Identifying subgroups and risk among frequent emergency department users in British Columbia.

Objective: Frequent emergency department (ED) users are heterogeneous. We aimed to identify subgroups and assess their mortality.
Methods: We identified patients ≥18 years with ≥1 ED visit in British Columbia from April 1, 2012 to March 31, 2015, and linked to hospitalization, physician billing, prescription, and mortality data. Frequent users were the top 10% of patients by ED visits. We employed cluster analysis to identify frequent user subgroups. We assessed 365-day mortality using Kaplan-Meier curves and conducted Cox regressions to assess mortality risk factors within subgroups.
Results: We identified 4 subgroups. Subgroup 1 ("Elderly") had median age 77 years (interquartile range [IQR]: 66-85), 5 visits/year (IQR: 4-6), median 8 prescription medications (IQR: 5-11), and 24.7% mortality. Subgroup 2 ("Mental Health and Alcohol Use") had median age 48 years (IQR: 34-61), 13 visits/year (IQR: 10-16), and 12.3% mortality. They made a median 31 general practitioner visits (IQR: 19-51); however, only 23.7% received a majority of services from 1 primary care physician. Subgroup 3 ("Young Mental Health") had median age 39 years (IQR: 28-51), 5 visits/year (IQR: 4-6), and 2.2% mortality. Subgroup 4 ("Short-term") had median age 50 years (IQR: 34-65), 4 visits/year (IQR: 4-5) regularly spaced over a short term, and 1.4% mortality. Male sex (all subgroups), long-term care ("Mental Health and Alcohol Use;" "Young Mental Health"), and rural residence ("Elderly" in long-term care; "Young Mental Health") were associated with increased mortality. Conclusions: Our results identify frequent user subgroups with varying mortality. Future research should explore subgroups' unmet needs and tailor interventions toward them.

INTRODUCTION

Background

High health care users account for disproportionate costs: experience in the United States and internationally indicates that the top 10% of patients account for 68% of health care spending. , “Super‐users” of health care, and particularly of emergency departments, are therefore a cost‐containment priority. , , Frequent ED users comprise 4.5%‐8% of ED patients, yet account for 21%–30% of visits. , They incur disproportionately high costs because of high ED and other health service use. , They also make higher acuity visits and are admitted and die more often than non‐frequent users. , , , ,

Despite the recognized impact of frequent ED use, there is a knowledge gap in how best to address this issue. Interventions described to date (eg, case management, care plans) are rarely tailored to patient‐specific needs. , , , Greater primary care linkages have been proposed; however, over 93% of North American frequent users are already attached. , , , , , Interventions liaising frequent users with primary care physicians at discharge have had mixed effects on ED use. ,

Existing literature indicates that frequent users are heterogeneous. They have a bimodal age distribution with peaks at 25–44 and >65 years. They make varied ED presentations including for pain, mental health, , , chronic medical illness, , , , and substance use. , , , , Costs may increase with combinations of characteristics: for instance, homelessness, multiple physical conditions, mental illness, and substance use. Characteristics and risk also vary by degree of use. Previous analyses suggest that frequent users with ≤17 visits/year may be older, have more chronic medical illness, and be admitted and die more often than extremely frequent users (≥18 visits/year), who may be younger with more substance use and mental illness‐related visits. , , Clinical heterogeneity suggests different care needs; however, these have been incompletely explored to date.

Cluster analysis is a method that has been broadly applied to identify distinct meaningful groups, including patients at risk for disease, health care practitioners with common characteristics, and patterns at a cellular level. , , , Cluster analysis has not been applied to the question of identifying frequent user subgroups with distinct care needs.

Importance

Frequent ED users are high‐cost, high‐risk, and heterogeneous patients. Plans to manage these patients rarely address root causes of frequent use. Interventions that improve health and reduce health care use could yield considerable cost savings and importantly improve patient care and outcomes. There is a need to develop a nuanced and multidimensional understanding of heterogeneous frequent user subgroups as a first step in planning targeted and effective interventions.

Goals of this investigation

The study objective was, first, to identify and characterize frequent ED user subgroups based on demographic, clinical, and health care usage patterns within a comprehensive linked administrative database in British Columbia (BC), Canada. Our second objective was to assess risk factors for 365‐day mortality among these frequent user subgroups.

METHODS

Study design, setting and participants

This was a retrospective administrative database study capturing patients who visited an ED in BC between April 1, 2012 and March 31, 2015. We split the data into 4 fiscal year groupings. We created a cohort by identifying all patients aged ≥18 years who made ≥1 ED visit based on ED visit records within the National Ambulatory Care Reporting System (NACRS) database.

Data sources

Our study database linked patient‐level data for each patient in our cohort (NACRS) to hospitalization data within the Discharge Abstract Database (DAD), physician billing data in the Medical Services Plan (MSP), provincial prescription records (PharmaNet), and mortality data (Vital Statistics). Population Data BC (PopData) housed and linked databases using personal health number, age, sex, and postal code. PopData, which originated as the provincial Ministry of Health‐funded BC Linked Health Data Set, is a multiuniversity resource supporting linkage and access to individual‐level, de‐identified data for research. PopData undertakes validation and quality assurance/control, and employs rigorous, standardized linkage procedures, using a combination of deterministic and probabilistic approaches. , Integrated data checks within PopData's linkage procedures minimize false positives and negatives, and the use of a linkage coordinating file eliminates the risk of propagation error. All patients received an anonymized study identification number that was consistent across the data. The University of British Columbia Clinical Research Ethics Board approved this study.

