UNDERSTANDING MEDICINE 2.0

Social Network Analysis and the VECoN System

Sam Stewart and Syed Sibte Raza Abidi

NICHE Research Group, Dalhousie University, 6050 University Ave., Halifax, NS, Canada

Keywords:

Web 2.0, Social network analysis, Data visualization, Knowledge translation.

Abstract:

Web 2.0 provides new and valuable tools to the world of medicine, and Social Network Analysis (SNA)

can provide insight into how these communication networks function. This paper explores the potential for

SNA methods to explain the communication patterns in the Pediatric Pain Mailing List, including identifying

content experts and isolating potential subgroups of interest. These results are incorporated into VECoN, a

novel network visualization designed to improve the standard network exploration process by presenting the

network graphically and incorporating SNA statistics into the presentation.

1 INTRODUCTION

Experiential healthcare knowledge manifests in a va-

riety of modalities: clinical case studies, problem-

based discussions between clinicians, experience-

based insights, diagnostic heuristics et cetera. This

knowledge accounts for the intrinsic experiential

know-how, insights, judgements and problem-solving

strategies of healthcare practitioners. Such knowl-

edge is not ‘published’ as evidence-based, yet it

holds vital insights into solving atypical clinical prob-

lems. The key issues related to experiential healthcare

knowledge are: (a) how to formulate a community

of healthcare practitioners; (b) how to explicate and

share their experiential healthcare knowledge; and (c)

how to put value on experiential healthcare knowl-

edge, especially for clinical decision making, since it

is not systematically evaluated in the same manner as

evidence-based studies.

In the realm of Web 2.0, the emergence of

Medicine/Health 2.0 presents ‘virtual’ community-

driven environments to create and share healthcare

knowledge. The key idea is that the community cre-

ates and validates experiential knowledge in an or-

ganic manner, and applies it to provide feedback to

its effectiveness. Web 2.0 based knowledge sharing

mediums include online discussion forums, email-

based mailing lists, web blogs, et cetera. Through

these mediums, healthcare practitioners are able to

articulate and share clinical, operational and even

psycho-social experiences, along with insights and

know-how about particular healthcare topics. The ef-

ﬁcacy of this approach is that healthcare practitioners,

originating from different backgrounds and expertise

levels, can engage, collaborate and share their experi-

ential knowledge for the beneﬁt of the entire commu-

nity.

Given the virtual nature of the community it is

of interest to get insights about the knowledge shar-

ing dynamics (active participants, key contributors

of knowledge, topics of interest, inﬂuential mem-

bers, et cetera) of the virtual community, as it helps

to put a value on the knowledge created and shared

there. We argue that a study of the community’s

communication patterns and of the knowledge con-

tent being shared can provide insights into experi-

ential knowledge sharing dynamics of a specialized

healthcare practitioner community. Social Network

Analysis (SNA) allows us to analyze the communi-

cation networks within a socially-connected commu-

nity (such as an online healthcare practitioner com-

munity) and highlight the key actors, interest groups,

sub-networks, content seekers and experts, collabora-

tion opportunities, communication barriers, and other

network attributes. SNA focuses on analyzing the at-

tributes that arise out of the structural properties of a

social network, rather than the properties of the actors

themselves, providing an overview of the community

and how people communicate within it.

In order to make the results of the SNA useful

to the PPML community at large, a visualization tool

has been developed to allow the members to visually

navigate the network and explore both the conversa-

tions on the mailing list and the social connections

Stewart S. and Sibte Raza Abidi S..

UNDERSTANDING MEDICINE 2.0 - Social Network Analysis and the VECoN System.

DOI: 10.5220/0003167100700079

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2011), pages 70-79

ISBN: 978-989-8425-34-8

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Depiction of the experiential knowledge sharing framework and how SNA can leverage the knowledge sharing to

inform knowledge creation.

that have arisen out of it. VECoN (Visual Exploration

of Communication Networks) is a tool that provides

the means to visualize the social aspects of the net-

work, along with functions to navigate to individual

threads on the mailing list. A beta release of VECoN

has been produced as a proof of concept of how we

can provide further insights into the different uses of

Web 2.0 technologies in healthcare.

