Frederick J. Bloom Jr, MD, MMM

 To determine whether diagnostic coding shifts might undermine apparent improvements resulting from the 2007 Healthcare Effectiveness Data and Information Set (HEDIS) measure on avoidance of antibiotics for the treatment of adults with acute bronchitis (

International Classification of Diseases, Ninth Revision, Clinical Modification

Time series analysis within a primary care network for 3 successive winter seasons from 2006 to 2009. Methods: All initial adult visits with a primary diagnosis code of 466.0 or 490 (bronchitis, not otherwise specified) were analyzed. Multivariable analysis accounted for clustering of observations by physician.

 The percentage of visits treated with antibiotics declined significantly for code 466.0 (76.8% to 74.4% to 27.0% of visits over the 3-year study period; 

 <.0001 for trend) but did not decline for code 490 (86.6% to 87.6% to 82.1% of visits; 

 = .33 for trend). Use of the 490 code rose significantly over the study period, from 1.5% of total bronchitis visits in year 1 to 84.6% of total bronchitis visits in year 3. As a result, the odds of an antibiotic prescription for codes 466 and 490 combined decreased slightly in year 3 compared with year 1 (odds ratio 0.88; 95% confidence interval 0.78-0.99).

 While performance on the specific HEDIS measure improved dramatically during this study period, overall antibiotic prescribing did not decline substantially. Quality measures that assess performance on specific diagnosis codes are imperfect and do not account for shifts in diagnosis coding.

Quality measures that are linked to specific disease diagnoses can create unintended shifts in diagnosis patterns that offset apparent gains in quality.

		Diagnosis shifts are important consequences of quality measures that have narrow diagnostic specificity.

		Programs should examine both intended and unintended effects of quality measures, specifically examining changes in behavior for related conditions not captured by the measure.

Acute respiratory tract infections (ARIs), including acute bronchitis, are the most common reasons for outpatient office visits by adults in the United States.

 Although numerous studies and guidelines have advised against routine use of antibiotics for these predominantly viral illnesses, 75% of all antibiotics prescribed in community settings are for ARIs.

 Extensive data have shown that overprescription of antibiotics is a major contributor to emerging antibiotic resistance,

is costly,5 and exposes patients to potential side effects.6 Efforts utilizing public health campaigns, provider education, and practice guidelines like the Centers for Disease Control and Prevention’s “Principles for Appropriate Antibiotic Use for ARIs in Adults”7 have attempted to reduce antibiotic use for ARIs with varying degrees of success.

 Antibiotic prescriptions for acute bronchitis have been particularly refractory to change.

 Fewer than 10% of cases of acute bronchitis are bacterial in etiology, and antibiotic treatment has been shown to have minimal, if any, benefit.

 Guidelines state that acute bronchitis in healthy adults does not require antibiotic treatment, yet rates of antibiotic use remain very high. National Committee for Quality Assurance data from 2006 estimated that 71% of commercially insured patients diagnosed with acute bronchitis received antibiotics.

In recognition of this issue, avoidance of antibiotics for the treatment of adults with acute bronchitis (defined as 

] code 466.0) was included as a component of the Healthcare Effectiveness Data and Information Set (HEDIS) reported to the National Committee for Quality Assurance in 2007. Data obtained from HEDIS indicators are used by employers and consumers to compare quality measures across health plans, and can have an impact on pay-for-performance reimbursement at both the individual and practice level. Desire to score well on the HEDIS measure is an additional financial incentive for providers and insurers to reduce antibiotic use for bronchitis. However, it remains to be seen if the HEDIS measure will result in true reductions in antibiotic use, or if clinicians will alter diagnostic coding for visits to justify antibiotic prescribing.

Using data from a large, integrated health system, we examined trends in antibiotic use for acute bronchitis from 2006 to 2009, years surrounding the implementation of the HEDIS acute bronchitis measure. In this network, an effort to improve quality of care and performance on the HEDIS measure led to an electronic health record—based intervention that attempted to provide real-time feedback to the physician that antibiotics were not necessary in uncomplicated acute bronchitis. The display text for code 466.0 was changed to “ACUTE BRONCHITIS (ABX NOT INDICATED)” and for code 490 it was changed to “ACUTE BRONCHITIS (COMPLICATED).” We first sought to determine current rates of antibiotic use for acute bronchitis within this health system. We then examined whether shifts in coding patterns occurred during this period, and if the shifts were associated with antibiotic prescription rates. Finally, we sought to determine which patient and provider factors were associated with receipt of an antibiotic in order to identify potential targets for future interventions.

