needs of the patients.
Interactive visualization has proven to be a useful method for analyzing datasets
across a wide range of disciplines, including in the health care domain,[2]
[3] and holds great promise for advancing the state-of-the-art in health care Many visualization
tools for health care applications operate on a wide variety of structured data.[4]
[5]
[6]
[7] Prior work in visualizing structured data from patients with various types of abdominal
pain includes that of Rao et al, which involves the extraction of diagnostic paths
from electronic health record (EHR) data.[8] While such work can be very effective, much of it is not directly applicable to
the visualization of largely unstructured text from an online patient forum. Sorbello
et al present the utility of using structured text—MeSH terms—with a visual analytics
interface for pharmacovigilance; however, they do not directly address the problem
of extracting structured information from free text.[9] Tools such as Jigsaw[10] enable the interactive extraction and visualization of named entities and their
co-occurrence from document collections, and the work of Sampathkumar et al shows
the utility of an ontology-based approach for visualizing data from online health
forums.[11]
Ontologies are controlled vocabularies that represent knowledge about a domain of
interest.[12] They offer richer representations than other controlled vocabularies (e.g., taxonomies,
thesauri) because they enable relationships beyond hierarchical and synonymous. Ontologies
have a long history in medicine and biological research,[13]
[14]
[15] and are used for a variety of purposes, e.g., classifying literature for information
retrieval, mapping and integrating diverse data sources, aggregating/clustering information,
and natural language processing applications.[16]
[17] Biomedical ontologies tend to focus on representing encyclopedic knowledge about
a given domain. For example, UBERON[1] contains approximately 20,000 concepts ranging from very granular anatomy (e.g.,
cell membrane) to larger systems (e.g., digestive system).
We have created an ontology to help organize the content of the IBD Partners forum
and make it more suitable for computational analysis, and developed an interactive
visualization prototype utilizing the ontology to enable the interactive exploration
of the patient-generated forum content.
Methods
Forum Data
The data snapshot used when creating the CCFA forum ontology consists of 97 research topics (i.e., user posts consisting of a proposed research question and a description of
the question), and 121 user comments made by fellow patients on proposed questions,
for a total of 17,322 words. An example research topic post is the following:
Question:
Nicotine has shown to be effective for UC [ulcerative colitis] in some individuals,
both prior- and nonsmokers. What is the mechanism? Does nicotine affect the microbiome,
the immune system, or both?
Description:
Big Pharma will not take on the role of studying nicotine as there is no $$$ in it.
Few studies with small sample sizes have been done but more research is needed.
Each research topic also has an anonymized user ID (400 unique users), the number
of votes for each topic (1,246 total votes), and one of nine predefined categories
(diet, medications, procedures and testing, environment, alternative therapies, lifestyle,
genetics, exercise, and other) selected by the topic creator.
Ontology Creation
We initially performed an analysis of the forum data using some basic linguistic processing
on the text, such as calculating word and phrase frequencies. However, the results
did not effectively capture the forum conversation. Phrases such as inflammatory bowel, controlled trial, and disease activity appeared frequently, which simply confirmed the obvious: patients were discussing
IBD and its research. These frequencies did not capture the nuance of specific lines
of research the patients were interested in.
We therefore created an ontology of the forum conversation to provide a structure
that more effectively captures the depth and breadth of research topics in which the
patients were interested. To create the ontology, we first conducted an in-depth,
manual exploration of the forum text. Specifically, we applied content analysis to
the forum text, sifting through manifest content (i.e., what is seen directly in the
text, such as the occurrence of a particular word) to find latent content (i.e., underlying
meaning, connotation, nuance). According to Wildemuth, “An example of latent content
is the level of research anxiety present in user narratives about their experiences
at the library.”[18] In other words, a user may not directly state, “I am so anxious.” Instead, the anxiety
may be implied, e.g., “My heart won't stop beating so fast” or “I wish I could relax.”