Study definitions and variables

Frequent users

We defined frequent users as adult patients ≥18 years who were within the top 10% of ED use when all patients who visited an ED were ordered by number of visits made within each fiscal year, consistent with a definition established by the Canadian Institute for Health Information (Supplementary Figure 1).

To determine visit counts, we first attempted to identify and remove scheduled revisits. Clinicians on our team determined that the majority of scheduled ED revisits in BC are made for intravenous antibiotics for cellulitis. We therefore developed an algorithm to identify ED visit strings occurring within 48 hours of one another with an initial diagnosis of cellulitis. To verify the accuracy of our approach, we examined diagnoses associated with 48‐hour repeat ED visits: cellulitis was the most common specified diagnosis (12.1%), followed in frequency by abdominal pain (4.0%), follow‐up examination (3.0%), and other medical care (2.3%).

We selected the top 10% of patients based on their ED visit count for each fiscal year, after presumed scheduled revisits had been removed. We present a flow chart outlining study inclusion in Supplementary Figure S2. In a sensitivity analysis, we also identified and characterized the top 5% and 1% of patients within our frequent user cohort.

Neighborhood income quintile and rural/urban residence

PopData determined neighborhood income quintile using a postal code‐based algorithm and rural residence based on “0″ in the second position of the postal code. Both are standardized methods employed by Statistics Canada.

The Bottom Line

Over a 3‐year period, frequent emergency department users in British Columbia represented 10% of ED visits, and could be classified into 4 frequent user subgroups with unique demographic, clinical, and visit pattern profiles. Frequent users had higher mortality (11.3%) compared to non‐frequent users (3.8%). Tailored interventions should be aimed at addressing the unique needs of these subgroups.

Index visit

We chose April 1, 2013 to March 31, 2014 as our study year, as this was the most recent year of available data that would provide us with a complete baseline period and follow‐up period for all included patients. We defined frequent users’ final ED visit between April 1, 2013 and March 31, 2014 as each individual's index visit. We chose patients’ final visits as their index visits to mirror the clinical situation in which a treating emergency physician might use a frequent user patient's visit history in the preceding year to assess their present risk, including reasons for past presentations and ED visit patterns. We examined baseline demographic and clinical characteristics for each patient in a preindex period that was unique for each individual patient (the 365‐day period before the index visit). We determined all independent variables for cluster and survival analyses during this preindex period. We examined mortality within the 365‐day period following each individual's index visit. We structured our analysis such that it would mirror information that might be available to an emergency physician when caring for a patient in the ED and reflect outcomes of interest to emergency physicians (eg, within a time frame that could be influenced by interventions initiated during that visit). Although the exact 365‐day preindex period used to determine baseline variables and the 365‐day postindex period to determine mortality were unique for each patient, data available for our study ranged from April 1, 2012 to March 31, 2015.

Majority source of care

We used the general practitioner specialty code to identify primary care physicians and visits using MSP data. We then created a Majority Source of Care variable that described continuity of primary care. Majority Source of Care is a binary variable that identified whether ≥50% of patients’ services were provided by a single general practitioner, among patients who received ≥3 family doctor services in 1 year. This variable is a recognized standard for measuring primary care continuity. , ,

Long‐term care residence

We used a previously developed algorithm based on MSP billing codes to identify patients’ residence in a long‐term care (LTC) facility. We identified LTC residence if the patient received the following fee item codes attributable to a general practitioner: “Visit nursing home 1 or multiple patients,” “Nursing home visit,” “LTC institution visit,” and “LTC facility visit.”

Regularity index

The regularity index is a previously developed measure quantifying the spacing between visits. This index describes whether patients made visits that were regularly spaced relative to one another or visited sporadically. We calculated the regularity index using the following equation: 1/ 1+variance of visits, with visit variance calculated using the time in days between each visit. We measured this over a 365‐day time horizon preceding each patient's index visit. This generated a score from 0 to 1, where indices closer to 1 reflected more regular spacing. For example, 2 people who made 12 ED visits will have different “regularity” if one makes those visits randomly (irregular) and the other makes 1 visit every month throughout the year (regular).

Measure of complexity: number of prescription medications

We summed the number of distinct generic medication names listed in provincial PharmaNet records. We used number of prescription medications as a measure of patient complexity.

Measure of complexity: number of diagnostic categories

We assessed the number of diagnostic categories (International Classification of Diseases, 10th edition [ICD‐10] chapters) attributable to ED diagnoses for which frequent users’ presented as a measure of complexity. This method is a previously validated measure of the complexity of patients’ health needs. ,

Primary data analysis

Cluster analysis and clustering variables

We employed cluster analysis to explore whether subgroups existed within our frequent user cohort. Cluster analysis is a methodology to identify de novo patterns and to organize data into subgroups that maximize within‐subgroup similarities and between‐subgroup differences. We employed an iterative process to determine the number and nature of included variables. We integrated information from our team's clinical experience to identify meaningful and distinct subgroups, as is commonly done in cluster analysis. Furthermore, we applied information from a comprehensive characterization of our cohort using the same linked provincial data set to inform variable selection (in submission). For instance, our parallel analysis identified that frequent users had a high prevalence of mental illness, had greater numbers of prescription medications, and made more primary care visits than non‐frequent users. We included variables in our clustering algorithm pertaining to patients’ ED visit patterns and clinical characteristics to reflect information available to emergency physicians at the point of‐ care to guide clinical decisionmaking. We excluded patients with missing information from this analysis.