We conducted an analysis of antibiotic prescription patterns for adults with ARIs over a 3-year time period. The initial study population included all adults who sought care for any ARI at any of 33 primary care practice sites within the network.

Data for these analyses were collected from the integrated electronic health record that is utilized across all outpatient and inpatient practice sites within the healthcare system. We extracted data for all visits meeting inclusion criteria for the 3 successive periods of October 1 through March 31 for the years 2006 to 2009. Eligible encounters included outpatient visits to any of the 33 primary care practice sites by adult patients with an ARI (defined below). For each visit, we extracted the date of the visit, primary care site, patient demographic variables (age, sex), primary and secondary diagnoses for the visit, documented vital signs, and the presence of specific comorbidities based on the electronic problem list. In addition, we captured the prescription of an oral antibiotic at each eligible visit based on documentation in the electronic health record.

Visits for ARIs were identified based on the primary diagnosis code, using any of the following 

 codes: 460.0 through 487 (acute respiratory infections,

other diseases of the respiratory tract, pneumonia, and influenza), 490 (bronchitis not specified as acute or chronic, hereafter bronchitis NOS), 034.0 (streptococcal

sore throat), 079.0 (adenovirus), 784.1 (throat pain), and 786.2 (cough). For the analyses in this study, we limited our visits of interest to those with a primary diagnosis code of 466.0 (acute bronchitis) and 490 (bronchitis NOS). In addition, we excluded any visits for acute bronchitis within 30 days of a prior visit for an ARI as defined above. Additional exclusion criteria, based on the HEDIS measure, included age <18 years or >65 years, and the presence of specific comorbidities and/or antibiotic-responsive secondary diagnoses. Specific comorbidities included HIV, cystic fibrosis, malignancy, chronic bronchitis or emphysema, or other chronic respiratory diseases, and were identified based on the problem list maintained in the electronic health record and documented as of the date of visit. Antibiotic-responsive secondary diagnoses were documented at the time of the visit for acute bronchitis. The full list of diagnoses is included in the 

Analyses were limited to visits with a primary diagnosis of either acute bronchitis (466.0) or bronchitis NOS (490). Descriptive statistics were performed to examine the distribution of patient characteristics across all participating sites. We next performed unadjusted comparisons of antibiotic prescription rates across each of the study years using the Mantel-Haenszel chi-square analysis for trend. Analyses were conducted for visits coded as 466.0 and 490 separately and combined. Next, generalized estimating equations (PROC GENMOD) were used to control for clustering of antibiotic prescribing patterns by provider, and to evaluate independent predictors of antibiotic prescribing, including study year and

 diagnosis code. All statistical computations were performed using the SAS statistical software program (version 9.2; SAS Institute Inc, Cary, North Carolina). This study was approved by the Institutional Review Board at the participating clinical site and at the University of Pennsylvania and University of California, San Francisco.

There were 18,542 eligible visits for acute bronchitis (combining 

 codes 466.0 and 490) in the years 2006-2009. 

 displays the total number of visits for bronchitis; when stratified by 

 code, visits for bronchitis were coded almost exclusively as 466.0 for the years 2006-2007 (98.5% of cases) and 2007-2008 (98.4% of cases). The electronic health record diagnosis display text was changed in October 2008. In 2008-2009, the frequency of visits coded as 466.0 declined, and the use of code 490 rose significantly and accounted for a total of 84.6% of visits for bronchitis (

 displays the characteristics of patients

diagnosed with acute bronchitis, stratified by study year and diagnosis code. Overall, vital sign abnormalities were generally uncommon among patients across the 3 study years and comparing both diagnosis codes. Similarly, the documented presence of other comorbidities (not considered exclusion criteria for the HEDIS measure) was uncommon among patients diagnosed with either 466.0 or 490.

summarizes factors associated with the use of code 466.0 versus code 490. In both unadjusted and adjusted models, the odds of the 490 diagnosis went up substantially in year 3 compared with year 1 (odds ratio [OR] 370; 95% confidence interval [CI] 97-1416). Female patients had a lower odds of diagnosis with 490 versus 466.0 (OR 0.7; 95% CI 0.6-0.9); and current smokers had a higher odds of diagnosis with 490 versus 466.0 (OR 1.38; 95% CI 1.16-1.65). No other patient characteristics were associated with the choice of 466.0 versus 490 diagnosis codes.