2 EXPERIENTIAL HEALTHCARE

KNOWLEDGE SHARING

FRAMEWORK

From a health knowledge management perspective,

an online mailing list or discussion forum provides a

collaborative learning environment in which domain

experts can disseminate their wealth of knowledge

and experience, and junior health practitioners can

learn by leveraging the explicated experiential knowl-

edge. This brings into relief an experiential knowl-

edge sharing framework (Figure 1) that allows a com-

munity of healthcare practitioners to interact and col-

laborate to create and share experiential knowledge,

while organizing the knowledge in terms of domain

speciﬁc topics. Each topic is pursed by a group of

practitioners who interact through online tools, such

as email or discussion forums. The discussion around

a topic can be organized in terms of a ‘discussion

thread’: a series of emails/posts by healthcare practi-

tioners around a speciﬁc topic. Using this framework

analysis of emergent social networks within the com-

munity can be performed (where the social network

depicts the collaboration/communication dynamics of

the community) and analysis of the content of the dis-

cussions can be done through the use of intelligent

text analytics and text/knowledge creation methods.

It is anticipated that subsequent SNA of the knowl-

edge sharing behaviour of the virtual community will

provide useful insights into the patterns of knowl-

edge ﬂow amongst healthcare practitioners. The re-

sultant social network will provide an objective mea-

sure of the effectiveness of the online knowledge shar-

ing medium to support collaborative learning.

The objective of the SNA is to provide meaning-

ful insight into the ﬂow of communication within the

network. This will be provided at both the micro and

macro level. Centrality measures provide insight into

the roles individuals play within the network, and will

help identify content experts, those members that are

actively communicating with their peers and facilitat-

ing knowledge transfer. Clique and structural equiv-

alence analysis is performed at the macro level, iden-

tifying subgroups of actors that are tightly connected.

The presence of subgroups may represent particular

sub-topics of interest, or groups of clinicians that are

not fully communicating with the rest of the network.

The social network analysis and visualization

framework is outlined in Figure 2. The following sec-

tions will outlined the process in detail.

3 BACKGROUND

3.1 Social Network Analysis

Social Network Analysis (SNA) is the analysis of

the relations between actors, i.e. understanding the

underlying social structure of a community of prac-

tice. SNA utilizes the principles of graph theory from

the world of mathematics to represent communication

networks in terms of actors (nodes) and ties between

actors (edges) (Wasserman and Faust, 1994) (Hanne-

man and Riddle., 2005). Traditional statistical anal-

UNDERSTANDING MEDICINE 2.0 - Social Network Analysis and the VECoN System

PPML

Archives

ThreadsActors

Social

Network

Group

Level

SNA

Information

User

Level

SNA

Information

Social

Network

Visualization

(VECoN)

Figure 2: SNA and visualization process. The network in-

formation is extracted from the PPML database, where it

is processed to produce the network. The network is used

to general social networking information, which is incorpo-

rated with the network to produce the VECoN system.

ysis focuses on actors as independent units, and ana-

lyzes them in terms of their personal attributes. SNA

instead focuses on the structures that emerge out of

the relations between actors, and not on the actors

themselves.

One aspect of SNA that is important to this project

is the difference between one and 2 mode networks.

In one-mode networks the nodes in the network are

homogeneous, i.e., they all belong to the same class.

This is the traditional network layout, in which nodes

represent people and ties represent some sort of so-

cial construct that connects them: friendship, advice,

work, et cetera. Two-mode networks contain two dif-

ferent classes of network, and ties exist only from one

mode to another. These most commonly occur when

one set of nodes represents people, and the second set

represents events, and ties go strictly from one mode

to another (indicating that a person has attended an

event). A two-mode network that represents people

and the events they attend is sometimes called an afﬁl-

iation network, and is the structure of the VECoN sys-

tem. The network represented in this project presents

both healthcare practitioners and discussion topics as

two independent classes of nodes, and the ties from a

practitioner to a topic indicates that that user has com-

municated on the connected topic.

Because the majority of SNA methods involve

one-mode networks, a common component of two-

mode network analysis is to transform the data into a

one-mode network. A new network is created of the

members of the network, and a tie is created between

them if they both communicate on the same thread. It

is preferable to analyze the two-mode network when

possible, as some information is lost in the transfor-

mation, but due to the existing body of literature for

SNA it is necessary to perform much of the analysis

on the one-mode network. Future research will be di-

rected at adapting methods to two-mode networks.

This paper will begin by analyzing the network at

the micro level, using centrality measures to identify

the active members of the network. It will then move

on to identifying potential subgroups of actors within

the network using clique and structural equivalence

analysis.