We included the following 10 variables, assessed during the 365‐day preindex period, in our clustering algorithm:

Total number of ED visits.

Number of months in the year that the patient visited an ED, a measure of visit spread.

Regularity index.

Number of ICD‐10 ED discharge diagnosis pertaining to mental and behavioral disorders (Chapter V). Substance use disorders were included in this category.

Number of ICD‐10 ED discharge diagnosis pertaining to circulatory system diseases (Chapter IX).

Number of ICD‐10 ED discharge diagnosis pertaining respiratory system diseases (Chapter X).

Number of ICD‐10 ED discharge diagnosis pertaining digestive system diseases (Chapter XI).

Number of distinct ICD‐10 ED discharge diagnostic chapters assigned to each patient.

Number of different medications prescribed.

Age at time of index visit.

We scaled all variables such that they all had a mean of zero and a standard deviation of 1 to prevent any single variable from overinfluencing the clustering algorithm. We employed the Kmeans method to identify variables that best differentiated our subgroups, as this approach was most suited to the size of our data set and number of variables we included in our algorithm. , We evaluated the optimal subgroup number using the elbow method and pseudo‐F test (Supplementary Figures 3–4 and Supplementary Table S1). , , Using these methods, we determined that 4 subgroups were the optimal number of groupings to describe our data.

Outcome measures and survival analysis

Our primary outcome was 365‐day mortality among frequent user subgroups. We present Kaplan‐Meier curves as graphical representations of 365‐day mortality stratified by subgroups. To determine patient characteristics associated with a higher mortality hazard, we calculated adjusted hazard ratios for mortality using multivariable Cox proportional hazard regressions stratified by subgroups. We used a backwards stepwise method to select which covariates to include. We considered a combination of clinical importance, optimizing the Bayesian Information Criterion number and minimizing collinearity to create our final models (Supplementary Tables S2–3). We calculated Schoenfeld residuals to test the proportional hazards assumption for variables included in our Cox models (Supplementary Table S4). Subgroup 1 demonstrated non‐proportional hazards; we therefore stratified our Cox model in this subgroup by LTC residence. We included the following covariates in our final models: sex, LTC residence, number of general practitioner visits, urban/rural residence, and neighborhood income quintile. We also included the number of individual general practitioners visited in our Cox model for subgroup 4, based on superior model performance using our variable selection algorithm. We examined outliers using DfBeta residuals. We excluded patients with missing information from the survival analysis.

We performed all analyses in R (R Development Core Team, 2011) and used the R‐packages “cluster” and “survival.” , We used α < 0.05 as the threshold for statistical significance. We provide further detailed methods in the Supplementary Materials.

RESULTS

Overall, 1,196,353 patients made at least 1 ED visit, and 205,136 were classified as frequent users over our 4 years of data. During our study year April 1, 2013 to March 31, 2014, we identified 58,491 frequent users with a median age of 53.03 years (interquartile range [IQR]: 34.93, 72.19), 25,784 (46.6%) of whom were male.

Frequent user subgroups

Our cluster analysis identified 4 frequent user subgroups. We present their characteristics pertaining to the clustering variables, demographics, and health care use in Table 1.