For 2006-2009, antibiotics were prescribed for 74.9% of all patients with acute bronchitis (466.0 + 490 codes combined), and the overall rate declined from 76.9% in 2006-2007 to 74.6% in 2008-2009 (

). Compared with the first year of the study, the odds of antibiotic prescriptions were slightly lower in the second year (OR 0.9; 95% CI 0.80-1.01) and third year (OR 0.88; 95% CI 0.78-0.99) of the study (

). Across all 3 years, antibiotic prescriptions were significantly more common for visits coded as 490 (82.3% of visits) than for visits coded as 466.0 (71.5% of visits) (OR 2.18 for 490 vs 466.0; 95% CI 1.92-2.47; 

 <.0001). When analyzed by visit year, the percentage of visits where antibiotics were prescribed was consistently high for visits coded as 490 at 86.6%, 87.6%, and 82.1% of visits in 2006-2007, 2007-2008, and 2008-2009, respectively (

 = .33 for trend). In contrast, the percentage of visits where antibiotics were prescribed was high for visits coded as 466.0 in years 2006-2007 (76.8%) and 2007-2008 (74.4%), but dropped substantially to 27.0% in years 2008-2009 (

No patient factors were significantly associated with the odds of an antibiotic prescription, including patient age, sex, ethnicity, current smoking, or presence of comorbidities (Table 3). However, in comparing provider types, patients seen by a nurse practitioner or physician assistant were more likely to have an antibiotic prescription compared with patients seen by a physician (OR 2.77; 95% CI 1.69-4.52).

Acute bronchitis is a very common reason for outpatient office visits.

 Extensive data support avoidance of antibiotics for uncomplicated acute bronchitis,

 and quality improvement measures such as HEDIS have been designed to incentivize evidence-based practice for this condition.

 In this study of outpatient visit data from an integrated health plan, we found that a large decrease (ie, improvement) in antibiotic prescribing for acute bronchitis (466.0) as defined by the HEDIS measure was largely accounted for by diagnostic coding shifts. Antibiotic use for visits coded as bronchitis NOS (490) remained high throughout the period 2006-2009, and antibiotic prescription rates for the combined group of

 codes 466 and 490 showed little change over time.

The high rates of antibiotic use (74.9% overall) for acute bronchitis seen in our study are comparable to those found in earlier studies with HEDIS plans as well as with national ambulatory office visit surveys.

 Previous studies have elucidated many provider, patient, and health system factors that contribute to continued antibiotic use for bronchitis, including physician diagnostic uncertainty (ie, use of antibiotics due to difficulty in ruling out community-acquired pneumonia),

 and patient demographics and knowledge factors, and expectation of antibiotic treatment.

 Efforts at provider and patient education including decision support tools and management algorithms have shown some efficacy in lowering rates of antibiotic use.

The HEDIS bronchitis measure has significant potential to alter practice patterns, as health systems and providers have a financial incentive to prescribe fewer antibiotics. However, it is crucial to ensure that improved performance on the HEDIS measure translates to a true reduction in antibiotic use. This study suggests that diagnostic coding patterns should be analyzed when assessing HEDIS outcomes, and perhaps should be considered in future revisions of the HEDIS bronchitis measure.

The very simple electronic health record coding intervention had a significant effect on the use of antibiotics for acute bronchitis (466.0) after implementation in the third year of the study. However, this small effect was overshadowed by the large impact of coding changes on the overall use of antibiotics for patients with bronchitis. The use of diagnostic coding pattern shifts to improve ratings is a potential concern in the reporting of HEDIS measures. A study from 2000 found that after an educational intervention to reduce antibiotic use for acute bronchitis among family physicians, rates of antibiotic prescription declined, and there was no significant shift in coding from acute bronchitis to other ARIs such as sinusitis or pneumonia; however, coding shifts between acute bronchitis and bronchitis NOS were not assessed.

