Emergency department involvement in the diagnosis of cancer among older adults: a SEER-Medicare study

Abstract Background Internationally, 20% to 50% of cancer is diagnosed through emergency presentation, which is associated with lower survival, poor patient experience, and socioeconomic disparities, but population-based evidence about emergency diagnosis in the United States is limited. We estimated emergency department (ED) involvement in the diagnosis of cancer in a nationally representative population of older US adults, and its association with sociodemographic, clinical, and tumor characteristics. Methods We analyzed Surveillance, Epidemiology, and End Results Program–Medicare data for Medicare beneficiaries (≥66 years old) with a diagnosis of female breast, colorectal, lung, and prostate cancers (2008-2017), defining their earliest cancer-related claim as their index date, and patients who visited the ED 0 to 30 days before their index date to have “ED involvement” in their diagnosis, with stratification as 0 to 7 or 8 to 30 days. We estimated covariate-adjusted associations of patient age, sex, race and ethnicity, marital status, comorbidity score, tumor stage, year of diagnosis, rurality, and census-tract poverty with ED involvement using modified Poisson regression. Results Among 614 748 patients, 23% had ED involvement, with 18% visiting the ED in the 0 to 7 days before their index date. This rate varied greatly by tumor site, with breast cancer at 8%, colorectal cancer at 39%, lung cancer at 40%, and prostate cancer at 7%. In adjusted models, older age, female sex, non-Hispanic Black and Native Hawaiian or Other Pacific Islander race, being unmarried, recent year of diagnosis, later-stage disease, comorbidities, and poverty were associated with ED involvement. Conclusions The ED may be involved in the initial identification of cancer for 1 in 5 patients. Earlier, system-level identification of cancer in non-ED settings should be prioritized, especially among underserved populations.


Index date
To characterize the health-care episode in which the cancer was first detected, we used a published and validated algorithm (18) that identifies the earliest International Statistical Classification of Diseases, Tenth Revision or International Classification of Diseases, Ninth Revision claim code associated with the patient's tumor site within ±1 month of the SEER diagnosis month.This index date approximates the start of the cancer care episode (ie, when care delivery for suspected cancer began).Using this method, we successfully assigned index dates for 86.9% of study participants.For the remaining 13.1% of patients, the index date was imputed as the 15th of the SEER diagnosis month (Table 1).

Outcome
Following precedent (3), our primary (binary) outcome was whether each patient visited the ED in the 30 days before their index date, which we considered "ED involvement" in their initial cancer diagnosis.First, a 1-month lookback period starting from the index date (day 0) to 1 month (day -30) before was established.Next, ED utilization during the lookback period was defined as either any inpatient hospitalization with an associated ED claim or any outpatient claims with an ED revenue code (0450-0459).To explore the timing of ED utilization, this interval was subclassified as 0 to 7 (including ED visits on the index date) or 8 to 30 days before the index date.For patients whose index date occurred during an emergency (nonelective) in-patient hospitalization, this interval was set to 0.

Exposure variables
We classified patient demographic, socioeconomic, and clinical characteristics using SEER data: age at diagnosis, sex, race and ethnicity (Non-Hispanic White, Non-Hispanic Black, Hispanic, Asian American [including individuals of Chinese, Japanese, Filipino, Korean, Vietnamese, Asian Indian, or Other Asian ethnicity], American Indian or Alaska Native, Native Hawaiian or Pacific Islander, mixed race, other or unknown race), partnership status at diagnosis (married or partnered, not married/partnered, unknown), year of diagnosis, SEER region (East, Midwest, South, or Western United States), urban-rural residence (metropolitan counties, urban nonmetropolitan counties, rural counties), and census-tract level poverty (<5%, 5%-10%, >10%-20%, or >20%) in the year of diagnosis.Summary tumor stage (localized, regional, distant) is generally captured after initial presentation but was included as a proxy for the extent of disease at presentation.A claims-based modified Charlson Comorbidity Index score was calculated based on all claims in the year before diagnosis (19,20).

