Impact of an addiction medicine consult team intervention in a Canadian inner city hospital on acute care utilization: a pragmatic quasi-experimental study

The Enhanced Multidisciplinary Care for Inner City Patients Seeking Acute Health Care (EMCIH) study used a two-arm, pre-post, longitudinal quasi-experimental study design; details are described in our published protocol [44]. A pragmatic, non-randomized design was chosen because individual-level randomization was not feasible for a site-level service innovation, and insufficient comparable and implementation-ready sites existed within the same Canadian province to permit site-level randomization. A community-engaged research approach was used throughout the study period; the study protocol, implementation, and interpretation received iterative guidance from a Community Advisory Group of people with lived experience of both substance use and the acute healthcare system.

Intervention arm participants were recruited from among inpatients and outpatients within two weeks of their initial assessment by the intervention site’s AMCT in Edmonton, Canada. Consult requests to the AMCT were initiated by the patients’ most responsible physician and could come from any inpatient service or the emergency department (ED). During the recruitment period, 611/1448 (42%) of AMCT referral requests originated from the internal medicine service, 422 (29%) from the ED, 162 (11%) from surgery, 91 (6%) from the hospitalist team, and the remainder from a variety of other specialty services. AMCT referral criteria were active substance use related concerns and/or social instability. These more inclusive intervention eligibility criteria were applied due to a large overlap in substance use and unstable housing in the target population based on intervention site clinical experience, and a lack of reliable pre-study data on which hospitalizations would benefit from referral. Control arm participants were recruited from inpatients (medical, surgical, and psychiatric) and outpatients (emergency and urgent care) seen in two acute care facilities serving a similarly disadvantaged patient population in the nearby city of Calgary, Canada. Potential participants were identified by treating clinicians. In both study arms, research assistants confirmed study eligibility (problematic substance use, unstable housing, and/or unstable income), explained the study, obtained informed consent for the overall study, and obtained additional consent for retrieval and linkage of administrative health service data to primary survey data. Research assistants then completed a detailed contact information sheet detailing both traditional and non-traditional means of contact (e.g. name of outreach worker, usual places frequented) to facilitate participant follow-up.

The intervention site AMCT was launched in July 2014, in response to patient-voiced unmet care needs. The intervention development team considered substance-use related care, health promotion, and interventions addressing the social determinants of health to be equally important. Feedback from our Community Advisory Group, and later from our parallel process evaluation, confirmed the importance of whole person care in acute health service delivery and the need to incorporate this approach into the AMCT model. The resulting multidisciplinary consultation service was co-designed with patients, clinicians, and other stakeholders. At study inception, the core clinician complement included rotating physicians with addiction medicine expertise (one physician available 0800 h–2100 h, seven days/week), a nurse practitioner, and a social worker; an addiction counselor and peer support worker were added in the latter half of the study period. Emphasizing bridging support to promote care continuity during the hospital-community care transition, the AMCT provided management of intoxication, withdrawal, and/or acute pain; initiation or continuation of opioid agonist treatment; referral to addiction treatment and recovery programs; overdose prevention interventions; screening for sexually transmitted and blood borne infections; and, social stabilization supports including assistance with housing, income, and/or obtaining healthcare coverage, photo identification or health care card. The AMCT provided consultation on referral by the attending clinical team to both hospital inpatients and ED patients, the latter of whom were offered follow-up assessment in a transitional clinic on site. The intensity, duration, and specific components offered were tailored to patient needs and preferences. Usual care at the control sites consisted of assessment and treatment by an attending team without specific expertise in addiction medicine or related individualized care plans to manage the unique health and social needs of patients who use substances; attending teams could access a nurse-clinician on select days for addiction medicine advice.

Participants completed a baseline quantitative survey administered by a research assistant. Survey variables (see supplementary file 1) included sociodemographic information (e.g. gender, ethnicity, housing status), substance use behaviour (e.g. alcohol and drug [45, 46] consumption patterns, experience of overdose), perceived health (EQ-5D [47]), and unmet need for care (Perceived Needs for Care Questionnaire [48]). Although follow-up survey data were collected for the purposes of secondary outcome analysis, only baseline survey data were used for the analyses presented herein.

The research team collaborated with the provincial administrative health service data custodian to retrieve and link administrative data for consenting participants. Data on healthcare encounters were retrieved for a 540-day period including 180 days prior to and 360 days following the baseline study enrolment date for each participant.