Because our study was observational, we cannot conclude that the implementation of the HEDIS measure and health system response were directly responsible for the observed code shifting and change in antibiotic prescribing patterns. It is possible that other factors influenced the change from use of code 466.0 to use of code 490, such as differences in patient presentation over the years, or concerns about postinfluenza bacterial illnesses in the wake of the influenza epidemic in 2008-2009. In addition, it is possible that provider turnover during the study period may have contributed to the observed change in coding patterns noted in 2008, if newer providers had different coding habits. Another limitation involves the abstraction of data on bronchitis diagnosis and antibiotic use from medical records, which may overestimate or underestimate actual antibiotic use rates depending on errors in the coding. However, previous studies to determine the accuracy of quality measurements such as HEDIS have shown that review of 

 coding and claims data provides information comparable to that obtained by full chart review and that claims data review appears to underestimate actual antibiotic use by providers.

Our study highlights the importance of ensuring that quality assessment measures obtain a complete picture of their designated outcomes. While the HEDIS bronchitis measure is an important step to encourage adoption of evidence-based management of acute bronchitis and reduce antibiotic overuse, the potential to work around its data collection measurements, either intentionally or accidentally, should be acknowledged. In future versions of the HEDIS bronchitis measure, the inclusion of bronchitis NOS (490) may be useful to obtain a complete assessment of antibiotic practice patterns.

 From the Department of Medicine (SR, JPM), Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA; Department of Medicine (RG, JHM), University of California, San Francisco, San Francisco, CA; and the Geisinger Health System (TH-A, FJB, TG, MSS), Danville, PA.

Supported by grant R01-CI000611 from the Centers for Disease Control and Prevention and grant K24-AI073957 from the National Institutes of Health.

Articles

A quality measure reduced antibiotic use for patients with acute bronchitis but led to use of an alternative diagnosis, offsetting most of the observed improvement.

Unintended Consequences of a Quality Measure for Acute Bronchitis

 To estimate cost savings associated with ProvenHealth Navigator (PHN), which is an advanced model of patient-centered medical homes (PCMHs) developed by Geisinger Health System, and determine whether those savings increase over time.

A retrospective claims data analysis of 43 primary care clinics that were converted into PHN sites between 2006 and 2010. The study population included Geisinger Health Plan’s Medicare Advantage plan enrollees who were 65 years or older treated in these clinics (26,303 unique members).

Two patient-level multivariate regression models (with and without interaction effects between prescription drug coverage and PHN exposure) with member fixed effects were used to estimate the effect of members’ exposure to PHN on per-member per-month total cost, controlling for member risk, seasonality, yearly trend, and a set of baseline clinic characteristics.

 In both models, a longer period of PHN exposure was signifi cantly associated with a lower total cost. The total cumulative cost savings over the study period was 7.1% (95% confi dence interval [CI] 2.6-11.6) using the model with the prescription drug coverage interaction effects and 4.3% (95% CI 0.4-8.3) using the model without the interaction effects. Corresponding return on investment was 1.7 (95% CI 0.3-3.0) and 1.0 (95% Cl —0.1 to 2.0), respectively.

 Our fi nding suggests that PCMHs can lead to significant and sustainable cost savings over time.

 Geisinger Health Systems has implemented its own version of patient-centered medical homes called ProvenHealth Navigator (PHN) since 2006. Our results demonstrate that the longer the member has been exposed to PHN sites, the lower the cost of care. The return on investment (ROI) has not yet signifi cantly exceeded the break-even point. However, because the magnitude of savings from PHN depends on the length of members’ exposure to PHN, it remains to be seen whether the cumulative ROI can eventually exceed the break-even point.

 Cost savings are achievable by redesigning primary care, but it takes time to reap the benefits.

At this time of national focus on the need to simultaneously improve quality and reduce cost, novel solutions are needed. The fragmented healthcare delivery system further adds to the challenge.

 The key stakeholders—physicians, patients, health insurers, and hospital administrators, as well as the federal government&mdash;now recognize the need for healthcare redesign to increase the quality of care while containing the cost of care.

In particular, reorienting the current healthcare delivery system to one in which greater emphasis is given to more effective primary care is likely to be an important solution to the problem.

