Feasibility and impact of implementing a private care system’s diabetes quality improvement intervention in the safety net: a cluster-randomized trial

Kaiser Permanente’s “A.L.L. Initiative” (Aspirin, Lovastatin (any statin), Lisinopril (any angiotensin converting enzyme-inhibitor (ACE) or angiotensin receptor blocker (ARB)); hereafter called “ALL”) is a system- and clinic-level QI intervention. Implemented throughout KP in 2003, ALL was designed to increase rates of patients with diabetes who are appropriately prescribed cardioprotective aspirin, statins, and ACE/ARBs. (Because the evidence for aspirin changed, we did not target it here.) At KP, ALL used electronic health record (EHR)-based tools coupled with top-down strategies to incentivize provider uptake. Its overarching strategies were to facilitate providers: (1) identifying patients with diabetes who were indicated for the ALL medications but not taking an indicated medication and (2) prescribing these medications. ALL was highly successful in KP: an internal study estimated a >60 % reduction in cardiovascular disease (CVD) events among targeted adults taking the ALL medications for 1–2 years [4, 5]. We selected ALL as a “test case” for studying cross-setting translational implementation based on its strong underlying evidence [4, 6‐23], its alignment with national treatment guidelines, its impressive impact at KP, the simplicity of its strategies, and preliminary evidence that it could be adapted for the safety net [6].

To adapt ALL for implementation in CHCs, we underwent an iterative, year-long process involving researchers, electronic health record (EHR) programmers, and CHC staff and providers, as previously described [24]. In brief, while the overall intervention aims and EHR tool functions remained the same as in KP’s implementation, each intervention component was customized to fit the CHCs’ needs and EHR capabilities. Each study CHC used the combination of components that best fit its existing workflows. KP’s top-down implementation strategies [25] were likewise adapted to emphasize practice facilitation [26‐30], due to the CHCs’ different organizational structures and resources [31]. Table 1 summarizes the intervention components in both settings and how they were translated for implementation in the study CHCs. Details on the components of the implemented intervention were previously reported [24].

The 11 study CHCs are ambulatory primary care clinics managed by three Federally Qualified Health Center (FQHC) systems in the Portland, Oregon metropolitan area. All are members of OCHIN, Inc., a non-profit organization that provides health information technology to safety net clinics [32‐34]. The study CHCs vary in size and organizational structure: one is operated by a large academic medical center, though it is not located at that center; six by a county health department in urban locations; and four by a non-profit organization primarily serving suburban, Spanish-speaking populations. These clinics share a single EHR which is managed centrally at OCHIN, including regular data validation. EHR data for these analyses were extracted at OCHIN. Results from our extensive process evaluation will be reported in future manuscripts. [35]

In our staggered, cluster-randomized implementation strategy, six study clinics were randomized to implement the intervention in June 2011 (early implementation), and five 1 year later (late implementation). Cluster randomization was used because this is a clinic-level intervention with clinic-level outcomes. Randomization was matched on size of the clinics’ patient population and the FQHC system operating the clinic. The intervention’s effect was evaluated using an interrupted time-series design, analyzed with segmented regression models with calendar months as the unit of analysis [36].

Our outcome of interest was the proportion of patients with diabetes who were indicated for cardioprotective medication(s) (denominator) and had an “active” prescription for the indicated medication(s) (numerator), calculated monthly. We measured two clinic-level prescribing rates: (1) the proportion of patients indicated for ACE/ARB and statin who had an active prescription for both, and (2) the proportion indicated for statin only who had an active statin prescription.

The early implementation effects on the two prescribing rates of interest were estimated using segmented regression models with two groups—the early implementation clinics were modeled as the intervention group and the late implementation clinics as the control group. We used data from 24 monthly intervals: 12 in the pre-intervention period (June 2010–May 2011) and 12 in the post-intervention period (June 2011–May 2012). Models included a constant term, a term to model the slope of the pre-intervention linear trend, a variable to indicate study group (intervention vs. control), terms to estimate change in level from pre- to post-intervention and change in slope of the trend, and interaction terms (group by pre-intervention trend, group by change in level, and group by change in trend) to test for significantly greater improvement in prescribing rates in the early implementation clinics, compared to the controls. The change in level provides an estimate of the immediate effect of the intervention; the change in trend, an estimate of its effect across time post-implementation.

Analyses of late implementation effects involved the late implementation clinics only, with level and trend of the pre-intervention period serving as controls for the post-intervention period. These analyses included 36 monthly intervals: 24 in the pre-intervention period (June 2010–May 2012) and 12 in the post-intervention period (June 2012–May 2013). Models contained a constant term, a term for the slope of the pre-intervention linear trend, and terms estimating pre/post-intervention changes in level and trend of outcome rates.

Preliminary qualitative analyses indicated that providers modified ALL medication prescribing based on patient age and gender. (Because of the teratogenic risks, some providers were wary of prescribing to younger females, and others were reluctant to prescribe statins to younger patients in general; details of these qualitative results pending in future papers). To account for potential differences in prescribing rates based on distribution of these factors across clinic groups, we adjusted all models for gender (percentage of denominator who were female at each time point), age (percentage of denominator age 18–39 on the first day of each month), and their interaction. All regression models also included a first-order autoregressive term to control for serial autocorrelation [36]. All analyses reflect tests of statistical significance with a two-sided α of 0.05 and were conducted using SAS Enterprise Guide 6.1 (PROC AUTOREG; SAS Institute Inc., Cary, NC, USA).