Main outcome measures were obtained from administrative health data. The primary outcome measure was reduction in high ED use, where the baseline measurement period was defined as 180 days prior to baseline study enrolment date, and the follow-up measurement period was defined as 180 days falling between six and 12 months after the baseline study enrolment date. Consistent with other studies of complex populations with high acute case use, “high ED use” was defined as more than two visits to either an ED or urgent care (UC) facility during either 180-day period [49, 50]. We considered using ED frequency counts across the 540-day study period, but decided to create a change variable omitting the 180-day period immediately after intervention start because we anticipated that 1) frequency of ED visits would be non-normally distributed [26, 51] and 2) ED visits would increase as a function of the acute care-based intervention before subsequently decreasing.

Secondary acute care utilization outcome measures included reduction in inpatient admissions and reduction in hospital length of stay, calculated using the same 180-day pre- and post-periods used for the primary outcome measure.

Covariables were identified a priori based on their potential to exert clinically meaningful effects on the intervention and/or outcome. Most covariables were collected via the baseline participant survey. These included age, gender (male, female, transgender), Indigenous ethnicity (coded as Indigenous if responses of “First Nations”, “Métis”, “Inuit”, “Cree”, “Aboriginal”, “Non-status Indian”, or “Treaty” were given to the question “What ethnic group do you identify with?”), and unmet need for care (as measured by the Perceived Needs for Care Questionnaire) [48]. Shelter use was recorded as present if a participant reported sleeping in a hostel or shelter in the previous six months; it was chosen as our indicator of housing instability as the most common sleeping arrangement for those without housing. Opioid use was recorded as present if any non-prescribed use of intravenous or non-intravenous opioids was reported in the previous six months. Two non-survey covariables were also included. Clinical Risk Group (CRG [52], a diagnosis-based indicator of medical complexity that takes into account disease chronicity and severity and its expected impact on morbidity, mortality, and future service needs) was retrieved from the recruitment hospital admission or ED/UC visit and was coded low, medium or high risk. Recruitment location was documented by research staff as inpatient (ward) or outpatient (ED, UC).

We hypothesized a clinically significant intervention effect of a 20% reduction in high ED use, which a power analysis indicated would require 103 participants per study arm for detection (two-sample test of proportions, α = 0.05, power = 80%). We anticipated 50% study attrition as well as administrative data validation challenges (consistent with studies of similar cohorts [53, 54]), and therefore set a recruitment target of 300 for each study arm.

Generalized linear mixed models were used for analysis. The unit of analysis was the participant and all analyses were performed in R [55]. Summary statistics (e.g., frequencies, percentages, means, standard deviation [SD], median, interquartile range [IQR] represented as 25th percentile, 75th percentile) described covariables and outcomes by study arm. Initially, two-sample tests of proportions, chi-square tests, and t-tests compared outcomes by study arm and covariables. This was followed by multiple logistic regression analyses to compare outcomes across study arms, adjusted for age, self-identified gender, Indigenous status, housing status, opioid use, unmet need for care, CRG category, and recruitment location. Given the small number of participants that identified as transgender (n = 2), transgender was combined with the female category in all regression analyses. As CRG had 11 missing values, missing values were assumed to be missing at random and five imputed datasets using the multivariate imputation by chained equations (MICE [56]) package were created and regression models pooled. Odds ratios (ORs) and associated 95% confidence intervals (CIs) are presented and a p-value less that 0.05 was considered statistically significant.

A total of 572 patients (302 intervention arm, 270 control arm) were enrolled into the study and completed baseline survey data collection between August 2014 and June 2016 (Fig. 1). At the time of recruitment over 90% of the sample reported recent alcohol or substance use, and over 70% reported unstable housing, unstable income, or both (Table 1). Almost all participants (n = 554; 97%) consented to the retrieval and linkage of their longitudinal administrative health services data. After removing invalid personal healthcare identifiers, identifying study withdrawals, and a small number of control arm participants potentially exposed to local case management co-interventions, the final sample included 516 participants (290 intervention arm, 226 control arm) for analysis (Fig. 2). Attrition due to death, study withdrawal, or inability to contact was similar across both study arms. There were no significant deviations from the original study protocol throughout the data collection period.

