Using network analyses to characterise Australian and Canadian frequent attenders to the emergency department

To explore and compare the characteristics of frequent attenders to the ED at an Australian and a Canadian tertiary hospitals by utilising a network analysis approach.


Introduction
Frequent attenders to the ED are a vulnerable population with physical health, mental health and socioeconomic needs, 1 experiencing a greater likelihood of mortality compared to non-frequent attenders. 2 Although frequent attenders make a disproportionate number of ED visits and are associated with significant healthcare expenditure, 3 their needs are often unmet. 1 Many interventions that have aimed to reduce healthcare utilisation by frequent attenders have been unsuccessful. 4 This complex population is heterogeneous, limiting the effectiveness of generalised interventions. Frequent attender subgroups need to be sufficiently identified to create highly tailored interventions that address their specific needs. 1 Few studies have detailed the relationships between frequent attender visit • Younger, middle-aged and older frequent attenders had high rates of mental illness/ substance use, a combination of mental and physical illnesses, and physical illnesses, respectively. • A network analysis approach revealed frequent attender subgroups unique to the Melbourne and Toronto hospitals.
characteristics. Such relationships can be easily visualised using a novel approach, the network analysis, enabling the identification of important subgroups.
Additionally, there is little research directly comparing frequent attenders in different countries. The comparison of frequent attenders across existing studies is limited by the wide variation in designs and frequent attender definitions. Such a comparison provides an opportunity to determine the generalisability of interventions in different jurisdictions and explore the influence of systemlevel factors on frequent ED use. Our primary aim is to characterise frequent attenders at an Australian and a Canadian tertiary hospitals by utilising network analyses.

Study design and patient population
We conducted a descriptive, retrospective cohort study using patient data from a major hospital in Melbourne, Australia, and from Toronto, Canada. All adult patients who visited the ED between 1 January and 31 December 2019 were included in the study. As both hospitals do not have dedicated paediatric wards, we excluded patients under the age of 18 years. Visits that had missing data were excluded from the analysis. Our study was approved by the Sunnybrook Research Ethics Board (project number: 1516) and Alfred Hospital Ethics Committee (project number: 631/19).

Study setting
The hospital in Melbourne is an academic, tertiary referral hospital that has over 69 000 annual emergency presentations. 5 Its catchment area includes metropolitan Melbourne, offering a range of tertiary-level services, including trauma, cardiovascular, oncological and psychiatric care, but does not provide inpatient obstetrics and gynaecology services. 5 Its ED has a short stay unit that allows short-term observation and treatment for selected patients. The hospital in Toronto is also an academic, tertiarylevel hospital, with more than 61 000 emergency visits per year, providing specialty services in trauma, interventional cardiology, stroke, oncology, neurosurgery, psychiatry and highrisk obstetrics and gynaecology. 6 It receives patients within the Greater Toronto Area and nearby regions. In contrast to the Melbourne hospital, the Toronto hospital's ED does not contain a short stay unit.

Data sources
At the Melbourne hospital, we used administrative data from the Cerner electronic health record. At the Toronto hospital, we used data collected from the ED Information System and the National Ambulatory Care Reporting System.

Outcome measure
To identify frequent attenders, we counted the number of ED visits each patient made, during the 364 days prior to their last ED visit in 2019. If their total number of visits (including their last visit) was four or more, they were defined as a frequent attender. This is consistent with the most common definition of frequent attendance in the literature. 7