TABLE 1

Demographic and visit characteristics for each subgroup

	Elderly	Mental health and alcohol use	Young mental health	Short‐term
Subgroup characteristics
Age, median (IQR)	77 (66–85)	48 (34–61)	39 (28–51)	50 (34–65)
Number of visits to the ED, median (IQR)	5 (4–6)	13 (10–16)	5 (4–6)	4 (4–5)
Number of discharge diagnosis mental illness chapters, median (IQR)	0 (0–0)	0 (0–3)	0 (0–0)	0 (0–0)
Number of discharge diagnosis circulatory chapters, median (IQR)	0 (0–1)	0 (0–0)	0 (0–0)	0 (0–0)
Number of discharge diagnosis respiratory chapters, median (IQR)	0 (0–0)	0 (0–1)	0 (0–0)	0 (0–0)
Number of discharge diagnosis digestive chapters, median (IQR)	0 (0–0)	0 (0–1)	0 (0–0)	0 (0–0)
Number of prescription medications, median (IQR)	8 (5–11)	8 (5–11)	4 (2–6)	3 (2–5)
Number of different discharge diagnosis chapters, median (IQR)	3 (2–4)	4 (3–6)	2 (1–3)	2 (1–2)
Number of months in the year visited ED, median (IQR)	4 (3–4)	7 (6–8)	3 (3–4)	1 (1–1)
Regularity index, median (IQR)	3.0 × 10−4 (1.0 × 10−4‐ 9 × 10−4	1.3 × 10−3 (7 × 10−4‐2.4 × 10−3)	3 × 10−4 (1 × 10−4‐8 × 10−3)	1 (0.8–1)
Patient demographic characteristics
Total patient number, n (% of all top 10% users)	21761 (39.3)	4278 (7.7)	28164 (50.8)	1166 (2.1)
Sex
Female, n (%)	11525 (53.0)	2093 (48.9)	15388 (54.6)	569 (48.8)
Male, n (%)	10228 (47.0)	2183 (51.0)	12776 (45.4)	597 (51.2)
Unknown, n (%)	8 (0)	<5 (0)	<5 (0)	<5 (0)
Neighborhood income quintile
1st quintile, n (%)	6112 (28.1)	1754 (41.1)	8734 (31.1)	242 (20.8)
2nd quintile, n (%)	4626 (21.3)	850 (19.9)	5934 (21.1)	233 (20.0)
3rd quintile, n (%)	3935 (18.1)	699 (16.3)	4942 (17.5)	233 (20.0)
4th quintile, n (%)	3462 (15.9)	454 (10.6)	4372 (15.5)	208 (17.8)
5th quintile, n (%)	3361 (15.4)	397 (9.3)	3685 (13.1)	239 (20.4)
Unknown, n (%)	255 (1.2)	118 (2.8)	404 (1.4)	10 (0.9)
NA, n (%)	10 (0)	6 (0)	93 (0.3)	<5 (0.1)
Urban/rural
Urban, n (%)	20820 (95.7)	4116 (96.2)	26489 (94.1)	1090 (93.5)
Rural, n (%)	938 (4.3)	162 (3.8)	1661 (5.9)	76 (6.5)
Unknown, n (%)	<5 (0)	0 (0)	11 (0)	0 (0)
ED visit characteristics (all visits)
Arrive by ambulance
Ground ambulance, n (%)	43382 (36.6)	22866 (35.6)	22660 (15.8)	178 (3.1)
No ambulance, n (%)	75273 (63.4)	41422 (64.4)	120864 (84.2)	5554 (96.9)
Triage level (CTAS)
1 (Resuscitation), n (%)	889 (0.7)	332 (0.5)	532 (0.4)	2 (0.0)
2 (Emergent), n (%)	26681 (22.5)	10736 (16.7)	19676 (13.7)	248 (4.3)
3 (Urgent), n (%)	62803 (52.9)	34293 (53.3)	69415 (48.3)	1293 (22.6)
4 (Less‐urgent), n (%)	24433 (20.6)	16061 (25.0)	45971 (32.0)	3128 (54.6)
5 (Non‐urgent), n (%)	3289 (2.8)	2609 (4.1)	7205 (5.0)	844 (14.7)
Unknown, n (%)	593 (0.5)	281 (0.4)	773 (0.6)	217 (3.8)
Top 5 ICD‐10 ED diagnosis, (non‐missing), n (%)
	Chest pain, unspecified (R074), 4357 (3.7)	Other and unspecified abdominal pain (R104), 3804 (5.9)	Other and unspecified abdominal pain (R104), 7566 (5.3)	Other specified medical care (Z5188), 363 (6.3)
	Other and unspecified abdominal pain (R104), 3685 (3.1)	Mental and behavioral disorders due to use of alcohol, acute intoxication (F100), 2485 (3.9)	Chest pain, unspecified (R074), 3546 (2.5)	Urinary tract infection (N390), 311 (5.4)
	Urinary tract infection (N390), 2896 (2.4)	Chest pain, unspecified (R074), 2040 (3.2)	Cellulitis (L039), 2758 (1.9)	Follow‐up examination after unspecified treatment for other conditions (Z099), 273 (4.8)
	Cellulitis (L039), 2485 (2.1)	Anxiety disorder, unspecified (F419), 1326 (2.1)	Lower back pain (M545), 2568 (1.8)	Acute pyelonephritis (N10), 259 (4.5)
	Pneumonia, unspecified (J189), 2285 (1.9)	Lower back pain (M545), 1220 (1.9)	Urinary tract infection (N390), 2213 (1.5)	Cutaneous abscess, furuncle and carbuncle, unspecified (L029), 209 (3.6)
Top 5 ICD‐10 ED diagnostic categories, (non‐missing), n (%)
	Symptoms, signs and abnormal clinical and laboratory findings (XVIII), 22660 (19.1)	Symptoms, signs and abnormal clinical and laboratory findings (XVIII), 12387 (19.3)	Symptoms, signs and abnormal clinical and laboratory findings (XVIII), 22787 (15.9)	Diseases of the musculoskeletal system and connective tissue (XIV), 769 (13.4)
	Endocrine, nutritional and metabolic diseases (IX), 9874 (8.3)	Mental and behavioral disorders (V), 8545 (13.3)	Injury, poisoning, and certain other consequences of external causes (XIX), 16291 (11.3)	Factors influencing health status and contract with health services (XXI), 730 (12.7)
	Injury, poisoning, and certain other consequences of external causes (XIX), 9731 (8.2)	Injury, poisoning, and certain other consequences of external causes (XIX), 5827 (9.1)	Mental and behavioral disorders (V), 8001 (5.6)	Injury, poisoning, and certain other consequences of external causes (XIX), 571 (10.0)
	Diseases of the circulatory system (X), 6490 (5.5)	Diseases of the respiratory system (XI), 3590 (5.6)	Diseases of the musculoskeletal system and connective tissue (XIV), 7060 (4.9)	Symptoms, signs, and abnormal clinical and laboratory findings (XVIII), 411 (7.2)
	Diseases of the skin and subcutaneous tissue (XIII), 5265 (4.4)	Factors influencing health status and contract with health services (XXI), 3455 (5.4)	Diseases of the skin and subcutaneous tissue (XIII), 6205 (4.3)	Diseases of the digestive system (XII), 377 (6.6)
Discharge disposition
Discharged home or place of residence, n (%)	79153 (66.7)	50847 (79.1)	124562 (78)	5556 (25)
Admitted or transferred, n (%)	39094 (32.9)	12585 (19.6)	18087 (12.6)	169 (2.9)
Left before completion of treatment, n (%)	360 (0.4	870 (1.4)	908 (0.6)	7 (0.1)
Died, n (%)	81 (0.1)	10 (0)	15 (0)	<5 (0)
Other health care utilization
Number of admissions per person in last 365 days, median (IQR)	2 (1–3)	3 (2–5)	1 (1–2)	1 (1–1)
Median time admitted (days, [IQR])	5 (2–12)	3 (1–8)	3 (1–7)	2 (1–4)
Long‐term care residence, n (%)	1340 (6.2)	59 (1.4)	72 (0.3)	<5
Number of general practitioner visits, median (IQR)	25 (16–39)	31 (19–51)	13 (8–21)	9 (6–14)
Number of individual general practitioner visited, median (IQR)	8 (5–11)	14 (9–19)	7 (5–9)	5 (3–7)
Majority source of care, n (%)	12288 (56.5)	1012 (23.7)	10040 (35.6)	410 (35.2)

