Behavioral Health Factors as Predictors of Emergency Department Use in the High-Risk, High-Cost Medicare Population

It is well established that a minority of the population drives the majority of health care spending (1). In the Medicare population, approximately 85% of costs are incurred each year by 25% of beneficiaries (2). A large and distinct cluster of these high-cost patients is identified as “frequent care,” with frequent hospitalizations and emergency department (ED) use (3). These patients often have multiple chronic medical conditions and have relatively high rates of mortality, and thus they are termed “high risk.”

Only two previous studies have examined frequent ED use in the general Medicare population (4,5). Conducted by Colligan and colleagues, both found that mental illness was one of the strongest predictors of frequent (4) and persistent (5) ED use. Psychiatric disorders and substance abuse were also common among high-cost Medicare patients deemed “frequent care” (3). What is less clear is whether behavioral health factors (psychiatric illness, substance abuse, and psychosocial problems) in the high-cost, high-risk Medicare population predict intensity of ED use and whether there are distinct behavioral health subgroups that can be targeted for intervention.

To address this knowledge gap, we examined the role of behavioral health factors as predictors of any ED use and the amount of ED use in a high-risk, high-cost Medicare population at our urban academic medical center. We sought to ensure detection of the presence of behavioral health factors in this population by including use of medications that may be psychotropic as markers of behavioral health issues and by searching the entire record (including inpatient and outpatient encounters and ED encounters) for the presence of diagnoses related to behavioral health. We also constructed a hypothetical model that predicted ED use rates when behavioral health factors were excluded to estimate the contribution of such factors to ED use rates.

Methods

Study Population and Data Sources

This study involved patients in the Medicare Case Management for High-Cost Beneficiaries Demonstration Project (CMHCB-DP) at Massachusetts General Hospital (MGH). MGH is an urban tertiary academic medical center with 950 beds and an ED that handles 90,000 visits annually. Patients were eligible for inclusion in the CMHCB-DP if they had an MGH primary care physician and were covered by Medicare or by Medicare and Medicaid (dually eligible; mostly disabled patients over age 65), with hierarchical condition category (HCC) risk scores of ≥1.5 and annual costs of at least $8,000 or HCC risk scores of ≥2.0 and a minimum of $6,000 annual medical costs in 2005. Patients were excluded from the CMHCB-DP if they had end-stage renal disease or were on dialysis, were receiving hospice care at the start of the program, resided in a skilled nursing facility or nursing home, were enrolled in Medicare Advantage or had Medicare as a secondary payer, or lost eligibility for Medicare Part A or B. The CMHCB-DP ran between August 1, 2006, and December 31, 2011. This study included the subset of CMHCB-DP patients continuously enrolled in the CMHCB-DP for at least 12 months between February 1, 2007, and October 31, 2011; it analyzed data from claims reflecting activity during the 12 months following the date of each patient’s initial enrollment. A comprehensive description and evaluation of the MGH CMHCB-DP has been published elsewhere (6).

This project was a retrospective quality improvement activity conducted under the aegis of institutional clinical leadership in concordance with relevant HIPAA, HITECH, and other constraints; confidentiality and patient privacy were fully protected. A data use agreement between MGH and the Centers for Medicare and Medicaid Services (CMS) for CMHCB-DP, which specified restrictions on use of the data, was signed before the study was conducted. As required by that agreement, data used in the study were destroyed.

Data for our study were obtained from four distinct sources: CMS CMHCB database, which provided patient-level data, such as enrollee gender, date of birth, start and end dates of enrollment, monthly enrollment status and reason for disenrollment or ineligible status, date of death, and HCC risk score (a claims-based measure generated by CMS to determine overall illness burden [7]); the MGH ED information system, which provided ED visit–level data; the MGH comprehensive electronic medical record (EMR), which provided a list of diagnoses, procedures, and medications; and the hospital’s electronic billing system, Transitions Systems Inc., which provided the diagnosis codes (ICD-9-CM) associated with the patient’s MGH clinic, hospital inpatient, and outpatient encounters.