Statistical analyses
All analyses were performed using SAS, version 9.4, statistical software (SAS Institute Inc, Cary, NC).Descriptive statistics were calculated overall, by imputed vs exact index date, and by ED visit timing before the index date (0-7 days or 8-30 days).We suppressed cell counts below 11 and rounded proportions to whole numbers for in-text callouts.With ED involvement as the outcome, we used modified Poisson regression with robust variance estimation to estimate prevalence ratios and corresponding 95% confidence intervals (CIs) using the SAS PROC GENMOD statement (21,22).A random intercept was included for clustering by registry, but the intraclass correlation coefficient was low (<1%); therefore, it was treated as a fixed effect in all models.All analyses were conducted overall and by cancer site, excluding patients with unknown values of covariates, except for unknown partner status, which we treated as a separate category.Variables included in all models were age at diagnosis; year of diagnosis; tumor stage; SEER region and registry; imputed vs exact index date; number of outpatient-, inpatient-, and ED-visit-days in the year before diagnosis; and tumor site for "all cancers" models.Covariate adjustment sets for each additional exposure of interest were defined using directed acyclic graphs (23) (Supplementary Methods, available online).

Sensitivity analyses
Because 13% of our population had their index date imputed as the 15th of the SEER diagnosis month due to lack of cancerrelated claims, we excluded them from all main analyses.

Results
The starting population size was 2 110 0096 SEER region residents of any age.After applying all exclusion criteria, including restricting to those individuals 66 years of age and older with continuous Medicare fee-for-service coverage, the final study population was 614 748 people (Supplementary Figure 1, available online).The average age at diagnosis was 76 years, half the population was male, 81% were non-Hispanic White, 8% were non-Hispanic Black, 5% were Hispanic, 4% were Asian American, fewer than 1% were Native Hawaiian or Other Pacific Islander or American Indian or Alaska Native.Marital status at diagnosis was distributed evenly, with 33% of each: married or partnered, unmarried, unknown.About half had a Charlson Comorbidity Index score of 0. Patients without an exact index date were slightly older and more frequently female, diagnosed with laterstage cancer, and residing in high-poverty regions (Table 1).Across all cancer sites, 23% of patients had ED involvement, with marked variation by tumor site: lung cancer at 40%, colorectal cancer (CRC) at 39%, breast cancer at 8%, and prostate cancer  1).Older patients (30% aged >85 years vs 16% aged 66-69 years) and unpartnered patients (30% vs 18% in patients with spouses or partners) had higher ED involvement.ED involvement increased from 22% in 2008 to 24% in 2014, then decreased to 23% in 2017.Non-Hispanic Black patients (28%) as well as American Indian or Alaska Native and Native Hawaiian or Other Pacific Islander patients (25%) had the highest ED involvement (vs 23% of non-Hispanic White patients).ED involvement varied by cancer stage at diagnosis (48% distant metastasis, 25% regional metastasis, 10% localized).Patients with a Charlson Comorbidity Index score of 3 or higher had higher levels of ED involvement (36% vs 19% among patients with Charlson Comorbidity Index score of 0).Patients residing in the highest-poverty regions were also more affected (28% vs 20% in the lowest-poverty regions) (Table 2).Among individuals with ED involvement, 77% visited the ED within 7 days of their index date.This proportion varied by tumor site (84% CRC, 78% lung, 60% breast, 54% prostate) and stage at diagnosis (83% distant metastasis, 78% regional metastasis, 64% local) (Figure 1).In adjusted models, ED involvement was 3% more prevalent for women overall and for all tumor sites that affect both sexes.Increasing age was associated with ED involvement overall (prevalence ratio ¼ 1.68, 95% CI ¼ 1.66 to 1.71 for patients ≥85 years of age vs patients <70 years of age), with especially notable associations for breast cancer (prevalence ratio ¼ 2.12, 95% CI ¼ 1.99 to 2.25) and prostate cancer (prevalence ratio ¼ 2.66, 95% CI ¼ 2.48 to 2.85).Relative to non-Hispanic White patients, the adjusted prevalence of ED involvement was 30% higher for non-Hispanic Black patients, 22% higher for Native Hawaiian or Other Pacific Islander patients, 17% higher for Hispanic patients, and 16% higher for American Indian or Alaska Native patients.Relative to married or partnered patients, those without a spouse or partner were 23% more likely to have ED involvement, with stronger adjusted associations for patients with breast cancer or prostate cancer.ED involvement increased with recency of year of diagnosis overall (prevalence ratio ¼ 1.16, 95% CI ¼ 1.14 to 1.17 comparing 2016-2017 with 2008-2009) and for patients with CRC, lung cancer, and prostate cancer.Distant-stage cancer was associated with ED involvement overall (prevalence ratio ¼ 2.38, 95% CI ¼ 2.34 to 2.41), with substantial variation by site (breast cancer prevalence ratio ¼ 8.95, 95% CI ¼ 8.57 to 9.35; CRC prevalence ratio ¼ 1.67, 95% CI ¼ 1.65 to 1.71; lung cancer prevalence ratio ¼ 2.04, 95% CI ¼ 2.00 to 2.08; prostate cancer prevalence ratio ¼ 3.83, 95% CI ¼ 3.65 to 4.01).Relative to the lowest comorbidity scores, patients with the highest comorbidity scores had 21% higher adjusted prevalence of ED involvement overall and for CRC, lung cancer, and breast cancer, but comorbidity status was not associated with ED involvement for patients with prostate cancer.Compared with the largest metropolitan areas, residing in smaller metropolitan and nonmetropolitan counties was associated with ED involvement for patients with prostate cancer and breast cancer.Patients residing in higher census-tract-level poverty areas were 18% more likely to have ED involvement (vs patients in the lowest poverty group), with stronger associations for patients with prostate cancer and breast cancer (Table 3).
When we restricted the analysis to patients with an exact index date, our results were materially identical to the size, direction, and significance of main analysis estimates.(Supplementary Tables 1 and 2, available online).

