Social Network Construction and Analysis in Healthcare
RedaAlhajj, Jon Rokneand et al.
Department of Computer Science, University of Calgary, Calgary, Alberta, Canada
Synonyms
social interactions, interaction network, data mining, knowledge discovery, healthcare solutions, medical referral.
Glossary
Graph: a set of nodes and edges connecting the nodes.
Network:a graph that assigns some semantics to the nodes and kind of interaction for the links.
Matrix: a grid with predefined rows and columns.
Pattern: a sequence of symbols.
SNA: (social network analysis) is to study the characteristics of a social network.
GP: General practitioner.
SP: Specialist practitioner.
HIS: (hospital information system).
Data mining: extracting implicit information from a domain.
Medical referral: transferring a patient to another doctor who may understand his case better.
Centrality: refers to a set of SNA metrics.
Definition of the Subject
One of the important processes in the health care domain in countries such as Canada is the medical referral process. It is the process of referring patients to doctors with a practice specialty[20]. We have studied the referral process being used by Alberta Health Services (AHS) in Calgary. When a patient has a health problem and he/she visits a General Practitioner (GP) the GP may decide that the problem of the patient requires a more specific expertise. Accordingly, the GP will refer the patient to a specialist (SP) who is specializing in the patient’s type of illness or has special expertise for performing a particular procedure. In Canada the GP additionally functions as a gatekeeper for SP’s in that access to SP’s is only by referral.
The larger and more complex the integrated health care system is, the more complicated and, hence, important, the problem of finding an appropriate specialist becomes [22]. Partly due to this complexity there is frequently an unacceptable delay in being admitted to a specialist consultation.
Calgary has one of the largest integrated health care systems in Canada serving 1.2 million people. More than 25000 staff members and 2200 physicians provide services at over 100 locations. The method used to find the appropriate specialist in Calgaryif the GP has no previous knowledge of the right SP, is either by: contacting different specialists, searching in previous referrals records and try to match the symptoms, or consulting with any other GPs who might be knowledgeable about some specialists. This causes the following problems: (1) A delay in the referral process especially if the GP has just established a practice. This is due to the problem of eliciting information from established GP’s. (2) A GP might not refer patients to a specialist with the right skills thus requiring further visits to the GP. (3) Patients might be referred to an appropriate specialist, but the waiting time is excessive and there might be other specialistsavailable with the right skills and shorter waiting times unknown to the referring GP.
Introduction
One of the main goals of SNA is to study interactions between actors, typically people as individuals or groups in order to improve organizational role and structure especially for organizations that depend on collaboration. Recently, many researchers have added data mining techniques as a preprocessing step to the SNA. SNA and mining as an example recently been used in the health care domain to improve the quality of primary care as reported byFattore et al. [12]. They conducted a study using SNA techniques to provide an evaluation of the effects of organizing GP’s into networks on their behavior with respect to prescribing drugs.
In this article, we apply SNA and data mining techniques to solve some of the issues occurring in the medical referral process. Three social networks of GPs and SPs are developed and then analyzed to discover doctors’ communities and hidden patterns that will be useful to solve the issues relating to the process.
A SNA generally benefits from data mining techniques as a pre-processing step to prepare the social network data. The main data mining technique to be used here is clustering which is the assignment of a set of observations into subsets (called clusters) so that all the observations in the same cluster are similar in some sense. In Anderson & Jay [1] SNA and data mining techniques were applied to study the relationship between physician networks and the utilization of computer based hospital information systems (HIS). They concluded that physicians were being influenced to use the new HIS because of the strong social relationship between the physicians through the networks.
Key Points
From a social (time lost) and economic (cost in staff time to establish a referral) point of view it is desirable to improve the referral process. In this paper, the focus is on speeding up the referral process in the consultations stage between physicians when there is a need for a referral. Since GPs face problems in finding the right peers to consult with for a referral, we have introduced a model to discover peers’ communities who can be recognized as the significant consultation group. In other words, using a dataset of previous referrals, we discover peers who could be targeted by other GPs for referral consultations. In addition, we analyze the social network of SPs to discover hidden patterns that are hard to discover by looking at the raw data. These patterns relate to disease discovery. We then, analyze the bipartite graph that contains nodes as GPs and SPs. We analyze this graph in two different models: 1-mode and 2-mode network models. The outcome of this analysis is to predict future relations between GP’s with other GP’s, SP’s with other SP’s and most importantly GP’s with SP’s.