3.1.1 Centrality

The goal of centrality analysis is to recognize the most

important actors in the network; those actors that are

at the centre of the action in terms of communication

between individuals. Three different centrality statis-

tics are going to be presented; the actors identiﬁed

through these measures occupy key roles in the net-

work, and will be considered content experts.

Degree centrality is the simplest of the centrality

measures, calculating the number of ties one actor has

to the others. For the two-mode network this is the

number of threads each actor communicates on, and

for the one-mode network it represents the number of

other actors that actor has communicated with.

Closeness centrality extends the idea of degree

centrality beyond one step. Closeness centrality con-

siders an actor central to the network if they can reach

all the other nodes in the network in as few steps as

possible. The calculation of closeness centrality dif-

fers little between one and two mode networks. If

two actors are one-step away in the one mode net-

work, they are necessarily two steps away in the two-

mode network. Likewise, if two actors in the one-

mode network share a common partner, then each of

them must have shared a thread with that actor in the

two-mode network. Though the normalization is dif-

ferent (Borgatti and Everett, 1997) the general rank-

ing is the same. As such only one-mode closeness

will be presented.

Betweenness centrality deems nodes central if

they are hubs of information. Where closeness deems

a node central if it can quickly reach other nodes, be-

tweenness deems a node central if it is used as path

between other nodes. A node has a high between-

ness score if it falls in the shortest path between many

other pairs of nodes. As with closeness centrality be-

tweenness centrality does not signiﬁcantly differ be-

tween one and two mode networks (Borgatti and Ev-

erett, 1997).

HEALTHINF 2011 - International Conference on Health Informatics

The next step in the SNA is to try and isolate po-

tential subgroups within the network. With a topic

as large as pediatric pain there may be evidence of

subgroups in which actors are more active around a

certain topic of interest. If the network were broken

into groups, one would expect a lot of communication

within groups, and relatively little communication be-

tween groups.

3.1.2 Structural Equivalence

The goal of structural equivalence (SE) is to identify

nodes that occupy similar roles within the network.

Formally, two nodes are SE if they have the same ties

to all other nodes in the network. If two nodes are SE

then one can replace the other without interfering with

the ﬂow of information in the network. In reality true

SE is rare, so approximate SE needs to be measured.

A simple measure would be to count the proportion of

matching ties, or the number of tie changes required

to make two nodes SE. There are several measures

available, but for this project a simple count of the

number of similar ties is used.

Regardless of which SE measure is used, a SE ma-

trix is developed, which records the SE between all

the actors. This matrix is used to group similar actors

using a hierarchical clustering algorithm. The result

is a binary tree, or dendogram, depicting a hierarchi-

cal ranking of similarities, as in ﬁgure 7. Cutting the

tree off at a particular level results in partitions being

created from the clusterings. The red blocks in ﬁgure

7 represent the cutpoint at which the clusters are cre-

ated. Assigning the actors to these groups creates a

blockmodel.

A blockmodel is a partitioning of the network into

exclusive, non-overlapping groups, such that nodes

within the group are approximately SE. For a block-

model there tends to be a lot of communication within

the blocks and relatively little between them. Once

the optimal block model is determined the active

blocks can be further investigated to determine the

content that makes certain blocks unique.

3.2 Network Visualization

The visualization of networks is a key component

of SNA, and as such there is a rich literature base

describing methods of presenting networks visually.

Linton Freeman (Freeman, 1999) documents the his-

tory of social network visualization from a socio-

logical perspective, including theories on node lay-

out (both information-based and algorithmic theories)

along with the use of colour, size and shape to encode

network information. There are many current tools

for analytic network visualization, including UCINet

(Borgatti et al., 2002) and an extension for the R sta-

tistical language called statnet (Handcock et al.,

2003).

Previous work on social network visualization

has also been directed towards network navigation.

Examples include ContactMap (Nardi et al., 2002)

for identifying community groups within email con-

tacts, PieSpy (Mutton, 2004), which provides a real-

time visualization of social networks for Internet Re-

lay Chat (IRC) members, and Vizster (Heer and

boyd, 2005), a tool for exploring the Friendster

(www.friendster.com) social networking site. These

tools are all designed for 1-mode networks, for ex-

ample, the nodes in the Vizster program all represent

users of Friendster, and the ties represent friendship

links between them. In contrast, this project is visual-

izing a 2-mode network, where the ﬁrst class of nodes

represent mailing list members and the second rep-

resents threads, and the links between a node and a

thread indicate that a certain list member has commu-

nicated on that thread.