CTAS, Canadian Triage and Acuity Scale; ED, emergency department; ICD, International Classification of Diseases; IQR, interquartile range.

Subgroup 1 (“Elderly”) had a median age of 77 years (IQR: 66–85), was more frequently female, and made a median of 5 visits/year (IQR: 4–6). They had high indicators of complexity: median of 8 prescriptions (IQR: 5–11) and made visits related to a median of 3 diagnostic categories (IQR: 2–4). Endocrine and metabolic disorders accounted for 8.3% of their visits, and circulatory complaints for 5.5% of their visits.

Subgroup 2 (“Mental Health and Alcohol Use”) had a median age of 48 years (IQR: 34–61), was more frequently male, and made a median of 13 visits/year (IQR: 10–16). Mental health accounted for 13.3% of visits. Alcohol‐related visits accounted for 6.1% of visits, and 3.9% related specifically to alcohol intoxication. , Patients in Subgroup 2 were relatively medically complex, with a median 8 prescription medications (IQR: 5–11), and ED visits related to a median of 4 different diagnostic categories (IQR: 3–6). They made a median 31 general practitioner visits (IQR: 19–51); however, only 23.7% received a majority of services from 1 primary care physician.

Subgroup 3 (“Young Mental Health”) had a median age of 39 years ([IQR]: 28–51), was more frequently female, and made a median of 5 visits/year (IQR: 4–6). They had moderate complexity relative to other subgroups, as indicated by a median 4 prescription medications (IQR: 2–6) and a median of 2 different diagnostic categories (IQR: 1–3). Mental illness accounted for 5.6% of their ED visits; specifically, alcohol‐related presentations accounted for 1.7% of their visits. ,

Subgroup 4 (“Short‐term”) had a median age of 50 years (IQR: 34–65), was more frequently male, and made a median of 4 visits/year (IQR: 4–5). They had low indicators of complexity (median 3 prescriptions [IQR: 2–5], and made visits related to a median of 2 diagnostic categories [IQR: 1–2]). They made very regularly spaced visits over a short time (ED visits within a median of 1 month [IQR: 0.8–1] over the baseline year). Common diagnoses in this group included urinary tract infection (5.4%), follow‐up examination (4.8%), pyelonephritis (4.5%), and abscess (3.6%).

Our sensitivity analysis among the top 5% of frequent users identified 4 subgroups with similar characteristics (Supplementary Table S5). Among the top 1% of frequent users, 3 of 4 identified subgroups were similar (“Elderly;” “Mental Health and Alcohol Use;” “Young Mental Health”); however, we no longer identified a “Short‐term” subgroup. Instead, we identified a “Respiratory Illness” subgroup comprising older patients (median age 60 [IQR: 48–70]) who had made a median of 7 (IQR: 5–9) respiratory‐related visits/year (Supplementary Table S6).

Frequent users’ mortality

We observed high mortality overall among frequent users (11.3% [6635/58,491]), compared to non‐frequent users in our study year (3.81% [18,993/497,936]). Furthermore, our analysis revealed different 365‐day mortality among frequent user subgroups following the index visit, demonstrated by Kaplan‐Meier curves (Figure 1). We observed the highest mortality in Subgroup 1 (“Elderly”) at 24.7%, followed by 12.3% in Subgroup 2 (“Mental Health and Alcohol Use”), 2.2% in Subgroup 3 (“Young Mental Health”), and 1.4% in Subgroup 4 (“Short‐term”). Causes of death also varied among subgroups (Table 2). The top causes of mortality were lung cancer for Subgroup 1 (“Elderly”), chronic obstructive pulmonary disease for Subgroup 2 (“Mental Health and Alcohol Use”), lung cancer for Subgroup 3 (“Young Mental Health”), and atherosclerotic heart disease for Subgroup 4 (“Short‐term”). Figure 2 illustrates percentages of patients who died within each subgroup, with subgroups separated visually by median age and size. We present Kaplan‐Meier curves for all variables, stratified by subgroup in Supplementary Figures S5–9.