Discussion
Our results indicate that about 4 in 10 patients with CRC or lung cancer and 1 in 15 patients with breast cancer or prostate cancer may have visited an ED at the beginning of their diagnostic episode, with the majority visiting the ED in the 7 days before the first indication of cancer appeared in their medical records.We also found that non-Hispanic Black patients, Hispanic patients, and Native Hawaiian or Other Pacific Islander patients; individuals who are older, unpartnered, or residing in lower-income neighborhoods; and those patients with more comorbidities were more likely to have ED involvement in their cancer diagnosis.
Our estimates are consistent with recent international studies of emergency presentation for cancer.In an analysis of data from 14 jurisdictions in 6 countries, the prevalence of emergency presentation ranged from 26% to 51% for lung cancer, 23% to 37% for colon cancer, and 9% to 20% for rectal cancer (3).Furthermore, a 2019 UK study found that between 2006 and 2013, 5% of patients with breast cancer, 11% of patients with prostate cancer, 16% of patients with rectal cancer, 34% of    (2009-2011)-where 32% of all patients with cancer (N ¼ 989) were admitted through the ED (12).Our results also update population-based estimates by Pruitt et al. (7), who observed that 29% of patients with CRC in SEER-Medicare (1992-2005) had an ED claim related to obstruction, perforation, or emergency inpatient admission in the same month as their registry date of diagnosis.
The prevalence of ED involvement varied substantially by tumor site and was more common in later-stage tumors.This finding aligns with previous literature (24).Although all 4 cancer types in this study were detectable through guidelinerecommended screening tests during the study period, screening adherence was higher for breast and prostate cancers, which had the lowest levels of ED involvement, and are also more often detected at earlier stages.Even in the absence of screening, tumors of the breast and prostate have specific (and often less medically acute) symptomology that can facilitate earlier, nonemergent diagnosis, such as breast lump or hematuria, respectively.Patients with lung cancer and CRC had the highest levels of ED involvement.In the absence of screening, lung cancer and CRC often progress unnoticed until they produce a dramatic and medically serious event, such as intestinal perforation, or "red flag" symptoms, such as hemoptysis or hematochezia, the experience of which may precipitate a patient visit to the ED.
Older age and the presence of multiple comorbidities were associated with ED involvement, a finding consistent with prior literature (24).Older age is associated with functional decline and cognitive impairment, which may affect a patient's ability to recognize and report cancer symptoms, or with social isolation, limiting the social and economic support needed for attending ambulatory preventive care services (25).Morbidities may make attribution of symptoms of underlying cancer harder and may also distract attention from investigation of new symptoms (26).In addition, older age increases the risk of frailty, which can increase the risk of falls and accidents that may lead to incidental diagnosis of cancer, as well as comorbidities, such that age and comorbidity may represent compounding influences that can complicate or delay cancer diagnosis (27,28).
Unmarried or unpartnered marital status was associated with ED involvement, as well.This finding has not been widely reported beyond 2 prior Norwegian studies (29,30).Although some emergency diagnoses may be considered unavoidable because of tumor biology (24) or driven primarily by symptomology (31), the increased risk of ED involvement for unpartnered patients observed across all tumor sites studied suggest the importance of nonbiological drivers of emergency diagnosis.Being married or partnered is associated with earlier stage at diagnosis (32) and improved patient survival (33), and this relationship is largely attributed to social support.Potential explanations for the relationship with emergency presentation include previous observations that unmarried patients are less likely to have a usual source of care (34) or to seek medical care for nonspecific symptoms (35).
We observed statistically significant adjusted associations between patient race, socioeconomic conditions and ED involvement, with a higher prevalence of ED involvement among patients who were non-Hispanic Black, Native Hawaiian or Other Pacific Islander, and living in regions with more than 20% poverty.These relationships were observed in previous US studies (6,7,10) and contribute to a wide body of literature describing disparities in burden and outcomes across the care continuum for patients with cancer in the United States.This relationship, which persisted after adjusting for tumor stage and other covariates, also points to nonbiological drivers of ED involvement and may be explained by higher ED utilization for routine care (36) or poor-quality primary care (37).These populations are also known to have lower-quality cancer care and worse outcomes independent of treatment.With regard to these outcome disparities, there may be a compounding effect of ED involvement, as it has also been found to be associated with reduced receipt of curative treatment (24).The finding of higher ED involvement among Native Hawaiian or Other Pacific Islander populations is novel and underlines the importance of separating these patients from the larger population of Asian Americans, whose levels of ED involvement were similar to non-Hispanic White patients (38).