The software being used to implement this project

is the prefuse toolkit in Java (Heer et al., 2005).

Prefuse was chosen because it provides a full Java

library, and previous implementations of prefuse, in-

cluding the Vizster program, have proven successful.

4 VeCON System

4.1 Visualization

The purpose of the visualization is to ﬁrst provide a

tool for visually exploring social networks, and sec-

ondly to provide some insight into the underlying so-

cial structure of the network. This section will outline

the graph-theoretic layout decisions for the network,

and then explain the visualization tools implemented

to help the exploration of the mailing list.

4.1.1 Graph Structure

The network is laid out using a force-directed lay-

out, in which the nodes repel one another and the

edges act as “springs” that hold the nodes together.

Because of this spring effect, the layout is also

sometimes referred to as a “spring embedding” al-

gorithm. Prefuse implements the Barnes-Hut algo-

rithm (Barnes and Hut, 1986) which allows for real

time calculation of spring-embedding forces. The al-

gorithm is an iterative process, and following the lead

of Vizster, this project chooses to not limit the num-

ber of iterations of the algorithm, resulting in a visu-

alization in which the nodes migrate to their optimal

UNDERSTANDING MEDICINE 2.0 - Social Network Analysis and the VECoN System

positions but continue to move subtly. The effect is “a

living or ‘breathing’ feel, connoting social energy or

playfulness.”(Heer and boyd, 2005).

Two changes that need to be made to the spring

embedding algorithm are i) an adaptation to two-

mode networks, and ii) dealing with the problem of

components (disconnected sections of the graph). The

two-mode issue is addressed by Krempel (Krempel,

1999), in which he suggests ﬁxing the second mode

and allowing the ﬁrst mode to vary. To this end, the

algorithm is adjusted slightly: the ﬁrst step is to dis-

perse the threads evenly around the space, and then

allow the actors to move according to the spring em-

bedding algorithm. This method has proven to be ef-

fective, but a more formal evaluation of its efﬁciency

is required.

The issue of components is addressed in Kamada

and Kawai’s seminal work on force directed layouts

(Kamada and Kawai, 1989). The solution is to par-

tition the space according to the number of compo-

nents in the graph, with each graph being allocated

space proportional to its size (number of nodes). The

forces are then calculated only on the nodes within

the component. This means that components do not

effect each other in terms of layout, and avoids the

“drift” that is caused when disconnected components

continue to push each other away (see ﬁgure 72 in

(Fruchterman and Reingold, 1991)).

The colouring of the nodes is deﬁned by their

mode: blue indicates actors and red indicates threads.

Following Vizster’s lead, nodes and their neighbours

are highlighted when the mouse hovers over them:

when hovering over an actor that actor’s threads are

highlighted, and conversely when hovering over a

thread the actors that communicated on that thread are

highlighted. The highlighting is done by increasing

the saturation of the colour (see ﬁgure 3).

4.1.2 Exploration

The visualization provides several different explo-

ration methods, which will each be described in de-

tail. The objective of each of the tools is to provide a

different way of exploring the mailing list to retrieve

pertinent messages.

Hover Over. As mentioned before, the visualiza-

tion implements a hover-over feature. When the

mouse hovers over a node, that node becomes ﬁxed,

and it and its neighbours are “highlighted”. A node

is highlighted by increasing the saturation. Figure 3

demonstrates the difference between the regular and

highlighted nodes.

Along with changing the representation of the

(a) Regular. (b) Highlighted. (c) Focus Node.

Figure 3: Highlighting is achieved by increasing the satura-

tion from 25 to 50. A focus node is activated by clicking it,

at which point it doubles in size.

nodes, the hover-over feature presents additional in-

formation about the node/edge in question. For ac-

tors it presents the actor’s name in the top-right cor-

ner of the visualization, and for threads it presents the

thread’s subject line (see ﬁgure 8). The node names

were not put on the nodes in order to avoid cluttering

the visualization: current node labels are numerical

identiﬁers, provided in order to differentiate between

actors.

On the right side of the visualization is the con-

trol bar, where users can manage their search features

and control the factors that effect the spring embed-

ding. At the bottom of the control bar is a text box

that presents the detailed content of the selected node.