FIGURE 1

Kaplan‐Meier curves stratified by subgroup

TABLE 2

Characteristics of patients who died within 365 days of index visit, by frequent ED user subgroups

	Elderly	Mental health and alcohol use	Young mental health	Short‐term
	(Med age: 77)	(Med age: 48)	(Med age: 39)	(Med age: 50)
Number of deaths (% of all frequent users within each subgroup)	5385 (24.7%)	527 (12.3%)	618 (2.2%)	16 (1.4%)
Sex
Female, n (%)	2516 (46.7)	204 (38.7)	253 (40.9)	3 (18.8)
Male, n (%)	2865 (53.2)	321 (60.9)	365 (59.1)	13 (81.3)
Unknown, n (%)	<5 (0.1)	<5 (0.4)	<5 (0)	<5 (0)
Age at death, median (IQR)	80 (70–88)	65 (54–78)	53 (43–61)	71 (57–83)
Neighborhood income quintile
1st quintile, n (%)	1444 (26.8)	199 (37.8)	212 (34.3)	<5 (25.0)
2nd quintile, n (%)	1202 (22.3)	113 (21.4)	137 (22.2)	5 (31.3)
3rd quintile, n (%)	995 (18.5)	103 (19.5)	93 (15.0)	<5 (6.3)
4th quintile, n (%)	879 (16.3)	45 (8.5)	84 (13.6)	<5 (0)
5th quintile, n (%)	809 (15.0)	50 (9.5)	78 (12.6)	6 (37.5)
Unknown, n (%)	56 (1.0)	16 (3.0)	13 (2.1)	<5 (0)
NA, n (%)	<5 (0)	<5 (0.2)	<5 (0.2)	<5 (0)
Urban/rural
Urban, n (%)	5118 (95.0)	513 (97.3)	571 (92.4)	13 (81.3)
Rural, n (%)	267 (5.0)	14 (2.7)	47 (7.6)	<5 (18.9)
Top 5 causes of death, n (%)
	Lung cancer, unspecified, (C349), 463 (8.6)	Unspecified cause of mortality, (R99), 43 (8.2)	Unspecified cause of mortality, (R99), 63 (10.2)	Atherosclerotic heart disease, (I251), 2 (12.5)
	Atherosclerotic heart disease, (I251), 245 (4.5)	COPD, unspecified, (J449), 31 (5.9)	Lung cancer, unspecified, (C349), 39 (6.3)	HIV, unspecified, (B24), 1 (6.3)
	Unspecified dementia, (F03), 190 (3.5)	Lung cancer, unspecified, (C349), 19 (3.6)	Malignant neoplasm of brain, (C719), 25 (4.0)	Liver cell carcinoma, (C220), 1 (6.3)
	COPD, unspecified, (J449), 186 (3.5)	Alcoholic cirrhosis of liver, (K703), 17 (3.2)	Malignant neoplasm of breast, (C509), 23 (3.7)	Lung cancer, unspecified, (C349), 1 (6.3)
	Acute myocardial infarction, unspecified, (I219), 182 (3.4)	Malignant neoplasm of prostate, (C61), 16 (3.0)	Alcoholic cirrhosis of liver, (K703), 22 (3.6)	Malignant melanoma of lower limb, (C437), 1 (6.3)
Long‐term care residence, n (%)	644 (12.0)	26 (4.9)	9 (1.5)	<5 (0)
Number of general practitioner visits, median (IQR)	32 (20‐51)	43 (25‐71)	18 (11‐29)	20 (12‐27)
Number of individual general practitioner visited, median (IQR)	9 (6‐13)	15 (10‐21)	8 (5‐10)	6 (5‐9)
Majority source of care, n (%)	2729 (50.7)	153 (29.0)	254 (41.1)	11 (68.8)

COPD, chronic obstructive pulmonary disease; ED, emergency department; IQR, interquartile range.

FIGURE 2

Bubble plot illustrating subgroup size, median age, and 365‐day mortality

In our adjusted analysis, male sex was associated with an increased hazard ratio for mortality among all subgroups. LTC residence was associated with an increased hazard ratio for mortality in the “Mental Health and Alcohol Use” and “Young Mental Health” subgroups. Residing in the lowest and second‐highest neighborhood income quintiles were protective of mortality in the “Mental Health and Alcohol Use” subgroup, and residing in a rural location increased the hazard ratio for mortality in the subgroups of “Elderly” not residing in LTC and “Young Mental Health.” Increased numbers of general practitioner physician visits were associated with a slight increased hazard ratio for mortality among patients in the “Elderly” subgroup not residing in LTC, the “Mental Health and Alcohol Use,” and “Young Mental Health” subgroups (Table 3).