Recently, patient-centered medical homes (PCMHs) have gained attention as a way to re-engineer the care-delivery process. In the broadest sense of the term, PCMHs refer to “provision of comprehensive primary care services that facilitates communication and shared decision-making between the patient, his/her primary care providers, other providers, and the patient’s family.”

 There is early but growing evidence that PCMHs have the potential to improve care for individuals, improve health of populations, and slow the growth in costs of healthcare by reviving primary care and bringing the patient and family to the center of the care-delivery system

The ultimate goal of PCMHs is to improve patient outcomes via redesigned primary care. The preventive aspect of primary care enhances the value proposition of PCMHs as a long-term intervention. Such an emphasis on primary care, therefore, implies that PCMHs may have the desirable effect of reducing cost of care. The early experiences of the PCMH demonstrations strongly suggest this.

 In this study, we explicitly explore this possibility by evaluating the effect of ProvenHealth Navigator (PHN), an advanced model of PCMHs developed and implemented by Geisinger Health System (Geisinger), since 2006 on cost of care.

PHN consists of the following 5 core program components: (1) patientcentered primary care, (2) integrated population management, (3) medical “neighborhood”

 (ie, alignment of key community partners, such as home health agencies, skilled nursing facilities, outpatient/ancillary services, hospital facilities, and community pharmacies), (4) comprehensive quality improvement, and (5) value-based reimbursement redesign that includes a quality, outcome-based pay-for-performance program. More detailed descriptions of how the PHN intervention was designed and implemented are published elsewhere.

PHN is designed to move resources further “upstream” in the primary care setting to reduce “downstream” costs from the highest acuity settings. In this system, primary care is considered upstream in the sense that it serves as the starting point in the chain of care delivery. In general, care becomes more expensive as patients move downstream to specialty and inpatient care. Therefore, successful upstream efforts are expected to result in reductions of inpatient care—related costs due to fewer uncontrolled exacerbations of chronic diseases and more effective care transitions that prevent hospital readmissions and unnecessary duplication of services.

A PHN site refers to one of the primary care clinics that has been designated as such by Geisinger and has undergone significant changes in its management and operations in accordance with the PHN practice redesign. PHN was implemented in several phases over the 5-year period from late 2006 through 2010. At its core, the PHN intervention can potentially reduce cost over time while improving quality via:

• Prevention: redesigned primary care using automation to leverage resources and enhance reliability for early detection and interventions prior to exacerbations and complications

• Chronic disease optimization: redesigned primary care using a high-touch, high-technology approach to manage emerging exacerbations and to minimize complications

• Comprehensive care management: health-plan— trained, practice-embedded “concentrated care–RNs” who serve as nurse case managers focused on proactive identification and management of the more medically complex patients

This analysis focuses on individual members’ exposure to PHN by measuring how long a member has been exposed to a PHN site and examining whether a prolonged exposure to PHN is associated with a lower cost of care over time. It is expected that the longer the patient has been exposed to a PHN site, the greater the cost savings.

, a number of primary care clinics in the Geisinger Health Plan (GHP) provider network were selected and converted to PHN sites during each phase of the intervention. As a result, primary care clinics became PHN sites at different times. We exploit this variation in time of PHN conversion to estimate the effect of PHN exposure on total cost. The fi gure also shows the trends in average permember per-month (PMPM) total cost among clinics in each PHN implementation phase. For baseline clinic characteristics in each phase, see 

Our data originated from GHP’s claims database covering the period between January 1, 2006, and December 31, 2010. The unit of observation was member-month (ie, there was a unique record for each member for each month during which the patient was a GHP member). If the member did not have any claims for that month, claim amounts were recorded as $0. Our sample was restricted to GHP’s Medicare Advantage plan members who were at least 65 years of age during the study period (because PHN had focused primarily on the Medicare population during this time) and enrolled in 43 primary care clinics that eventually became PHN sites by the end of 2010 (37 with physicians employed by Geisinger and 6 non- Geisinger primary care practices). About 30% of all patients treated by practices with Geisinger-employed physicians are covered by GHP.

Two measures of PHN exposure were developed: First, a member-level PHN exposure measure was calculated as the number of months a member had been enrolled in a PHN clinic as of a given time. This exposure measure was then broken into 5 categories (0, 1-6 months, 7-12 months, 13-24 months, and >24 months) in order to capture any non-linear relationship between the cost of care and PHN exposure. Second, a clinic-level PHN exposure measure was also calculated as the number of months in which the clinic had been a PHN site as of a given period since its PHN conversion on the phase start date as shown in Figure 1.