For actors a list of the threads they have communi-

cated on is presented, and for threads the conversation

is presented. When an edge is selected then that spe-

ciﬁc message is presented. The objective is to allow

the user to quickly browse through the conversations,

and upon ﬁnding the desired message to explore it in

more detail. Future work will pursue connecting the

visualization directly to the online discussion forum,

but as the forum is not yet developed this feature is

not implemented.

Point and Click. The objective of the point and

click option is to allow the user to explore a partic-

ular node in more detail. When a node is clicked dou-

bles in size, and the text associated with it becomes

ﬁxed. The user can then adjust the temporal ﬁlter to

see where that node ﬁts within the ﬁltered data. While

the node is in focus (i.e. while its size is doubled) the

text stays static, though the rest of the highlighting ef-

fects associated with hovering remain. A second click

on the node reduces it back to normal size and allows

the text to change freely again. Figure 3 demonstrates

what a focus node looks like.

Temporal Filtering. Adding a temporal ﬁlter pro-

vides a way to reduce the volume of information be-

ing presented. The temporal ﬁltering is accomplished

using a horizontal scroll bar, located within the con-

trol panel. The user can manipulate the lower-end or

HEALTHINF 2011 - International Conference on Health Informatics

upper-end of the bar to adjust the visualization, or can

set the bar to a speciﬁc window width and slide the

window itself. When the mouse is released the graph

is re-drawn with only those messages that are within

the window presented. The user can also manipulate

the dates manually by typing dates into the two date

boxes that are below the bar.

When the visualization is adjusted, only those

nodes that are connected to each other are presented,

reducing clutter by removing obsolete nodes. There

is still a potential for isolates in the ﬁltered network,

however. If an actor participates in a thread, but that

contribution is not made during the window, then the

edge will be removed but the actor will remain in

the network, and that contribution will still be pre-

sented in the hover-over effect. Future work should

explore how to incorporate the disconnected user into

the force-directed layout, but for now the user is as-

signed a separate segment of the space.

5 METHODS

5.1 Data

This project will use the archives of the Pediatric Pain

Mailing List (PPML), provided to the VECoN project

by the administrators of the mailing list. There are

over 13,000 messages in the PPML archives, dating

back to 1991; for this project a sample of the mes-

sages from 2007 and 2008 was used. The sample

has been parsed from simple ASCII text ﬁles and the

messages have been written into a MySQL database.

Along with sender, subject and date information about

each message, a thread number has been assigned to

indicate which messages combine to form conversa-

tions. The threads and senders will be used as nodes

in the network, with ties between them indicating that

a sender has communicated on a particular thread.

Currently the PPML administrators are in the process

of transferring the archives of the mailing list to an

online discussion forum. Once this forum is active

the VECoN system will be connected to it.

5.1.1 Social Network Visualization

Beyond simple network exploration the tool pro-

vides functions to explore the SNA dimensions of

the PPML. Within the control panel there are but-

tons to toggle centrality indicators and block mod-

elling. When activated the centrality indicators rede-

ﬁne the actor and thread nodes in the network, setting

the node size relative to that node’s degree. Currently

only degree centrality is implemented, but all three

centrality measures will eventually be added.

As well, the tool performs agglomerative cluster-

ing using the SE calculations from the blockmod-

elling. It is left to the user to decide how many blocks

to assign to the network, presenting the blocks as

“blobs” surrounding the component members. When

activated the blobs modify the spring embedding al-

gorithm to increase tensions within the blobs and

decrease tension outside them, resulting in tighter

groups of block members.

6 RESULTS

6.1 SNA

6.1.1 Micro Level Analysis

Table 1 lists the actors with the highest degrees in the

two-mode network, and table 2 lists the same for the

one-mode network.

Table 1: Actors with the highest degree centrality in the

two-mode network, i.e., the actors with the most number of

threads. The degree is normalized by dividing the degree

measure by the maximum possible degree.

Actor Degree NormDegree

771 31 0.1148

901 27 0.1000

782 26 0.0963

904 22 0.0815

920 20 0.0741

Table 2: Actors with the highest degree centrality in the

one-mode network, i.e., those actors that communicated

with many other members of the network. The degree is

normalized by dividing it by the max possible degree.