There is a known need to expedite the dissemination of effective interventions across all care settings [38‐40]. Doing so would facilitate the spread of proven interventions and QI approaches and reduce the need for care delivery systems to develop their own. Although this dissemination would be particularly useful to under-resourced clinics serving vulnerable populations in the USA and elsewhere, such clinics have historically been under-studied in dissemination and implementation science [41]. Instead, most previous QI efforts in CHCs and similar clinics were internally developed (a few exceptions cited here), and most cross-setting implementation research has focused on translation across similar care settings [28, 30, 41‐48].

We believe this was the first clinical trial of the feasibility and impact of translating a QI intervention developed and shown effective in a private, integrated care setting, for implementation in under-resourced clinics. We showed that such translation and implementation is feasible but may require substantial adaptation to meet local needs and structures. In brief, we adapted the intervention components for implementation in the study clinics, as directed by an iterative process involving clinic staff. KP’s key strategies—making it easier to identify patients “missing” an indicated medication, and to prescribe that medication—remained the same; we adapted the specifics of how these strategies were implemented (including adapting the tools) and supported [24, 31].

Lessons learned about adapting QI interventions for implementation in under-resourced clinics include: (i) Consider the strategies used to support uptake of an adapted intervention [25]. Here, KP used top-down directives coupled with financial incentives; the CHCs used on-site facilitation. Though not a difference in the intervention itself, this could influence its uptake. (ii) Clinic cultures and leadership styles (e.g., degree to which top-down directives are issued and followed) can influence adoption of practice change initiatives, and should be considered when adapting such interventions. (iii) Though difficult and time-consuming, collaborative decision-making by clinic leaders (related to how to adapt the intervention) may be essential to eventual uptake. (iv) Ensure that the intervention aligns with the clinic’s standards of care; if possible, integrate it into the official standard of care. (v) Consider that low-income and otherwise vulnerable patients can face barriers to acting on medical recommendations—barriers not easily addressed through alerts and panel tools—and adapt as possible to address these barriers.

The demonstrated feasibility of effectively implementing a proven intervention in CHCs has important implications for safety net clinics in the USA and internationally. As such clinics rarely have the resources to develop “home-grown” interventions, implementing those shown effective elsewhere could yield important efficiencies in efforts to improve health care quality and outcomes. Such cross-setting translation should be considered a realistic means of helping lower-income clinics implement cutting-edge, effective QI strategies. Although implementing strategies from other settings requires adaptation to meet the clinics’ needs, it is likely more efficient than these clinics developing their own “from scratch”—an efficiency that could support dissemination of effective strategies across CHCs and similar care settings.

Comparing the impact of implementing ALL in the study CHCs versus KP is difficult for several reasons (some parallel study limitations, below). KP targeted patients with diabetes aged ≥55 and/or those with coronary artery disease at high risk for CVD, whereas in the study CHCs, the intervention targeted all adult patients with diabetes (an adaptation the CHCs requested). KP was able to measure uptake of ALL using dispense and refill data; we could only access prescribing data from OCHIN’s EHR, which until recently lacked medication dispensing data. Thus, this study would consider a patient’s care as appropriate even if the prescribed medication was never dispensed—a method of assessing guideline concordance that was less stringent by necessity. (Such primary non-adherence occurs in about 5 % of KP patients with diabetes [49] but was likely higher among uninsured CHC patients—preliminary qualitative analyses suggest that even small co-payments for prescriptions created an important barrier). KP started with considerably lower rates of concordance with targeted outcomes and at a time when there was less evidence supporting the medications’ effectiveness. When KP implemented ALL in 2003, about 33 % of their target population were appropriately taking the ALL medications; this rose to 52 % by 2005, a 58 % relative increase [4]. In the study CHCs, at baseline, about 48 % of patients indicated for an ACE/ARB and a statin were prescribed both. One year post-intervention, this increased to 62 % in the early clinics and 57 % in the late clinics. While we had thus limited ability to directly compare the impact of implementing ALL in KP versus the study CHCs, our results nevertheless show significant improvement in the CHCs.

The relative increases in guideline-concordant prescribing reported here range from 16–39 %. While these improvements are significant, the overall rates of appropriate prescribing ranged from 57–67 % by the end of the reporting period, meaning that 33–43 % of patients still did not receive prescriptions that were considered indicated. We believe this highlights limitations in the field’s current technical capacity and understanding of how to build effective EHR-based QI/decision support interventions. From a technical perspective, patients may be incorrectly identified as indicated for a given medication for myriad reasons that are potentially difficult to capture in an EHR-based algorithm, including: patient seeks pregnancy, complex concurrent medical conditions, potential medication interactions, and medication(s) prescribed elsewhere/entered in the chart in a way that does not trigger the algorithm. This illustrates the challenges inherent to striking a balance between false positives and false negatives when creating decision support algorithms in complex populations. This may also suggest that the implemented intervention was more successful than it appears, because some patients in the rate denominators could not be given the target medications. We plan to conduct further research on optimizing decision support tools adapted for safety net clinics.