Data analysis
All analyses were undertaken using visit-level data. Stata (version 16.1; StataCorp, College Station, TX, USA) was utilised to generate descriptive statistics. Data points were collected in broad categories with regard to patient demographics, referral status and arrival mode. The Australasian Triage Scale (ATS) was utilised at the Melbourne hospital and the Canadian Triage Acuity Scale (CTAS) at the Toronto hospital. The ATS and CTAS are five-level triage scales, with category one and five representing the highest and lowest acuity presentation, respectively. 8,9 We considered a visit to be triaged at high acuity if the triage level was level 1, 2 or 3, and low acuity if the triage level was 4 or 5. We utilised International Classification of Diseases (ICD-10-AM) 10 codes at the Melbourne hospital and ICD-10-CA 11 codes at the Toronto hospital to identify primary diagnoses.
We conducted network analyses on the younger, middle-aged and   older frequent attender groups at both hospitals using R (JASP 0.14.1 interface). 12 The R-package qgraph 13 was utilised to visualise the networks. In a network analysis, the variables are depicted as nodes (i.e. circles) and the pairwise associations or connections between variables are called edges (i.e. lines). In the present study, the blue and red edges represent positive and negative correlations, respectively. Thicker edges indicate stronger relationships between nodes. We used partial correlation coefficients to determine the associations between nodes after controlling for all other variables. As a result, the edges linked to a node are equivalent to the outcomes of a multiple regression analysis. 14 Further explanation of the network analysis methodology is detailed in Appendix S1 and Tables S1-S3.
We performed a sensitivity analysis for the network analysis on younger frequent attenders with highly related diagnoses. To assess the accuracy of edge weights and test how often an edge was recovered, we performed non-parametric bootstrapping 15 with 2500 iterations for each network. We used the Rpackage bootnet 16 to generate edgeweight accuracy plots with 95% confidence intervals.

Results
In 2019, 51 002 and 46 024 individuals visited the EDs of the hospitals in Melbourne and Toronto, respectively. For each patient, we counted the total number of visits during the 364 days prior to their last visit. The sum of these visits for all patients was 79 164 at the Melbourne hospital and 68 365 at the Toronto hospital. After determining the number of frequent attenders and their visits, we excluded visits with missing data and patients under 18 years. At the Melbourne hospital, 5.7% of 51 001 patients were frequent attenders with 17 475 visits, and at the Toronto hospital, 4.7% of 46 013 patients were frequent attenders with 11 573 visits ( Fig. 1). Bimodal distributions of frequent attenders' age were observed at both hospitals. We defined the younger age group (18- Abdominal pain (2.5) 39 years) to incorporate the first peak, the older age group (70 years and older) to include the second peak, and the middle-aged group to include individuals aged between 40 and 69 years. The three age groupings used in the present study incorporate the average ages of three frequent attender subgroups identified in a recent cluster analysis. 17 At the Melbourne hospital, 29.7%, 41.6% and 28.7% of 17 475 visits were made by younger, middle-aged and older frequent attenders, respectively (Table 1). This differs from the Toronto hospital, where 18.9%, 40.7% and 40.4% of 11 573 visits were made by younger, middle-aged and older frequent attenders, respectively. The percentage of visits that were selfreferred and triaged at high acuity did not differ considerably across all age groups for frequent attenders at the Melbourne hospital. At the Toronto hospital, however, younger frequent attender visits had high rates of self-referral compared to the other age groups and the proportion of visits triaged at high acuity increased with age. At both hospitals, the ambulance use rate increased with age and the rate of visiting after hours decreased with age. Middle-aged frequent attenders at the Melbourne hospital and younger frequent attenders at the Toronto hospital had the highest percentage of left-without-being-seen (LWBS) visits.

Younger frequent attenders
Bootstrapping plots (Figs S1-S4) demonstrated that edge weights were stable and accurate. For younger frequent attenders, the top 10 diagnoses included several related to mental illness and substance use. At both hospitals, acute alcohol intoxication was associated with arriving by ambulance and visiting after hours. Symptoms/ signs involving emotional state and multi-drug intoxication were associated with ambulance use at the Melbourne hospital. At the Toronto hospital, the abdominal pain node connects the male node with a red edge (inferring an association with being female) and the self-referred node with a thick red edge (inferring a strong association with being referred from another service). This suggests that frequent attenders with abdominal pain were female and often referred from another service. LWBS visits were associated with after-hours visits at both hospitals and with low triage acuity only at the Melbourne hospital, given the red edge between the LWBS and high triage acuity nodes (Fig. 2). Because mental illness and substance use are highly comorbid, 18 we performed a sensitivity analysis combining the former into a single diagnosis (Table S4). At both hospitals, we found a strong association between mental illness/substance  use and being female, having a high acuity triage, and visiting after hours (Fig. S5).

Middle-aged frequent attenders
For middle-aged frequent attenders, the top 10 diagnoses included a mixture of physical illnesses, behavioural disorders and substance use. Notably, the diagnoses of joint pain and cellulitis in the limb were associated with low triage acuity at both hospitals. Unique to the Melbourne hospital were the associations between symptoms/signs involving emotional state and high triage acuity as well as arrival by ambulance. At both hospitals, the LWBS visits were associated with being male, after-hours visits and low triage acuity (Fig. 3).