Historical Background
Interactions between medical doctors allows for the transfer and share of experience to the benefit ofpatients. These interactions have been studies by researchers [1] and they are an essential part of local and global healthcare systems. Developing systematic methods to study physicians networks in order to elicit information on the interactions has always been the aim of these studies. The traditional way of eliciting the information was adhoc and manual since it was mostly based on snail mail, telephone and fax.
The referral process is defined by Shortell & Anderson [1] as the permanent or temporary transfer (including sharing) of responsibility for a patients care from one physician to another. The reason for a referral is in the majority of cases that a receiving physician is more experienced and specializing in the patient’s illness than the referring physician. Patients are most often referred by a GP to aSP, but it is also possible that patients are referred by a SP to another SP, by a GP to another GP or by a SP to a GP.
Proposed Solution and Methodology
In this section, we introduce our proposed solution which consists of three main components: (1) Social network analysis of General Practitioners to elicit information about their referral experiences, (2) Social network analysis of Specialists to find out various parameters that are useful for the referral process and (3) Social network analysis of General Practitioners and Specialists combined together to find the most effective modes of communication and cooperation for a referral.
Social Network Analysis of General Practitioners
New GPs often encounter difficulties in finding appropriate information about specialists to whom they may refer cases needing expert consultation or treatment. In this section, we alleviate this problem by designing a social network for general practitioners to study the relationship between general practitioners in terms of making consultations with each other about referrals. Analyzing this social network later on will allow us to identify the most powerful general practitioners group that has been sought for referral consultations. In other words, this group will be considered as the most knowledgeable group of specialists in the community since they are consulted by most of the other GPs seeking informed opinions about referrals. This identified core group may then be targeted by new GPs for information on referrals. This analysis may potentially save a significant amount of time because the GPs will not have to keep consulting with random GPs who are not familiar with the possible specialists in their area.
We use the following steps to visualize and then analyze the social network of general practitioners:
Step 1. We use a previous year’s worth of data on the relationships between physicians as represented by an incidence matrix where rows and columns represent GPs, and the matrix data is either one or zero which indicate whether a GP has consulted with another GP regarding a referral to a specialist or not.
Step 2.We define a threshold Θ= 6 and consider only those GPs who has been consulted >= Θ times. The reason for this threshold is because after looking at the overall data, we realized that those GPs who had been consulted less than six times (average) will not be significant as information sources. Thus, we consider them as noise. We have chosen the threshold value after studying the distribution of consultations between physicians.
Step 3. We then apply a k-means clustering algorithm using MATLAB on the incidence matrix to find five sub-groups of GPs who have been consulted in similar manners. The number of clusters was chosen after trying different numbers and since the k-means does not always result in a reasonable output, we had to check the clustering assignment manually to make sure every cluster has the correct GPs members as based on referral pattern similarities.
Step 4.For each sub-group, we find the consultation relations among each pair of GPs in the internal sub-group as well as with the external sub-groups. In other words, we explore the consultation patters of which GP has consulted with another GP. This means members of group1 may consult with group2 or members of group1 may consult with members of the same group.
Step 5.Since our study is based on the previous year we extract twelve consultation matrices where each matrix corresponds to a previous month to strengthen the analysis. We repeat Step 4 for each of these matrices.
Based on this step, we visualize the social network for each month where nodes represent the groups resulting from the clustering process in step 3 and edges connect two nodes if any member of a group has consulted with another member from the other group.
Step 6.Finally, we calculate the indegree centrality measure for each SN and then find the average of these indegree centrality measures of the twelve matrices to identify which group is the central group based on the average.
To validate our result and analysis about which group is the most significant based on the whole network structure, we define four consultation indices as following:
These calculations will allow us to elicit which group is the most powerful in receiving consultation requests. This will in turn validate our findings using the indegree measure.
Consultation index1:shows the interaction behavior in seeking consultations from other groups among the all the nodes in the network. For each group, we calculate the value of the consultation index1 by dividing the number of consultations sought by this group by the number of overall sought consultations for all the groups in the network.