Identification of Enrollees With Selected Behavioral Health Conditions

Diagnoses from all visits (MGH clinics, hospital inpatient, and outpatient) were gathered; all codes found in DSM-IV-TR (8) or the mental disorders section of ICD-9-CM were considered as diagnoses of behavioral health conditions. Behavioral health diagnostic codes were sorted into 11 mutually exclusive categories: adjustment, affective, anxiety, axis II, eating, neuropsychiatric, other, psychosis, sleep, substance abuse, and unknown. [A list of ICD-9-CM codes in each of these categories is available in an online supplement to this article.] In addition, we created the variable “number of behavioral health diagnosis categories” (none, one, or two or more). Data on all prescribed medications, current and past, documented in the EMR were gathered, and medications were divided into two groups: psychotropic and nonpsychotropic. Medications with multiple possible uses, such as valproate (which can be used for the treatment of seizures or mood disorder) were included as psychotropics. The psychotropic medications were further divided into six mutually exclusive categories: anxiety, cognitive, mood, psychosis, sleep, and substance abuse [see online supplement]. In addition, we created the variable “number of behavioral health medication categories” (none, one, or two or more). Subcategorization of behavioral health diagnostic codes and psychotropic medications was conducted by two investigators (JBW and JBT, both board certified in psychiatry), and disagreements between subcategorization were resolved by consensus.

Outcome Measure and Predictor Variables

The primary outcome measure was the number of MGH ED visits made by each patient during the first 12 months of CMHCB-DP enrollment. We separately calculated the probability of ED use and the intensity of ED use. We defined a frequent ED user as one with four or more annual ED visits. We separately characterized the ED visits stratified by frequency, including discharge disposition, discharge diagnosis (top five most commonly occurring, as well as primary and secondary diagnoses of behavioral health conditions), and length of stay.

Control variables included sex, age at enrollment, HCC score (stratified into four categories defined by the first, median, and third quartiles of HCC scores), poststudy survival, and year of enrollment.

Statistical Analysis

Analyses were performed with SAS, version 9.4, and two-tailed p values less than .05 were considered statistically significant in all analyses. Preliminary analysis of the data included simple counts and proportions to summarize patient characteristics for those who had an ED visit (one, two, three, and four or more visits) versus no ED visits. Similarly, for patients who had ED visits, we summarized ED visit characteristics.

We used a multivariable logistic regression model for the dichotomous outcome of any ED use (absent or present). To model the number of ED visits when at least one ED visit was present, we used a multivariable Poisson regression model. Reference levels for variables included in both models were identified for sex (reference category: female), age at start of enrollment (<60 years), HCC score (2.0≤HCC<2.3), poststudy survival (no documented death as of October 2011), year of enrollment (2007), number of behavioral health diagnosis categories (none), and number of behavioral health medication categories (none). Reference levels within each variable were set in the same way for both the logistic and Poisson regression models. The coefficients of independent variables were rendered as odds ratios (ORs) and rate ratios (RRs), with 95% confidence intervals for the logistic and count portions of the models, and standard forest plots were developed.

To better understand the impact of individual behavioral health diagnosis categories on the presence and count of ED visits, we ran a second set of multivariate logistic and Poisson regression models, respectively. In these models, we replaced the variable “number of behavioral health diagnosis categories” (none, one, or two or more) with the 11 individual behavioral health diagnosis categories. In these models, the reference category for behavioral health diagnosis categories was none present.

Finally, we addressed the “counterfactual” question, “If none of the patients had any behavioral health diagnosis categories or medications, all else equal, what fraction would have at least one ED visit?” To do so, we set the behavioral health diagnosis category and medication variables all to zero and generated a predicted ED use probability for each patient by using the original coefficients and the same values for the other independent variables. Similarly, we used the original coefficients from the Poisson model to form predicted ED visit counts for each patient after setting behavioral health diagnosis category and medication variables to zero. This allowed us to answer a related counterfactual question, “If none of the patients had any behavioral health diagnosis categories or medications, how many ED visits would we expect from the cohort?”