Wildemuth notes, “Sometimes there is no existing theory or research on your message
populations; you may not know what the important variables are. The only way to discover
them is to explore the content.”[18] In other words, it may be impossible to identify themes without first immersing
one's self in the text, allowing the themes to be revealed as one becomes more intimate
with the conversation. This is reflected in the fact that the most common predefined
category assigned by users to their proposed research topic was other (34 out of 97, over 40%), implying that the categories did not fully capture the
breadth of their interests and discussions.
The manual analysis was performed by a single member of the research team with significant
experience in content analysis. Spreadsheet software with manual entry was used to
keep track of the analysis. After completing the content analysis, it was clear that
no existing ontology would adequately represent the patient conversations. Most biomedical
ontologies provide encyclopedic objective knowledge about a particular subject, whereas
the CCFA forum text describes personal patient experiences, emotions, and desires.
The goal of CCFA physicians and researchers is to understand their patients' needs
and wants, and an effective ontology needs to reflect this goal to help bridge the
gap between how clinical practitioners, researchers, and patients view their conditions.
Our ontology structured and classified the raw information in the forum. Concepts
(e.g., medication, surgery, diet, and symptom) discussed in the forum became “classes” in the ontology. Although relationships
beyond hierarchical (e.g., medication treats symptom) are possible, our ontology does not currently include such relationships,
and thus functions primarily as a taxonomy—a hierarchical grouping of terms. As the
ontology is expanded, other types of ontological relationships (such as medication
treats symptom) will be added. Additional concepts (primarily from the CCFA website to align
with their approach to care) were included in anticipation of future forum conversations.
Where applicable, classes from two pre-existing ontologies, the Ontology for Adverse
Events (OAE)[2] and the Disease Ontology (DO)[3], were used. In total, 165 classes from the OAE and 36 from the DO were included.
During the ontology creation, we consulted with the IBD Partners team to ensure that
the ontology structure seemed appropriate. The resulting ontology describes a hierarchy
of 337 total classes, with seven top-level classes: comorbidity, diagnosis/monitoring method, IBD course, quality of life, risk factor,
symptom, and treatment method, and a maximum depth of 6. The ontology was created using Protégé,[19] exported in OWL format[4], and converted to an OBO Graph[5] using ROBOT[6] for easy ingestion into our visualization tool. Based on the content analysis, each
research topic was labeled with one or more terms from the ontology. Although the
initial content analysis was conducted based on the research topic question, description,
and comments, only the question and description were used for labeling. A chart showing
the frequency of each top-level ontology term when labeling the research topics, along
with immediate children for top-level terms that have a child with frequency greater
than 1, is shown in [Fig. 1]. The ontology structure and linkage to the research topics enable the interactive
visualization described in the next section.
Fig. 1 Ontology class frequencies as used to label patient-generated research topics. Top-level
classes, and immediate children for classes with a child with a frequency of at least
2, are shown.
CCFA Explorer
The CCFA Explorer is a browser-based tool developed using the D3 visualization library[20] that consists of three different interactive visualizations: (1) the CCFA forum
ontology, (2) an overview of the patient-generated research topics, and (3) a detailed
view of the forum text and other information about each research topic ([Fig. 2]). The ontology visualization enables researchers to understand the structure of
the ontology, see which areas of the ontology were more frequently discussed by the
forum users, and see how frequently different ontology terms were discussed together
in the same research topic. The research topic overview enables researchers to quickly
identify clusters of research topics that discuss similar ontology terms, and the
detailed view enables the researcher to read the forum text in-depth. In order to
understand relationships between ontology terms and research topics, users can select
visual elements representing ontology terms or research topics in each view. All three
views are linked to automatically highlight relationships from the various visual
elements in each view to the selected items. These linked views enable the researcher
to, for example, quickly examine the forum text associated with an ontology term of
interest, or determine which ontology terms are related to a cluster of research topics.
To develop effective interactive visualizations, Shneiderman's visual information
seeking mantra—overview first, zoom and filter, then details on demand—has been adopted by a wide range of data visualization tools.[21] We adopt this approach, providing overviews of the CCFA ontology and forum content,
along with the ability to filter and obtain detailed forum content based on patterns
and relationships discovered from interacting with the overviews.