The clinic-level PHN exposure variable is intended to capture the degree to which there might have been learning at clinics over time in accordance with PHN’s rapid-cycle innovation principle.

This is in recognition of PHN’s being a dynamic program that has undergone continuous modifi cations over time while maintaining its core components. As such, it is diffi cult to determine exactly at what point the PHN sites transitioned from practice transformation and redesign to fully functioning PHN sites. Nevertheless, we hypothesize that the longer a clinic has been designated a PHN site, the greater its impact on cost of care.

In addition to the member-level and clinic-level PHN exposure variables as described above, the covariates included age and hierarchical condition categories (HCC) risk score to capture each member’s case mix along with year-indicator variables to capture the temporal trend in total cost over time as well as month-indicator variables to capture seasonality. Other covariates included a set of baseline clinic characteristics as of 2005 (ie, percentage of male patients and average patient age in 2005, along with average total cost, HCC score, and number of inpatient admissions and readmissions in 2005). These variables were intended to capture any underlying differences among different clinics which may be correlated with each member’s total cost during the study period.

The main outcome measure was the total cost of care, defined as the total allowed amount (plan payment to providers plus copayments) for a given member in a given month (per-member-per-month [PMPM]) which included payments for inpatient and outpatient facility and professional services, as well as prescription drugs. Less than 2% of the total member- month observations in our sample had 0-value total cost, because capitation payments for certain services, such as behavioral health, were refl ected in the total cost calculations.

Because there are multiple observations for each member, a member fi xed-effects model was used to account for these repeated measures as well as to remove any unobserved and time-invariant member characteristics that could have biased our estimates. In this model, we exploit the within-member variation in the total cost and the PHN exposure variable over time. Thus, each member acts as his or her comparison. In other words, the question that our model seeks to answer may be stated as the following: relative to his or her own cost at 0 PHN exposure, what is each member’s expected total cost at 1 to 6 months, 7 to 12 months, 13 to 24 months, and greater than 24 months, of exposure?

Moreover, because members’ total costs vary depending on whether they have prescription drug coverage or not, a binary indicator variable for prescription drug coverage status during each month was also included as a covariate. Because of the fixed effects model used in this analysis, our estimate of the drug coverage effect relied on variation in the prescription drug coverage variable due to the change in each member’s drug coverage status over time. About 32% of the members picked up the drug coverage and dropped it at some point, or vice versa.

We hypothesized that the effect of PHN on total cost may differ depending on whether a member has prescription drug coverage in each period (ie, an interaction effect between prescription coverage and PHN exposure). For instance, a case manager might fi nd out whether his or her members have drug coverage or not, and if they have coverage, the case manager might encourage them to become more adherent to their drug therapy and therefore incur greater total cost. As such, we estimated 2 models: one with and another without the interaction effects. These 2 models might yield different estimates of total cost savings attributable to PHN, because in the model with interaction effects, the estimates depend both on the level of each member’s PHN exposure and the proportion of members who have drug coverage in each period. In the model without the interaction effects, the estimates depend only on the PHN exposure level.

Using the parameter estimates obtained from the fi xedeffects model, we calculated the expected total cost if none of the members had any PHN exposure (ie, member as well as clinic-level PHN exposure of 0 months) and compared them against the total costs given the actual levels of PHN exposure for each member in each given period (ie, the “observed” cost). This allowed us to estimate the effect of the PHN exposure on cost in terms of percent savings (see eAppendices A and B for the full regression model parameter estimates). The mean percent savings and the standard errors used to construct the 95% confi dence intervals (CIs) were obtained via bootstrap method with 200 replications.

As a sensitivity check, we also obtained a number of alternative estimates using different sample defi nitions. In particular, we obtained estimates with a restricted sample consisting of only those who did not switch their primary care clinics during the study period (about 80% of the full sample size). This allowed us to control for any clustering effects due to similarities among members enrolled in same clinics

 by including member-clinic fi xed effects in our regression model. The resulting estimates (available upon request) were similar to the reported estimates shown below.