Actor Degree NormDegree

901 74 0.3776

904 60 0.3061

782 57 0.2908

855 56 0.2857

920 53 0.2704

From the degree centrality analysis there seem to

be several actors that are quite active. The maxi-

mum normalized degree of 0.115 in the two-mode

network indicates that actors are not participating in

many threads, and a normalized degree of 0.378 in

the one-mode network continues to demonstrate there

does not seem to be a single user in the network that

communicates with all other users. The histograms in

UNDERSTANDING MEDICINE 2.0 - Social Network Analysis and the VECoN System

Distribution of two−mode Degrees

two−mode degree

Frequency

0 5 10 20 30

0 50 100 150

Distribution of Actor Degrees

Actor degree

Frequency

0 20 40 60 80

0 20 60 100

Figure 4: Distrubtion of degrees for the PPML data. The

histograms indicate that there are not many actors with high

degree centrality.

Closeness in Actor Network

Closeness

Frequency

0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60

0 10 20 30 40 50

Figure 5: Distribution of closeness scores for the PPML

data. The closeness scores follow a normal distribution.

ﬁgure 4 conﬁrm that suspicion; the majority of actors

in both the one- and two-mode networks have low de-

gree centrality measures.

Table 3 lists the highest closeness scores in the

network. The actors that have the highest closeness

Table 3: Closeness scores for the PPML network.

Actor Closeness

901 0.5957

904 0.5665

920 0.5552

782 0.5521

771 0.5506

scores are the same actors with high degree scores.

With a max closeness of 0.596 the network seems

to have a very high closeness measure. This means

that it is easy to get from any node to any other node.

These ﬁndings are conﬁrmed in the histograms in ﬁg-

ure 5. This histogram seems to follow a normal distri-

bution, centred at 0.419 with a standard deviation of

0.064, meaning that 97.5% of the members of the net-

work have a closeness centrality greater than 0.292.

Table 4 lists the highest betweenness scores from

the network. Once again, the same actors are promi-

nent for the betweenness calculation. The low maxi-

mum betweenness of 0.134 indicates that the network

does not depend on any particular actor to facilitate

Table 4: Betweenness scores for the actor network. As with

closeness, betweenness measures the same thing in the 1

and 2 mode networks.

Actor Betweenness

901 0.134

904 0.103

920 0.072

782 0.066

855 0.063

Distribution of Actor Betweenness scores

Normalized Betweenness

Frequency

0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14

0 50 100 150

Figure 6: Betweenness scores for the PPML network data.

communication. The histogram in ﬁgure 6 conﬁrms

that ﬁnding, with the majority of users having low be-

tweenness.

The centrality scores for the network indicate a

healthy and active knowledge sharing network. Low

degree centralities mean that there is not a single user,

or a set of users, that are required to initiate conver-

sation. High closeness scores mean that it is easy to

connect one user to another, either through a thread

they have both communicated on, or a short series

of “mutual friends”; this is key to facilitating knowl-

edge transfer, to allow users to connect directly to

the source of experiential knowledge in as few steps

as possible. Low betweenness scores compound this

ﬁnding by demonstrating that, when trying to connect

with someone through a series of users, there are al-

ways multiple paths available, without the need to al-

ways communicate through a single user. These re-

sults combine to describe a knowledge sharing net-

work that has a large number of healthy communica-

tors and an active user base.

Though the centrality results do not demonstrate

the presence of any dominant users, it is noteworthy

that the same actors appear at the top of all four sets of

centrality measures. Table 5 lists the top ranked actors

for each of the four centrality measures, and the same

four actors appear in each list. If further investiga-

tion into the content of the mailing list were required,

these actors should be the ﬁrst to be contacted, as they

seem to be the most active and most important to the

communication network.

HEALTHINF 2011 - International Conference on Health Informatics

Table 5: The ranking of actors provided by each of the three

centrality measures.

Degree Degree.2M Closeness Betweenness

901 771 901 901

904 901 904 904

782 782 920 920

855 904 782 782

920 920 771 855

125

127

115

192

188

110

193

194

191

190

187

189

178

180

138

134

171

108

139

154

161

197

140

196

185

186

149

173

163

167

172

102

195

164

174

104

105

109

143

182

130

103

128

132

145

148

141

131

121

122

144

181

184

117

118

175

176

100

120

107

126

156

158

151

142

153

137

152

135

162

177

129

136

169

170

183

179

116

133

168

157

146

160

111

114

112

113

159

147

155

166

165

119

150

106

101

124

123

Hierarchical Clustering of the Actor Network

Figure 7: A hierarchical clustering of the actors in the one-

mode network using structural equivalence. The red blocks

indicate the points at which the tree is cut to create the

blockmodel.