Consultation index2:is similar to the previous index, but here we consider the consultation requests coming to the group. This index shows the status of receiving consultations of each group comparing to the overall network.
Consultation index3:In this index, we want to see the interaction between the groups based on the comparison of the consultations coming to the group to the sum of the total consultations coming to the group and the total consultations made by the group. With this index, we can know the overall picture of how active the group is in terms of receiving and making consultations.
Consultation index4:studies the interaction between the group members themselves when they seek consultations from members inside the group. We divide the internal consultations by the total internal consultations of all the groups.
Social Network Analysis of Specialists
In this section, wish to identify and analyze the social network of specialists. This step will aid in finding some hidden patterns that will be helpful for the medical referral process. The nodes in this network are specialists and there is an edge between two nodes if they have received referrals from the same GP. Hence if a specialist, say ‘A’, has common GPs with all the nodes (specialists) in the network, it means that specialist ‘A’ will have connections with every other node in the network. This will result in his node appearing in the center of the network. Clearly, the more connections a specialist has, the busier he/she becomes. Furthermore if a specialist is busy only in a specific time period provides a hint that patients suffered from the particular disease for which the specialist is an expert in that period. A more detailed explanation of the disease discovery (i.e. epidemiology) is provided in the next section.
For the specialists’ social network, we use adjacency matrix, i.e., each specialist is assigned to both a column and a row in the matrix. In this case, the matrix will have two cells representing the fact that two SPs, say ‘A’ and ‘B’, have been referred to by the same GP. The two cells marked 1 for this case are in the intersections of column(A) and row(B) and the intersection or column(B) with row(A) to reflect that the two specialists have been referred to by at least one common GP. The produced matrix will be symmetric. In case the number of referred patients is needed the value in the entry will reflect the actual number of common referrals to contribute to the strength of relationship between the two specialists.
To visualize and analyze the social network of specialists, the following steps are used:
Step 1. We extract the adjacency matrix of the specialists based on the referral patterns of three different recent periods.
Step 2. We define a threshold Θ1= 7 and consider those SPs who only received at least a Θ1 number of common referrals. The reason for this is that specialists who received less than Θ1 common referrals would not be very socially effective in the network. The value Θ1 was chosen after studying how referrals are distributed among the SPs.
Step 3. To construct the SN of specialists, we calculate the number of the GPs who made referrals to each SP. A node represents a SP and a link is associated between two nodes if they have some GPs in common. The benefit of this analysis comes from the fact that if a SP has links with many SPs in the network, it means that this SP receives referrals from many different GPs and, hence, will be identified as a popular specialist in receiving referrals in the health region. The most popular specialist will always appear in the center of the network.
Step 4. We analyze the social networks in two different views: Visualization and Calculation views.
- Disease Discovery:
We can use SNA to discover the emergence of new diseases (epidemiology) in the community that might alert the health authorities that there is an emerging health issue. We first construct the social network of specialists for different recent periods (every month). The method used here is similar to the one discussed in “SN of Specialist” section except that the data set contains the referral patterns for those successful referrals. In other words, we consider only the data for patients who have been seen by specialists. We analyze the different networks for the different extracted periods by finding the center specialist in every network. If a specialist of the same area appears in the center in a particular month, this means, there is a significant need for each one of these specialists in different periods; this tells us that the community experiences specific diseases in each of the three different periods which makes some specialists too busy. This analysis is helpful for the health care system as it could be used a disease discovery tool. For instance, if cardiology specialists are is busiest specialist in terms of receiving referrals in a specific period, it indicates that many people have heart problems in that period and this could be caused by new factors such as diet, pollution etc. It could possibly result in a health care region issuing a bulletin warning people of the disease causing nature of the diet, pollution etc.
General Practitioners and Specialists Social Network
In the previous sections, we have considered the general practitioners social network and the specialist social network in isolation. Now, we are combining both networks together into one network and analyze it to find some hidden information that will be helpful for the medical referral system. The hidden information could be predicting the consultation patterns.
For this section, we use the following steps:
Step 1: When we draw the SN of the referral matrix, all what we get is a bipartite graph that does not provide us with much information that might be helpful in analyzing the consultation relationship. Instead, as the first step, we transpose the original referral matrix. Transposing (rotating) the matrix will switch the columns and rows of the original matrix.