Older frequent attenders
For older frequent attenders, the top 10 diagnoses were mainly physical illnesses. At both hospitals, high triage acuity was positively correlated with diagnoses such as congestive heart failure and dyspnoea. However, compared to the Melbourne hospital, the Toronto hospital had a greater number of associations between high triage acuity and diagnoses such as malaise and fatigue. A stronger relationship between congestive heart failure and high triage acuity was observed at the Toronto hospital. At both hospitals, cellulitis of the limb was associated with low triage acuity. The LWBS visits were associated with low triage acuity only at the Melbourne hospital and after-hours visits only at the Toronto hospital (Fig. 4).

Interpretation
The present study observed clinically important differences in frequent attender characteristics across the age groups and hospitals. However, general diagnostic themes emerged at both hospitals, namely mental illness and substance use in the younger group, and a combination of substance use and physical illnesses in the middle-aged group. Older frequent attenders were strongly characterised by physical (including chronic) illnesses.
The network analysis is a novel technique that has recently been used to study complex relationships in ED patients. 19 To our knowledge, this is the first study to apply a network approach to explore the characteristics of frequent attenders. It allowed the heterogeneity and differences between the Australian and Canadian frequent attender populations to be detailed in a visual analysis that was not immediately apparent from the tabulated data. For example, while younger frequent attenders have higher rates of visiting after hours, the reasons for this are unknown from Table 1. The network analysis revealed that acute alcohol intoxication is a relatively strong driver behind after-hours visits. Only at the Toronto hospital was a subpopulation of younger female frequent attenders with abdominal pain identified, who were referred by another service. While the Toronto hospital offers services in obstetrics and gynaecology, the Melbourne hospital does not, which may have contributed to this finding. At both hospitals, joint pain and cellulitis in the limb were the only diagnoses that were associated with low triage acuity, which may not require urgent care. Additionally, red edges between the high triage acuity and LWBS nodes were observed for the Melbourne hospital. This infers that frequent attenders at the Melbourne hospital who LWBS were triaged at low acuity, suggesting less urgent needs for this subpopulation.
Compared to the Melbourne hospital, there was a higher proportion of older frequent attenders at the Toronto hospital. This discrepancy may be explained by the differences between the Australian and Canadian healthcare systems. In Australia, patients with acute exacerbations of their chronic illnesses can quickly see their private specialist as an outpatient or be admitted to a private hospital. However, Canadian private health insurance does not cover physician services and there are no private hospitals. 20 Consequently, Canadian patients need to visit the ED for urgent care. They also have greater difficulty accessing specialised tests, such as cross-sectional imaging, compared to Australian patients. 21 Such barriers to healthcare access may result in less efficient urgent care, resulting in more ED visits.

Previous research
Our study is different from previous research. While previous studies have identified mental illness, substance use and chronic disease among frequent attenders, 22-24 they did not specify the burden of disease across age groups. At the largest academic hospital in the Netherlands, the most frequently occurring diagnosis among frequent attenders was purpura or coagulation defects, 24 which was not observed for any age group at the hospitals in Melbourne or Toronto. There is a limitation to this comparison, however, as ICD-10 codes were not used to record diagnoses in the Dutch study. Therefore, comparing frequent attender populations between different studies should be done with caution. Li et al. compared ED utilisation between the USA and Ontario; however, they did not characterise frequent attender populations. 25 Finally, previous research has analysed frequent attenders across multiple regions. 26 These large registry studies, however, did not explore the important nuances in frequent attender characteristics in different settings, and their findings may have limited generalisability for individual hospitals.

Strengths
Datasets used in the present study were robust with minimal missing data. We were able to compare frequent attenders in Melbourne and Toronto using the same methods and parameters. The hospitals in Melbourne and Toronto have similar characteristics, and both cities have populations with similar demographics, 27,28 which enabled an effective comparison. The network analysis is a novel technique that allowed the effective visualisation of important characteristics and subgroups of frequent attenders.