, the fi nal sample consisted of 26,303 members over the 5-year period, corresponding to 1,053,445 member-month observations in 43 primary care clinics that were designated as PHN sites at different points in time between 2006 and 2010. The median age of the members was 76 years, and 42% were male. Slightly less than half of the member-month observations fell in the 0 member-level PHN exposure category; about 12% fell in the highest category of greater than 24 months of PHN exposure. Table 1 also indicates that greater PHN exposure is associated with higher total cost and age, as well as greater proportion of membermonths with prescription coverage. This suggests that the effect of PHN exposure on total cost is confounded by member age and presence of prescription drug coverage benefit.

On average, GHP members in our sample maintained their membership for about 40 months out of the maximum possible 60 months, suggesting a stable enrollment pattern. About 20% of the members in our sample switched from a non-PHN site to a PHN site during our study period, while only about 1% switched from a PHN site to a non-PHN site.

 shows the regression coeffi cients and their corresponding 95% CIs for the key covariates in the 2 models (with and without the interaction effects between drug coverage and PHN exposure). In both models, the coeffi cients on the PHN exposure variables are consistently negative and get increasingly larger as the length of exposure increases, suggesting that longer PHN exposure is consistently associated with lower total cost. However, in contrast, the coeffi cient estimates on the interaction terms between drug coverage and PHN exposure are consistently positive, suggesting that there is a significant interaction between them.

shows the estimated percent savings and the corresponding bootstrapped 95% CIs as obtained from the regression model parameter estimates shown in Table 2. Table 3 clearly supports the hypothesis that a longer exposure to PHN is associated with lower total cost and therefore greater savings, regardless of whether the interaction effects were included in the model. The largest and statistically most signifi cant percent saving was achieved in the highest category of PHN exposure (>24 months) in both models.

ProvenHealth Navigator, Geisinger's version of advanced patient-centered medical homes, is associated with significant savings in total cost of care over time.

Reducing Long-Term Cost by Transforming Primary Care: Evidence From Geisinger's Medical Home Model

The primary care medical home has been promoted to integrate and improve patient care while reducing healthcare spending, but with little formal study of the model or evidence of its efficacy. ProvenHealth Navigator (PHN), an intensive multidimensional medical home model that addresses care delivery and financing, was introduced into 11 different primary care practices. The goals were to improve the quality, efficiency, and patient experience of care.

To evaluate the ability of a medical home model to improve the efficiency of care for Medicare beneficiaries.

Observational study using regression modeling based on preintervention and postintervention data and a propensity-selected control cohort.

Four years of claims data for Medicare patients at 11 intervention sites and 75 control groups were analyzed to compute hospital admission and readmission rates, and the total cost of care. Regression modeling was used to establish predicted rates and costs in the absence of the intervention. Actual results were compared with predicted results to compute changes attributable to the PHN model.

ProvenHealth Navigator was associated with an 18% (P <.01) cumulative reduction in inpatient admissions and a 36% (P = .02) cumulative reduction in readmissions across the total population over the study period.

Investing in the capabilities of primary care practices to serve as medical homes may increase healthcare value by improving the efficiency of care. This study demonstrates that the PHN model is capable of significantly reducing admissions and readmissions for Medicare Advantage members.

ProvenHealth Navigator (PHN), a multidimensional medical home model, was introduced into 11 Geisinger Health System primary care practices with the goal of improving the quality, efficiency, and patient experience of care for Medicare Advantage patients.

This transformed primary care model resulted in significantly fewer hospital admissions (18%) and readmissions (36%) when measured across the entire population.

Total care costs for the entire PHN population decreased 7%, but this decrease did not achieve statistical significance.

Medical home models seeking to change cost trends require a multidimensional transformation of primary care practice with intensive case management and a payer partnership.

Both policy makers and private payers in the United States have begun to recognize that improving care coordination across the fragmented healthcare delivery system is essential  to improve the quality and affordability of care. Related efforts include recent Medicare demonstrations examining the impact of external disease management programs and  payment reforms that reward integrated care organizations.

 An alternative approach—the patient-centered medical home&mdash;involves enhanced primary care practices as the locus  of integration and coordination of care. A version of the medical home model was originally described by the American Academy of Pediatrics and has more recently been refined   and delineated by a coalition of professional societies through a set of joint principles.