Study arms differed at baseline across a number of covariables (Table 1). More intervention arm participants used opioids and were Indigenous compared to control arm participants; whereas fewer intervention arm participants had used a shelter in the past six months. Intervention arm participants were more likely to be recruited as inpatients, and also had a higher level of medical complexity, as defined by the CRG.

With respect to the primary outcome, 28.3% (82/290) of the intervention arm participants experienced a reduction in high ED use compared to 33.6% (76/226) in the control arm (Table 2, p = 0.225). There was no difference between study arms in the proportion of participants experiencing a reduction in high ED use. When these bivariate results were adjusted by covariables in a logistic regression analysis, study arm remained non-significant and only CRG predicted the primary outcome (Table 3). During preliminary analysis we explored additional modeling using an ED visit count variable rather than a change variable, however, the result of this analysis were similar and thus are not included herein.

For secondary outcomes, bivariate results indicated that pre- to post-test reductions in inpatient admissions were more common among participants in the intervention arm (212/290, 73.1%, Table 2) compared to control arm participants (132/226, 58.4%, p = 0.001). However, in subsequent multivariable analyses adjusted by covariables (Table 3), study arm was no longer statistically significant, and recruitment location and baseline opioid use were the variables associated with this outcome.

Finally, more intervention arm participants exhibited reductions in the inpatient length of stay (212/290, 73.1%, Table 2) than control arm participants (138/226, 61.1%; p = 0.005). When adjusted by study covariables (Table 3), study arm was no longer statistically significant and age and recruitment location were the variables most associated with outcome.

Because of baseline site-level differences between the study arms, the research team and administrative data partners created a second geographic comparison group at the intervention site, using a diagnosis-based sampling frame definition developed for the study. Intervention site controls were matched with intervention recipients according to available variables of date and admission status (inpatient vs outpatient) of index encounter, complexity (CRG), age, and sex. Healthcare utilization measures retrieved for this additional cohort included ED use, inpatient admissions, and length of stay. Patterns of acute care use for the internal matched controls did not differ from those observed among the formal study control arm. Nevertheless, there may be important macroeconomic and/or healthcare throughput differences between the two cities in this study that may have influenced our results.

We employed a quasi-experimental study design for feasibility reasons; the null findings may in part be due to the lack of power inherent to this approach. Moreover, we detected baseline inequivalence in outcome measures. The multi-pronged complex intervention under study served multiple subgroups of patients (e.g. different primary substances of concern) with services tailored to individual needs. To address these differences at the individual participant level, we adjusted for observed baseline inequivalence across study arms and included clinically important covariables in the statistical models. However, a number of unmeasurable contextual factors are likely at play. We considered propensity weighting, but determined that this approach would not fully account for observed differences.

It is plausible that due to understandable limitations in the capacity of the newly formed intervention team, and benefits of the intervention perceived by referring clinicians, recruited participants were a group who not only met eligibility criteria but had a high level and acuity of need that stood out within eligible patients. This is reflected in the high level and acuity of need seen in the baseline questionnaire data and particularly high complexity (CRG) in the intervention group. Thus selection bias may have influenced our results.

The rollout of the EMCIH study at the control arm sites was met with excitement at the possibility of future intervention spread to other geographic regions within the same health system. Research staff on site were well received by staff and participants alike. A Hawthorne effect resulting from control arm participants responding favourably to their encounters with study and clinical staff following study enrolment cannot be ruled out; however, given the lack of support options available to participants beyond three structured quantitative data collection sessions and usual care, we believe a Hawthorne effect is unlikely to have exerted a significant enough influence to change acute care utilization.

The EMCIH study protocol occurred during the early intervention implementation period. The intervention has since undergone significant refinement (e.g. enhanced team composition, dedicated team expansion to the ED, addition of specific harm reduction supports such as managed alcohol programming, naloxone kit distribution, and a supervised consumption service) in response to referral patterns and consultation with program recipients and stakeholders. Intervention adaptation was also critical to address evolving population needs, most notably the advent of Canada’s opioid poisoning epidemic shortly after the launch of the intervention, and the subsequent prioritization of harm reduction and immediate access to opioid use disorder treatment options. The study findings reported herein reflect an earlier iteration of the intervention prior to intervention refinements having occurred. The team is engaged in ongoing program data collection and quality improvement work to assess the impact of such refinements.