TABLE 3

Multivariable models of 365‐day mortality stratified by frequent ED user subgroups

	Beta	95% CI	P value
Subgroup 1 (elderly) residing in long‐term care
Sex
Female	Ref	Ref	Ref
Male	1.51	(1.29–1.77)	3 × 10−07***
Number of general practitioner physician visits	1.00	(0.99–1.00)	0.08
Urban/rural
Urban	Ref	Ref	Ref
Rural	0.69	(0.44–1.09)	0.11
Neighborhood income quintile
1st quintile	0.89	(0.71–1.11)	0.29
2nd quintile	1.15	(0.91–1.45)	0.24
3rd quintile	Ref	Ref	Ref
4th quintile	0.83	(0.64–1.09)	0.18
5th quintile	0.93	(0.70–1.22)	0.58
Subgroup 1 (elderly) not residing in long‐term care
Sex
Female	Ref	Ref	Ref
Male	1.38	(1.31–1.47)	<2 × 1016***
Number of general practitioner physician visits	1.01	(1.01–1.01)	<2 × 1016***
Urban/rural
Urban	Ref	Ref	Ref
Rural	1.19	(1.05–1.36)	0.01
Neighborhood income quintile
1st quintile	0.95	(0.87–1.04)	0.27
2nd quintile	1.04	(0.95–1.14)	0.43
3rd quintile	Ref	Ref	Ref
4th quintile	1.06	(0.96–1.17)	0.26
5th quintile	1.00	(0.90–1.10)	0.94
Subgroup 2 (mental health and alcohol use)
Sex
Female	Ref	Ref	Ref
Male	1.62	(1.36–1.94)	1 × 10−07***
Long‐term care
False	Ref	Ref	Ref
True	2.63	(1.71–4.05)	1.1 × 10−05***
Number of general practitioner physician visits	1.01	(1.01–1.01)	<2 × 1016***
Urban/rural
Urban	Ref	Ref	Ref
Rural	0.62	(0.35–1.07)	0.09
Neighborhood income quintile
1st quintile	0.75	(0.59–0.95)	0.02*
2nd quintile	0.89	(0.69–1.17)	0.42
3rd quintile	Ref	Ref	Ref
4th quintile	0.64	(0.45–0.90)	0.01*
5th quintile	0.87	(0.62–1.22)	0.41
Subgroup 3 (young mental health)
Sex
Female	Ref	Ref	Ref
Male	1.79	(1.52–2.11)	2.8 × 10−12***
Long‐term care
False	Ref	Ref	Ref
True	4.34	(2.24–8.42)	1.4 × 10−05***
Number of general practitioner physician visits	1.01	(1.01–1.02)	<2 × 1016***
Urban/rural
Urban	Ref	Ref	Ref
Rural	1.36	(1.01–1.83)	0.05*
Neighborhood income quintile
1st quintile	1.25	(0.98–1.59)	0.08
2nd quintile	1.19	(0.91–1.55)	0.20
3rd quintile	Ref	Ref	Ref
4th quintile	1.01	(0.75–1.36)	0.93
5th quintile	1.14	(0.84–1.54)	0.40
Subgroup 4 (short‐term)
Sex
Female	Ref	Ref	Ref
Male	5.80	(1.59–21.12)	0.008**
Long‐term care
False	Ref	Ref	Ref
True	0.18	(0.00–Inf)	1.00
Number of general practitioner physician visits	1.02	(0.99–1.05)	0.27
Number of individual general practitioner physicians visited	1.24	(1.01–1.52)	0.04**
Urban/rural
Urban	Ref	Ref	Ref
Rural	2.38	(0.66–8.52)	0.18
Neighborhood income quintile
1st quintile	3.03	(0.33–27.62)	0.29
2nd quintile	4.38	(0.50–38.12)	0.17
3rd quintile	Ref	Ref	Ref
4th quintile	0.00	(0.00–Inf)	1.00
5th quintile	5.44	(0.65–45.32)	0.10

ED, emergency department.

LIMITATIONS

Our analysis is constrained by inherent limitations in large administrative database analyses. First, our use of NACRS to identify patients who made ED visits in BC will miss records from EDs not reporting to NACRS. Despite this, NACRS remains the most comprehensive provincial ED visit repository, and by 2015/16, our most recent year of data, 29 BC EDs contributed 1.57 million records to NACRS, accounting for 74% provincial coverage. Second, our reliance on ED diagnoses to characterize frequent users will be affected by accuracy and completeness of discharge diagnostic information. We observed 31.8% and 30.1% missing data for discharge diagnoses for frequent and non‐frequent users, respectively (we do not suspect that missing data are systematically different). Discharge diagnosis was a conditional mandatory field for Level 2 reporting facilities in BC during the study period, meaning that completion of either but not both of presenting complaint of discharge diagnosis was required. , Institutions have variable quality and standardization of NACRS coding in BC; training and medical expertise of coders can vary widely, introducing possible transcription and interpretation errors. Furthermore, additional diagnoses beyond the primary visit diagnosis may not be reliably captured. To mitigate coding inaccuracies in individual discharge diagnoses, we included broader diagnostic categories (ICD‐10 diagnostic chapters) in our frequent user characterization and cluster analysis, an approach that has demonstrated improved reliability. Third, although NACRS has a field for scheduled revisits, this is unreliably coded. We therefore developed an algorithm to identify presumed scheduled revisits for cellulitis treatment, based on our clinical experience. Although our method has not been validated, we examined all visits occurring within 48 hours of one another and did not observe that diagnoses other than cellulitis commonly appeared. Fourth, we could not explore important patient factors not captured within included databases (eg, housing insecurity/homelessness, ethnicity, employment, and individual/household income). Fifth, our analysis precedes the opioid overdose‐related public health emergency in BC declared in April 2016. The opioid epidemic may have changed the characteristics and risk profiles of frequent user subgroups. Nonetheless, our current analysis sheds important light on comorbidities (eg, mental illness and alcohol) and risk factors (eg, LTC residence) that likely continue to affect experiences and outcomes of patients with high health care use affected by opioid overdose. Finally, the generalizability of our BC analysis to other settings is unknown. Nonetheless, despite anticipated interjurisdictional nuances, commonalities between our results and existing frequent user characterizations in other locations suggest that our findings capture overarching characteristics with wide relevance. , Furthermore, our novel application of cluster analysis to identify population‐level patterns offers an important proof of concept that could be replicated in other population data sets to better understand nuances of frequent user subgroups in other settings.