Fig. 2 The CCFA Explorer interface: ontology (left), topic overview (middle), and topic details (right).
Ontology Visualization
Hierarchies are a specific form of ontology, in which each node may have at most one
parent, and multiple children. Visualization techniques for hierarchies include tree
maps,[22] icicle plots,[23] and tree diagrams (e.g., tidy trees[24]). Although such visualization techniques are effective for showing hierarchical
structure, they are not designed to show other types of ontological relationships.
Network diagrams offer the ability to encode different types of relationships via
different styles of links in the diagram. Due to this flexibility we adopted this
approach, although the current version of the ontology contains only hierarchical
“is a” relationships. Kamdar et al present research analyzing user interactions with
biomedical ontologies for different visualization types, including network diagrams,
and show that different users interact with ontologies differently.[25] Such research suggests that a suite of ontological visualization approaches may
be useful, especially when dealing with different user populations, which will help
inform our future work.
The CCFA Explorer force-directed network shows the ontology structure and indicates
the most prominent ontology terms ([Fig. 2], left). Each ontology term is represented by a node in the visualization, and links
(i.e., arrows connecting nodes) indicate “is a” hierarchical relationships (e.g.,
medicine is a treatment method). Node radius is proportional to the number of research topics labeled with that
ontology term. For any given ontology term, if a research topic has been labeled with
that term, the research topic is labeled with all ancestors of that ontology term.
Thus no child node will ever be larger than its parent. When the visualization initially
loads, node labels for top-level terms in the ontology are visible. Labels for other
nodes appear when the user hovers over a node, or upon user selection as described
in the Interactive Selection and Highlighting section.
Many researchers may already have an idea of what ontology terms they are interested
in. To facilitate rapid identification of predetermined areas of interest, the ontology
visualization includes a search box. The user can begin typing into the search box,
which shows suggestions for all matching ontology terms. Node labels for all matching
ontology terms will be shown and highlighted in red, enabling the user to investigate
nodes of interest.
Topic Overview
The topic overview uses t-SNE,[26] via the t-SNE.js library[7], to lay out circular glyphs representing each research topic ([Fig. 2], middle). t-SNE is a technique for dimensionality reduction that can be used to
lay out objects (e.g., research topics) in two dimensions based on their similarity
across a large number of dimensions (e.g., labeled ontology terms). We use t-SNE to
place research topics labeled with similar sets of ontology terms closer together,
which enables the user to visibly identify clusters of research topics labeled with
similar sets of ontology terms. The radius of each glyph is proportional to the number
of forum comments made in response to that research topic, and the outline thickness
is proportional to the number of user votes for that topic, enabling the user to identify
popular topics. The glyph color represents which of the nine predefined categories
was chosen by the research topic creator.
We introduce three modifications to the standard t-SNE layout to enable more effective
visualization of the CCFA forum data. (1) Because two or more research topics may
be labeled with similar sets of ontology terms, glyphs may overlap and occlude each
other. Such overplotting can make it difficult to see cluster sizes for very similar
research topics, and to see and select individual topics. We therefore apply a force-directed
layout for overlapping glyphs that separates the centers of each glyph while maintaining
some overlap to indicate closely related clusters ([Fig. 3A]). (2) Due to the hierarchical nature of the ontology, we enable weighting of higher-level
(closer to the root) or lower-level (closer to the leaves) ontology terms to determine
at which level in the hierarchy research topic glyphs are clustered. Weighting higher-level
ontology terms results in fewer clusters based on more general terms ([Fig. 3B]), and weighting lower-level terms results in more clusters based on more specific
terms ([Fig. 3C]). (3) Greater weights can be applied to the currently selected ontology terms, resulting
in clusters reflecting combinations of the selected terms. For example, [Fig. 3D] shows a layout emphasizing two selected ontology terms, with three clusters indicating
the presence of only the first term, only the second term, or both terms. This feature
enables, for example, easy selection of all research topics with a given set of ontology
terms.