6.1.2 Macro Level Analysis

For the PPML the blockmodel contains three groups,

two small groups with 15 and 22 members, and one

large group with 160 members (ﬁgure 7 contains the

dendogram). The image matrix produced by the block

model is available in table 6. An image matrix is a

table that represents the communication densities be-

tween blocks. This image matrix demonstrates that

there is a lot of communication within the two smaller

blocks, and a bit of communication between them

(0.183 density). For the large group in block 1, how-

ever, there is little communication, either within or

between blocks.

Table 6: Communication Densities Within and Between

Blocks.

Block 1 Block 2 Block 3

Block 1 0.04796 0.07958 0.06932

Block 2 0.07958 1.00000 0.10000

Block 3 0.06932 0.10000 1.00000

SE provides a partitioning of the network into

groups of “similar” actors. For the PPML it provides

three separate groups: two small, active groups and

one large, relatively inactive group. These groupings

provide some insight into the communication patterns

between users, and are of keen interest in determin-

ing the presence/absence of subgroups of interest in

the network. The small groups may represent an

active sub-group of experiential knowledge sharers,

who actively communicate with all other members of

the group. Further investigation of the content of the

communications within these two groups is required.

6.2 Visualization

Figure 8 is a capture of the VECoN system. The left

pane contains the network visualization, and the right

pane contains the spring-embedding controls, the ﬁl-

ter controls and the message pane where the contents

of the threads or the participating threads are pre-

sented.

7 DISCUSSION

The SNA provided some useful insight into the com-

munication patterns within the network, but further

analysis is required to fully ﬂush out what the results

mean in the scope of the project.

7.1 Isolates

There were 68 isolates recognized in the actor net-

work. These were actors whose messages received

no response on the mailing list. A further investiga-

tion reveals that some of these actors posted few times

to the network; 13 only posted to the PPML once.

Looking through the subject lines, there are several

messages that are not meaningful pediatric pain com-

munications (such as incorrect subscribe/unsubscribe

messages and job or conference announcements).

However, there are also meaningful messages which

received no response. There does not seem to be a

pattern to which messages were ignored, and after ﬁl-

tering the spam messages there were only nine mean-

ingful queries that were left unanswered. This is a

positive ﬁnding for a mailing list, as it is evidence

that knowledge seeking queries are being answered

most of the time. Moving forward it is imperative that

the PPML continues to be an active community by

incorporating new members and responding to their

queries.

UNDERSTANDING MEDICINE 2.0 - Social Network Analysis and the VECoN System

Figure 8: The PPML visualization. Note that the current network is restricted to messages between 2007-01-02 and 2007-03-

21.

7.2 Centrality

The centrality analysis has indicated that there is not

a central user or set of users that control the com-

munications on the network. For each of the four

measurements the highest scoring actors do not have

disproportionately high normalized values, indicating

that there is not one actor that communicates on ev-

ery thread, or connects disparate groups of actors.

Though the low degree centrality scores could be in-

terpreted as members being inactive, a more reason-

able interpretation is that there are many messages on

the mailing list that are spawning conversations be-

tween different actors. This is a very promising result

moving forward.

Detecting content experts or knowledge shar-

ing activities strictly using graph-theoretic principles

such as centrality analysis poses problems. If some-

one contributes to the list by asking many questions,

without providing answers, they are recognized as

central users, but should not be considered content

experts. In order to improve the prediction of con-

tent experts it is necessary to extract the underlying

semantics of the messages being communicated. Pre-

vious work, (Stewart et al., 2010), has worked on

extracting semantic representations of the messages,

and incorporating this information into the SNA could

help differentiate the content experts from the “ques-

tion askers.”

7.3 Structural Equivalence

The structural equivalence analysis broke the model

into 3 separate groups: two with high density commu-

nications and one with low density communication.

Further analysis into the content of the two groups re-

veals little to identify the two blocks, and there is little

to suggest that there are any speciﬁc sub-specialities

of interest within these two groups. Further investi-

gation should pursue other measures of SE, including

incorporating the 2-mode nature of the data.

7.4 Visualization

A beta release of the VECoN system has been pro-

duced, but it has not yet been tested by the PPML

members. Once the mailing list is available as a dis-

cussion forum and links between the visualization and

the forum are established the VECoN system will be

rolled out as a Java applet, and at that point research

will be conducted on its utility and on future addi-

tions. Currently future research is being conducted

to improve the spring-embedding algorithm and the

blockmodelling algorithm, along with adding more

centrality measures and a more intuitive control panel.