DISCUSSION

Our study confirms that distinct subgroups exist within the top 10% of the highest using ED patients, each with unique demographic, clinical, and visit pattern profiles. Our cluster analysis identified 4 frequent user subgroups: complex elderly, middle‐aged with comorbid mental health and alcohol‐related presentations, young with comorbid mental health, and middle‐aged with increased visits over a short term. The final subgroup likely comprises patients who make multiple visits related to a discrete time‐limited event, such as a complication, acute infection, or injury. Some of these patients likely represent those making scheduled revisits not captured by our algorithm focused on return visits for cellulitis. Our analyses also indicate that the 365‐day mortality among identified subgroups vary widely, with elderly frequent users and middle‐aged frequent users with mental health and alcohol use comorbidities having the highest risk. One quarter of patients in our “Elderly” subgroup, and 12.3% of patients in our middle‐aged, “Mental Health and Alcohol Use” subgroup died within 365 days of the index visit. In adjusted analysis, characteristics associated with increased hazard ratios for mortality included male sex (all subgroups), LTC residence (“Mental Health and Alcohol Use” and “Young Mental Health” subgroups), and rural residence (“Elderly” subgroup not residing in LTC and “Young Mental Health” subgroup). ED clinicians should consider supportive interventions and closer follow‐up plans for frequent users with these profiles.

Our results indicate that frequent user subgroups are important to consider distinctly: their unique characteristics and risk profiles likely indicate important differences in unmet needs and gaps in care that underlie frequent visits. Previous literature has identified heterogeneity with regard to age, reasons for presentation, and comorbidities. , Our study adds another level of understanding by using a clustering algorithm to identify distinct subgroups based on patterns within a comprehensive, linked provincial ED visit data set. Notably, our data‐driven approach confirms much of what is apparent clinically: mental health and alcohol use are important comorbidities among frequent users; and the elderly frequent users with complex medical histories have poor outcomes. In addition to confirming our clinical experience, our study highlights important and novel insights that challenge clinical intuition: nearly 1 in 8 patients within our middle‐aged, “Mental Health and Alcohol Use” subgroup died within 1 year. Our finding contradicts previous studies indicating that similar frequent users have lower rates of death. , The profile of frequent users who are perceived as “regular” patients at local EDs are often treated and discharged, with limited options for follow‐up, referrals, or ongoing care, and with minimal attention paid to whether follow‐up options actually meet individuals’ needs (eg, cultural sensitivity, accessibility). That these patients die often, and are relatively young, indicate that interventions could potentially have a considerable impact on life‐years saved. Our analysis corroborates previous findings that frequent ED users are also high users of primary care. , , , , , However, our study adds a unique dimension—the continuity, not just quantity, of primary care likely matters. , Only 23.7% of patients in our “Mental Health and Alcohol Use” subgroup had 1 physician serving as a majority source of care for their primary care visits. Investing in building trusting therapeutic relationships between frequent users with mental health and alcohol use comorbidities and primary care physicians should be explored as an intervention that could improve outcomes and save health care system costs.

In conclusion, our study provides profiles of frequent ED users who have a high overall mortality compared to non‐frequent users, but where risk is concentrated among patients who are complex and elderly and among middle‐aged frequent users with mental health and alcohol use comorbidities. Because of frequent users’ heterogeneity, one‐size‐fits‐all interventions are unlikely to succeed. Future studies should explore unmet needs and drivers of ED use among the subgroups identified, and should collaboratively develop, pilot and test interventions targeted to specific frequent user subgroups.

CONFLICT OF INTEREST

The authors declare no conflict of interest.

AUTHOR CONTRIBUTIONS

JM conceived the study, designed the analysis, obtained research funding, analyzed the data, interpreted results, and provided overall study oversight. FOS designed the analysis, analyzed the data, and interpreted results. MJM, MJS, KD, BRH, EG, CMH, and JH provided feedback on study design, data analysis, and results interpretation. KMM served as a methodological expert, designed the analysis, analyzed the data, and provided feedback on results interpretation. JM drafted the manuscript and all authors contributed substantially to its revision. JM takes responsibility for the paper as a whole.

DISCLAIMER

All inferences, opinions, and conclusions drawn in this article are those of the authors and do not reflect the opinions or policies of the Data Steward(s).

Supporting Information

Click here for additional data file.