Fig. 3 Modifications to the standard t-SNE layout: (A) force-directed layout of overlapping glyphs to increase cluster legibility, (B and C) differential weighting of ontology terms emphasizing (B) higher-level terms resulting in fewer, more general clusters and (C) lower-level terms resulting in a larger number of more specific clusters, and (D) emphasizing the currently selected ontology terms for clustering.
Topic Details
The topic details view is a scrollable list of panels for each research topic in the
forum. Each research topic panel contains the research question, description, and
comments for that topic, along with additional information such as the number of user
votes, color-coded user-selected category, and tags indicating the ontology terms
labeling that topic ([Fig. 4]). Users may select three different levels of details to display each research topic's
text: (1) question only, (2) question and description, and (3) question, description,
and comments. The list of research topics can be sorted by topic ID, user ID, number of votes, number of comments, and category. The list can also be filtered based on currently selected research topics or ontology
terms, as described in the Interactive Selection and Highlighting section. In addition,
the user can search for text in the search box, with the matching text highlighted
in red in each research topic panel.
Fig. 4 An example research topic in the topic details view.
Interactive Selection and Highlighting
The user can interactively select visual elements representing ontology terms or research
topics in any of the three views, and all views will be automatically updated to highlight
relationships to the selected items. These linked views enable the researcher to perform
actions such as finding all research topics labeled with a selected set of ontology
terms, or determining which ontology terms a selected cluster of research topics share
in common.
We define three types of possible relationships between ontology terms and research
topics: (1) the co-occurrence between two ontology terms is the number of research topics that have been labeled
with both terms, and therefore is an indication of which ontology terms are discussed
together by the forum users. For multiple selected ontology terms, the co-occurrence
between a term and the selection is the size of the union of the common research topics.
(2) The association between two research topics is the number of ontology terms that the two topics share
in common, and is an indication of how closely related the two topics are. For multiple
selected research topics, the association between a research topic and the selection
is the size of the union of the ontology terms they have in common. (3) The connection between an ontology term and a research topic is 1 if the topic is labeled with that
term, and 0 otherwise. For multiple selected ontology terms or research topics, the
connection is the sum of each individual connection.
In the ontology visualization, ontology terms can be selected by clicking on the node
for that term. In the topic overview, research topics can be selected by clicking
on the glyph for that topic. In the topic details view, research topics can be selected
by clicking on the panel for that topic, and ontology terms can be selected by clicking
on the tag for that term in any given topic. In all views, selected visual elements
are represented by dashed outlines for consistent representation of selections. Selection
in any view results in highlighting in all three views.
In the ontology visualization, the co-occurrence with any currently selected ontology
terms is represented by an inset circle for each node, with size proportional to the
co-occurrence and color proportional to the percent co-occurrence (co-occurrence divided
by total number of research topics connected to the selected ontology terms × 100)
with the selected ontology terms ([Fig. 5A]). Similarly, the association with any currently selected research topics is represented
by an inset circle with radius proportional to the association, and color proportional
to the percent association (association divided by total number of selected research
topics × 100) with the selected research topics ([Fig. 5B]). In both cases, labels are displayed for any nodes with a percent co-occurrence/association
of at least 25%. In the case of selected ontology terms and selected research topics,
highlighting research topic connections takes precedence in the ontology visualization.
Whenever there is a current selection being used for highlighting, a label is shown
in the visualization indicating what is currently being highlighted, e.g., “nodes
colored by co-occurrence with two selected ontology terms” or “nodes colored by connection
to three selected topics.” Automatic highlighting of the ontology visualization enables
the user to quickly find ontology terms that are discussed in the same research topics,
and which ontology terms are related to the selected group of topics.
Fig. 5 Interactive highlighting of the ontology visualization, enabling (A) highlighting of co-occurrences with a selected ontology term (drug), and (B) highlighting of connections to research topics selected in one of the other views.