Results

We identified 4,486 patients who enrolled in the MGH CMS CMHCB-DP between February 1, 2007, and October 31, 2011. We excluded 847 patients who had less than 12 months of enrollment and 19 patients with discontinuous enrollment, resulting in a final study cohort of 3,620 patients. Overall, 1,341 of 3,620 (37%) patients had a total of 2,587 ED visits during their first 12 months of participation.

Table 1 summarizes patient characteristics for our final study cohort. The mean±SD age was 76±11 years, and 53% were female. The mean HCC score was 3.0±1.1. Most patients had a current behavioral health diagnosis category or medication (69% (N=2,514 of 3,620) and 80% (N=2,891 of 3,620), respectively); only 12% of patients did not have either a current behavioral health diagnosis category or medication. Overall, 37% of patients made at least one MGH ED visit during their first 12 months of CMHCB-DP enrollment. Frequent users (four or more ED visits per year) were proportionately younger, had higher HCC scores, had shorter longevity, and were proportionately more likely to have two or more behavioral health diagnosis categories and two or more behavioral health medications.

Table 2 summarizes ED visit characteristics for each frequency group, as well as for all ED users. Although only 146 patients (4% of the cohort) were identified as frequent users (≥4 visits), these patients made a total of 800 ED visits, representing 31% of the total ED visits made by the cohort. Frequent users made from four to 15 ED visits per year. No difference was found between frequent and infrequent users in primary ED discharge diagnosis. Only 3% of ED visits (N=64 of 2,587) had a primary ED discharge diagnosis of a behavioral health condition. When secondary discharge diagnoses were included, 8% of ED visits had a discharge diagnosis (primary or secondary) of a behavioral health condition; in terms of patients, this represents 12% (N=155 of 1,341). However, 78% of these patients had at least one diagnosis of a behavioral health condition listed in the EMR (Table 1). Thus, for 85% of the patients who had a behavioral health diagnosis listed in the EMR, no ED discharge diagnosis of a behavioral health condition was given.

Figure 1 displays the adjusted ORs derived from the multivariate logistic model for any ED use. Characteristics of patients who had a higher likelihood of ED use, in descending order of relative significance, included death within 0–44 months after the first year of CMHCB-DP enrollment, presence of behavioral health medications, presence of two or more behavioral health diagnosis categories, and HCC scores higher than the median.

Figure 2 displays the adjusted RRs derived from the multivariate Poisson model for count of ED visits among those who made at least one ED visit during the study period. Of the four predictors found to be significant in the logistic model, three remained significant in the Poisson model; in the Poisson model, HCC scores did not significantly predict ED use rates. Patient age also predicted ED use rates; patients in their 70s had significantly lower rates of ED use, compared with the reference group (<60 years).

Table 3 summarizes the adjusted effects of significant predictor variables for any ED use (ORs) and number of visits among those who made at least one ED visit during the study period (RRs) when the variable “number of behavioral health diagnosis categories” was replaced with the 11 individual behavioral health diagnosis categories. ED use was significantly higher for patients with a diagnosis of adjustment disorder, neuropsychiatric disorder, psychosis, sleep disorder, and the unknown category (within the unknown category, the two most prevalent diagnoses were adult failure to thrive [ICD-9-CM code 783.7] and observation for unspecified suspected mental condition [ICD-9-CM code V71.9]). Rates of ED use increased significantly among patients with an axis II disorder, neuropsychiatric disorder, substance abuse, and other diagnosis (within the other category, the two most prevalent diagnoses were psychic factors associated with diseases classified elsewhere [ICD-9-CM code 316] and other psychalgia [ICD-9-CM code 307.89]).

Under the counterfactual condition (no behavioral health diagnosis categories or medications, simulating an otherwise identical population of 3,620 patients), the logistic model predicted that 752 (21%) patients would have had at least one ED visit. This contrasts to the 1,341 (37%) patients who made at least one ED visit. The visit intensity (Poisson) model counterfactual condition predicted a total of 1,017 ED visits, which differs from the 2,587 observed ED visits and which represents a 61% reduction in the number of ED visits.