This large, nationally representative study used high-quality data.Medicare patients are insured and should have access to primary and specialty care.Our findings, however, may not be generalizable beyond the US Medicare fee-for-service population.We were not able to include a growing portion of beneficiaries who participate in Medicare Advantage plans or patients insured only under Medicaid or other state or federal insurance systems (eg, US Department of Veterans Affairs) or the uninsured.These excluded populations may differ from our study population in terms of demographics, health-care access, cancer burden (39)(40)(41)(42), and (consequently) ED involvement.Barriers to care are often higher among uninsured and younger individuals, so the estimates presented here may be attenuated compared with the general population; this hypothesis is corroborated by reports of higher levels of emergency presentation than we observed in safety-net populations (11).Our use of disease stage at diagnosis as a proxy for extent of disease at presentation is imperfect and may be confounded by an unobserved third variable (ie, healthcare seeking preferences, quality of preventive care) that may be correlated with both delay in cancer detection and the propensity to visit the ED.Our use of a claims-based index date improves upon prior research anchoring on the cancer registration date of diagnosis, which is inconsistently defined, and may not reflect the initiation of cancer-relevant diagnostic health care (43) when ED utilization would be indicative of its involvement in identifying the cancer.In the absence of chart review, however, which would have been infeasible given the population-based nature of the study and sample size, it is impossible to know with certainty that ED clinicians played a role in the identification of a patient's cancer.It is possible that patients coincidentally visited the ED within 30 days of cancer identification, which would result in overestimation of ED involvement, or underestimation occurred by classifying as non-ED involved those patients who visited the ED in the days immediately following their index date.Nonetheless, the degree of misclassification that would have been required to explain the observed patterns of large variation in prevalence of ED involvement (eg, by disease stage and by cancer site) would have had to be implausibly large.Furthermore, our findings regarding tumor and patient factors associated with ED involvement concord fully with recent international and prior US studies, and they provide a timely update to an understudied topic in US cancer population research.
Even for cancers with available screening modalities and in a population with health-care coverage, a substantial proportion of patients visit the ED in the days and weeks before they are initially diagnosed with cancer, and ED clinicians may be the first to identify or suspect and coordinate workup of up to 40% of colorectal and lung cancers.Considering the burden of CRC and lung cancer, these results suggest that the ED plays an important role in cancer diagnosis in the United States.Patients whose cancer diagnosis starts or originates in the ED may reflect a constellation of factors, many of which are potentially modifiable at the patient and system levels, including insufficient awareness of prodromal symptoms of cancer or delayed help seeking, a failure of access to screening, suboptimal access and quality of the primary health-care system, and barriers to ambulatory investigations (8).The barriers to nonemergency diagnosis of cancer may be particularly applicable to populations historically targeted for marginalization, such as racial and ethnic minority groups, or those individuals living in lower socioeconomic neighborhoods and rural areas.Emergency diagnosis may also reflect or exaggerate known racial, ethnic, and socioeconomic disparities in cancer outcomes due to delayed treatment, undertreatment, and suboptimal access to preventive care and screening.Systemlevel efforts to improve the identification of cancer in non-ED settings, before the patient reaches the ED, may improve patient outcomes and health-care system efficiency.Further research is needed to better understand how emergency diagnosis is related to survival, quality of care, and patient experience in US populations.