Limitations
The present study used routinely collected data, which may not be granular enough to make meaningful conclusions about frequent attender needs. Our findings may not be generalisable to frequent attenders to community EDs, those in non-metropolitan areas, and those outside of Australia and Canada. However, our findings can guide local quality improvement projects to reduce frequent ED visits. Other frequent attender studies analysed variables relating to socioeconomic status and ethnicity; 22 however, this data was unavailable. Furthermore, data on paediatric patients were not analysed. The frequent attender definition used in the present study likely impacted the diagnoses identified, and changing the threshold for frequent attendance may yield different results. Additionally, as the Melbourne and Toronto hospitals offer different services, this may have influenced our findings. For example, the Toronto hospital provides care in obstetrics and gynaecology, which likely accounted for the younger females referred there with abdominal pain.

Clinical implications
Our findings could guide policymakers to develop interventions targeting frequent attender subgroups to reduce ED visits. As frequent attender populations in different countries are distinct, interventions need to be tailored to such populations with respect to their jurisdiction. Examples include optimising outpatient obstetric/ gynaecological care for young women presenting to the Toronto hospital with a quality improvement framework, or increasing frequent attenders' accessibility to specialists and investigations in the Canadian setting through tailored resource allocation. However, some interventions may be more generalisable across settings, such as improved joint pain management in the community, case management for older frequent attenders with chronic diseases and integrated care for those with mental illness and substance use. A systematic review reported that interventions in outpatient and primary care settings decreased ED utilisation among geriatric patients. 29 Additionally, assertive management services can improve outcomes in frequent attenders with substance use disorders and prevent hospital presentations. 30

Research implications
Using network analyses will greatly assist policymakers in knowledge translation through the effective visualisation of population characteristics. Health services could use this technique to more comprehensively investigate the factors associated with being a frequent attender (which tabulated data cannot easily convey) and to further facilitate jurisdictional comparisons. Similarities and differences between health services in different countries can provide insight into the impacts of the country's healthcare system on this vulnerable population. Furthermore, future research could investigate data not included in the present study, such as socioeconomic status, access to primary care and paediatric frequent attenders, and could stratify frequent attenders into groups based on variables other than age (e.g. time of presentation or triage categories). There is also the potential for future research to use the network approach to analyse smaller subgroups (e.g. older frequent attenders) to further explore the relationship between comorbidities and other variables of interest.

Conclusion
The present study described the characteristics of Australian and Canadian frequent attenders. The network analyses highlighted the differences between frequent attenders at the Melbourne and Toronto hospitals and revealed important subgroups at both sites. The age-specific needs of frequent attenders and their differences across jurisdictions should be considered in the design of tailored interventions.
publication of this article. Peerreview was handled independently by members of the Editorial Board to minimise bias.

Data availability statement
Additional data can be found in the supplemental material. The conditions of the ethics approvals stipulate that the raw data cannot be shared beyond the researchers of this study. Applications to the institutional ethics committees can be made if researchers wish to conduct further analysis.

Supporting information
Additional supporting information may be found in the online version of this article at the publisher's web site: Appendix S1. Network analysis: further methodology and rationale. Figure S1. Bootstrapping plots of edge-weight accuracy for younger frequent attenders at the Melbourne hospital and the Toronto hospital, after combining diagnoses relating to mental illness or substance use into a single diagnosis. Figure S2. Network analysis of younger frequent attender visits at the Melbourne hospital and the Toronto hospital. Among the top 10 diagnoses, those relating to mental illness or substance use were combined into a single diagnosis (Y1). Blue and red edges represent positive and negative partial correlations respectively. Thicker edges indicate stronger associations between variables. Threshold value = 0.037. The number of non-zero edges was 26 for the Melbourne hospital and 33 for the Toronto hospital. Figure S3. Bootstrapping plots of edge-weight accuracy for younger frequent attenders at the Melbourne hospital and the Toronto hospital. Figure S4. Bootstrapping plots of edge-weight accuracy for middleaged frequent attenders at the Melbourne hospital and the Toronto hospital. Figure S5. Bootstrapping plots of edge-weight accuracy for older frequent attenders at the Melbourne hospital and the Toronto hospital. Table S1. Top 15 diagnoses among frequent attenders at the Melbourne and Toronto hospitals. Table S2. Top 10 diagnoses among frequent attenders at the Melbourne and Toronto hospitals. Table S3. Threshold values and number of non-zero edges in the younger, middle-aged and older network analyses. Table S4. Top diagnoses among younger frequent attenders at the Melbourne and Toronto hospitals, after combining diagnoses relating to mental illness or substance use into a single diagnosis.