Discussion

We found that behavioral health factors were important positive predictors of both the probability of any ED use and the intensity of ED use in the CMHCB-DP population. Patients with two or more behavioral health diagnosis categories or two or more behavioral health medications had a significantly greater tendency to make at least one ED visit, compared with patients with no behavioral health diagnosis categories or medications, and were more likely to be frequent ED users. Patients with a diagnosis category of psychosis, a neuropsychiatric disorder, a sleep disorder, or an adjustment disorder were significantly more likely than those with other behavioral health diagnosis categories to use the ED. Substance abuse and affective (mood) disorder were not significant predictors of the probability of any ED use in this population. However, among patients who had at least one ED visit, a diagnosis category of substance abuse predicted higher intensity of ED use.

Most other studies have found that medical severity and comorbidity are the most powerful predictors of ED use (9–13). Our findings suggest that this is also true for a high-cost, high-risk subpopulation. Reduced survival after the index year of CMHCB-DP enrollment is a basic marker of medical severity and was the most powerful predictor of likelihood and intensity of ED use among those with at least one visit. The HCC score, which is a standard metric of comorbidity, was the next most powerful predictor of ED use.

Our findings also are consistent with studies suggesting a direct positive relationship between behavioral health factors, such as depression, substance abuse, or psychosis, and worsened general medical outcomes (14–17) and higher ED use rates. Most important, we found that at least for this subpopulation, independently and globally assessed behavioral health factors were jointly and significantly associated with both the probability of any ED use and the number of ED visits. This view was supported by our counterfactual simulation, which suggested that most patients (61%) would not have had any ED visits but for their behavioral health problems. This is not to suggest that behavioral health factors are unrelated to general medical issues, but the finding highlights the importance of including behavioral health factors when examining predictors of ED use in a frequent-user population.

Several courses of action are suggested by this work. First, future studies are needed to address the impact of individual behavioral health diagnoses on ED use. Future studies may also assess the relative impact of the diagnosis and psychosocial factors, because it has been shown that psychosocial factors also affect ED use rates among frequent users (9,18). Future studies and ongoing clinical care management efforts should include data from the entire record, not just the ED notes, when seeking to identify patients for whom behavioral health factors may affect ED use. Finally, given the impact of removing behavioral health factors from the counterfactual model, clinical quality improvement and cost control efforts, as well as research, may be initiated to determine whether identification and case management of patients with behavioral health risk factors affect ED use rates.

This study had several limitations. First, we examined the impact of categories of behavioral health diagnoses rather than individual behavioral health diagnoses. This was done because it was not possible to confirm diagnoses with standard tools. Therefore, we did not attempt to calculate the relative predictive power of individual diagnoses, such as delirium and dementia. Similarly, we could not verify that patients with an insomnia diagnosis met insomnia diagnostic criteria and what type of insomnia they had. In addition, we could not determine how adjustment disorder diagnoses were made or how accurate they were; we suspect that for at least some patients, a diagnosis of adjustment disorder was a proxy for the presence of a high level of psychosocial distress experienced by the patient or family. Also, we were not able to include the severity of behavioral health conditions, only the presence or absence of these conditions. Second, it was not possible to confirm that prescription of a medication with multiple possible uses (for example, valproate) was exclusively for psychotropic versus nonpsychotropic indications. Third, this retrospective study captured only visits to the MGH ED and, as such, may be an underestimation of the likelihood and intensity of ED use. Fourth, although the simulated removal of behavioral health factors predicted lower ED use rates, this study could not determine whether actual reduction in the frequency or severity of behavioral health factors— for example, by improved recognition and treatment—might reduce ED use. Finally, this study could not explain how or why behavioral health factors predicted ED use. One can reasonably speculate that the presence of psychiatric illness (including substance abuse) or psychosocial problems may interfere with some patients’ capacity to engage with a treatment plan or collaborate with their clinicians, but this is an area for future investigation.

Conclusions

Behavioral health factors—especially the count of psychiatric diagnosis categories and psychotropic medications and the presence of a diagnosis of psychosis, a neuropsychiatric problem, an adjustment disorder, or a sleep disorder—predicted greater ED use in a high-risk, high-cost Medicare population after the analysis was controlled for comorbidities and other relevant variables.