In the topic overview, the connection with any currently selected ontology terms is
mapped to glyph color saturation, normalized by the total number of selected ontology
terms ([Fig. 3D]). Similarly, the association with any currently selected research topics is also
mapped to glyph color saturation, normalized by the by total number of ontology terms
for that glyph's research topic (such that any selected topic will be fully saturated).
In the case of selected ontology terms and selected research topics, highlighting
ontology term connections takes precedence in the topic overview. Whenever there is
a current selection being used for highlighting, a label is shown in the topic overview
indicating what is currently being highlighted, e.g., “topic color saturated by association
with four selected topics” or “topic color saturated by connection to one selected
ontology term.” Automatic highlighting of the topic overview enables the user to quickly
find research topics related to ontology terms of interest, and discover research
topics with ontology terms in common.
In the topic details view, research topics can be optionally filtered by selected or connected. For selected, if any research topics are selected, only those topics will be shown. For connected, if there are any selected ontology terms or research topics, only topics with a
nonzero connection or association will be shown. In this manner, the user can quickly
drill down to see the forum text related to ontology terms or research topics of interest.
In addition, the same color map applied to the ontology nodes during highlighting
is applied to the ontology term tags for each research topic ([Fig. 6C]).
Fig. 6 Example use case with selection of ontology terms (A), selection of research topics related to those terms (B), and detailed inspection of selected research topics (C).
Discussion
After presenting the CCFA Explorer tool to members of the IBD Partners team, we received
useful feedback that will help inform our future work. In general, they thought that
the tool was a useful way to explore the CCFA forum data, and made it possible to
quickly identify major themes and popular research topics; however, they felt that
effective use of some of the tool's features may be too complex for users—both researchers
and others—who are unfamiliar with advanced interactive visual interfaces. Two themes
in particular that were identified to address this issue were (1) the utility of a
simplified patient-facing interface focused on helping forum users find similar patients
and more easily identify research topics relevant to them, and (2) a researcher-facing
interface focused on helping researchers in specific domains quickly identify information
related to their research area and generate summaries of relevant information that
can be easily presented to stakeholders. To this end, we intend to refine our tool
in various way. For example, the ontology visualization, while effective at showing
the overall structure of the ontology and highlighting relationships with the ontology
terms, is not very well suited for navigation to find ontology terms of interest.
We therefore plan to redesign our ontology visualization to make navigation easier,
while incorporating some of our current work in interactive highlighting. We also
plan to explore the use of text summarization techniques to include in a summary panel
that will present an infographic-like view of any currently selected ontology terms
or research topics.
Nelson et al present a useful rapid-prototyping model for refining user requirement
for dashboards in a health care setting that will help inform our work as we adapt
the interface for these specific user populations.[27] To aid in usability evaluation, we will also incorporate Dowding and Merrill's dashboard
visualization heuristics, designed for evaluation of information visualizations in
a medical setting.[28]
Another important line of future research will involve expanding the ontology to include
a wider variety of relationships than the strict hierarchical relationships currently
present. In addition, we will explore automatic and semiautomatic methods to analyze
the forum text and classify research topics based on existing ontology terms, or by
expanding the current ontology. This will enable more rapid ingestion of additional
research topics, as well as labeling the full forum conversation via comments. In
addition, it may be fruitful to combine the visualization of unstructured data, as
presented here, with structured data from an EHR, such as the work of Rao et al.[8]
Although the current version of the tool was developed with data specific to the structure
of the CCFA forum content (e.g., questions, descriptions, and categories), much of
the structure should be generalizable across a wide range of online discussion forums
(e.g., questions, descriptions, and comments can map to discussion threads, and user
IDs are typically associated with discussion thread content). It may therefore be
useful to employ an abstracted forum-content structure, enabling the investigation
of these techniques across a wider range of discussion forums involving patient-generated
content. For example, previous work has applied ontologies to the analysis of self-help
forums for chronic kidney disease.[29] Combining such ontology-based text mining approaches with the interactive visualization
techniques described in this article could enable more effective exploration, analysis,
and dissemination of online forum data across a wide variety of patient populations.