8 CONCLUSIONS

Experiential healthcare knowledge is a vital compo-

nent of the current healthcare system, and developing

new methods to facilitate the sharing of this knowl-

edge is vital to sustaining and improving the medi-

cal community. Medicine 2.0 technologies provide

online tools for facilitating knowledge sharing, es-

tablishing virtual communities of clinicians. Under-

standing the ﬂow of knowledge in these virtual com-

munities is key to developing new systems, and SNA

provides the necessary tools for understanding the

HEALTHINF 2011 - International Conference on Health Informatics

ﬂow of communication within the network. It has

provided a list of potential content experts within the

list, it has recognized several active subgroups, and it

has partitioned the network into disparate groups of

potentially different clinicians.

Further research should be directed at better un-

derstanding the members of the community. More

actor attributes, such as specialty, location, job de-

scription, et cetera, would provide better insight into

the structure of the network, and in particular into

the structure of the subgroups revealed through block-

modelling. Incorporating the semantic information of

the communications themselves would help differen-

tiate between acts of knowledge seeking and knowl-

edge sharing, and improve our overall understanding

of the experiential knowledge available in the net-

work.

Though the VECoN project is only in its beta

stages, preliminary results are promising. The net-

work has been visualized, and SNA tools have been

added. Future work will be on adding new SNA tools,

improving the visualization, implementing the sys-

tem online and making the interface more intuitive.

The ultimate goal of the VECoN system is to provide

a novel network exploration tool to help users make

new connections within the PPML community, and

ﬁnd new sources of experiential knowledge already

available in the network.

ACKNOWLEDGEMENTS

This work is carried out with the aid of a grant from

the International Development Research Centre, Ot-

tawa, Canada. The authors would like to acknowledge

Dr. Allen Finley for his contributions to the PPML

and his ongoing support of this research.

REFERENCES

Barnes, J. and Hut, P. (1986). A hierarchical o(n log n) force

calculation algorithm. Nature, 324.

Borgatti, S., Everett, M., and Freeman, L. (2002). Ucinet for

windows: Software for social network analysis. Har-

vard, MA: Analytic Technologies.

Borgatti, S. P. and Everett, M. G. (1997). Network analysis

of 2-mode data. Social Networks, 19:243–269.

Freeman, L. C. (1999). Visualizing social networks. Web.

http://www.cmu.edu/joss/content/articles/volume1/

Freeman.html.

Fruchterman, T. M. J. and Reingold, E. M. (1991). Graph

drawing by force-directed placement. Software-

Practice and Experience, 21:1129–1164.

Handcock, M. S., Hunter, D. R., Butts, C. T., Goodreau,

S. M., and Morris, M. (2003). statnet: Software

Tools for the Statistical Modeling of Network Data.

Seattle, WA. Version 2.1 . Project home page at

http://statnetproject.org.

Hanneman, R. A. and Riddle., M. (2005). Introduction

to social network methods. University of California,

Riverside, Riverside, CA.

Heer, J. and boyd, d. (2005). Vizster: Visualizing online

social networks. InfoVis.

Heer, J., Card, S., and Landay., J. (2005). prefuse: A toolkit

for interactive information visualization. In CHI.

Kamada, T. and Kawai, S. (1989). An algorithm for draw-

ing general undirected graphs. Information Process-

ing Letters, 31:7–15.

Krempel, L. (1999). Visualizing networks with spring em-

bedder: Two-mode and valued data. American Statis-

tical Association, Proceedings of the Section of Statis-

tical Graphics, pages 36–45.

Mutton, P. (2004). Inferring and visualizing social networks

on internet relay chat. InfoVis, pages 35–43.

Nardi, B., Whittaker, S., Isaacs, E., Creech, M., Johnson,

J., and Hainsworth, J. (2002). Contactmap: Integrat-

ing communication and information through visualiz-

ing personal social networks. Communications of the

ACM, 45(4):89–95.

Stewart, S., Abidi, S., and Finley, A. (2010). Pediatric

pain management knowledge linkages: mapping ex-

periential knowledge to explicit knowledge. Studies in

Health Technology and Informatics, 160:1184–1188.

Wasserman, S. and Faust, K. (1994). Social Network Analy-

sis: Methods and Applications. Cambridge University

Press.

UNDERSTANDING MEDICINE 2.0 - Social Network Analysis and the VECoN System