Figure 1 .
Figure 1.Number of patients with ED involvement in their diagnosis among Surveillance, Epidemiology, and End Results Program-Medicare patients diagnosed with breast, colorectal, lung, or prostate cancers (2008-2017), by timing of the ED visit relative to the index date (0-7 days vs 8-30 days before the index date).Figures shown for all patients combined and by tumor site (breast, colorectal, lung, prostate) and disease stage at diagnosis (localized, regional metastasis, or distant metastasis).ED ¼ emergency department.

b
Model 3 include model 2 variables and marital status.c Model 5 includes model 4 variables and Charlson Comorbidity Index.d Model 2 includes model 1 variables and county geography.

e
Model 4 includes model 3 variables and poverty category.

Table 1 .
Characteristics of SEER-Medicare patients diagnosed with breast, colorectal, lung, or prostate cancer (2008-2017) a a NA ¼ not available; SEER ¼ Surveillance, Epidemiology, and End Results.C. A. Thompson et al. | 3 at 7% (Figure

Table 2 .
Proportion and characteristics of patients with ED involvement in their diagnosis among SEER-Medicare patients diagnosed with breast, colorectal, lung, or prostate cancer (2008-2017) a (10)ents with colon cancer, and 39% of patients with lung cancer were emergency presentations (5).In the United States, Pettit et al. recently reported that 19% of patients with lung cancer, 9% of patients with breast cancer, 9% of patients with CRC, and 6% of patients with prostate cancer diagnosed statewide in Indiana (2013-2017) had evidence of ED utilization in the 6 months before diagnosis(10).Our results may differ from theirs because of differences in Indiana (which is not a SEER registry), the older age distribution of our sample, or other methodological differences.Other recent US studies were from single safety-net hospitals in New York City (2012-2014)-where 42% of 638 patients had CRC diagnosed through the ED (11)-and in Jacksonville, Florida

Table 3 .
Associations between patient sociodemographic, clinical, and tumor characteristics and ED involvement among SEER-Medicare patients diagnosed with breast, colorectal, lung, or prostate cancer (2008-2017) Prevalence ratios estimated using generalized linear models with log link and Poisson distribution.Model 1 independent variables included age group; sex; race category; sex; SEER region; disease stage at diagnosis; year of diagnosis; source of diagnosis date (registry vs claims); count of outpatient, inpatient, and emergency department visit days in the year before diagnosis (as 3 separate variables); and tumor site (for all cancers model only).ED ¼ emergency department; N/A ¼ not applicable; SEER ¼ Surveillance, Epidemiology, and End Results. a