Elements of and antecedents to the current concept of the medical home have been shown to be associated with higher quality of care and patient experience.

Closely related work on Wagner’s Chronic Care Model suggests the potential of this approach to improve the quality of care for patients with chronic conditions and prevent costly acute care.

 However, empirical evidence is scant about whether efforts to transform practices into medical homes will improve quality or yield healthcare cost savings. Two recent studies  demonstrate both the challenges of practicing transformation in primary care and the potential benefits in terms of patient and provider experience as well as preventable acute care utilization.

In this study, we evaluate the impact of a medical home model, Proven-Health Navigator (PHN), introduced for Medicare Advantage enrollees in 11 practices owned by Geisinger  Health System (GHS) in Pennsylvania. ProvenHealth Navigator is a new model of care designed to improve the quality, efficiency, and patient experience of care. It functions as a  partnership between participating primary care practices and Geisinger Health Plan (GHP). Central to the model is the transfer of population management capabilities, including  nurse case managers, from the health plan to the practice sites. This report focuses on the impact of the PHN on hospitalization and healthcare spending compared with a matched  set of practices, using 2 years of preintervention and 2 years of postintervention data.

We assembled medical claims data for services provided from January 1, 2005, through December 31, 2008,and paid through June 30, 2009, and demographic information for 15,310 members of GHP’s Medicare Advantage product. Using these data, we identified all claims for enrollees who were cared for by physician practices that implemented the  PHN model and enrollees who were cared for by matched physician practices during the study period. Enrollees who switched physician practices during the study period were  excluded from the analyses. Continuous enrollment was not required for inclusion in the study sample; analyses conducted on the continuously enrolled subpopulation yielded  qualitatively similar results.

This analysis was approved by the GHS institutional review board.

Located in rural northeastern and central Pennsylvania, GHS is a not-for-profit, integrated healthcare organization comprised of the Geisinger Clinic, which has nearly 800 employed physicians; 2 acute tertiary/quaternary care hospitals; GHP, which serves 190,000 commercial and 38,000 Medicare Advantage members; and numerous other clinical programs and facilities. Geisinger Health Plan also utilizes a network of more than 18,000 non-GHS providers and 80 non-GHS hospitals.

Geisinger Health System has an electronic health record (EHR) implemented systemwide for all ambulatory and inpatient care. This EHR also is used by GHP case managers and patients. These EHR capabilities were operational in all participating practices for several years prior to the launch of the PHN. All Geisinger-owned primary care practices, including the PHN sites, participated in a preexisting, EHR-enabled quality initiative to improve preventive, diabetes, and coronary artery disease care.

The PHN model has 5 functional program components: (1) Patient-Centered Primary Care Team Practice, (2) Integrated Population Management, (3) Micro-delivery Systems, (4) Quality Outcomes Program, and (5) Value Reimbursement System (

). A more detailed description of each component, with a comparison to the National Committee for  Quality Assurance (NCQA) Physician Practice Connections and Patient-Centered Medical Home (PPC-PCMH)standards, can be found in the eAppendix at 

Many of the elements required under the PPC-PCMH standards are provided in the Patient-Centered Primary Care Team Practice component. Access criteria are met through close  monitoring of performance on appointment standards (NCQA standard 1). Tracking and registry capabilities for several chronic diseases are embedded into the Primary Care Team Practice component’s EHR (NCQA standard 2). Reminders for preventive and chronic disease care are part of the quality improvement initiative described above (NCQA standards 3 and 8). Self-management support has been a central theme in the team-based care approach the practices use; disease and case management were added as part of the PHN model (NCQA standard 4). Electronic prescribing as well as test and referral tracking also are available in the EHR (NCQA standards 5, 6, and 7). Advanced communication capabilities for patients and providers are available through the electronic portals portion of the EHR system (NCQA standard 9).

Although the PHN model was created prior to the publication of NCQA’s PPC-PCMH standards, it does address all of the capabilities described in those standards. However, because our goal was to impact the quality, patient experience, and efficiency of care across the full continuum of care, not just in the office of the primary care physician (PCP), we believed that additional components and activities were necessary. The activities included in the 5 components are described further in Table 1. Several are worth noting. First, many of GHP’s population management activities were moved to the practice site. Geisinger Health Plan provided case managers for each practice at a ratio of 1 nurse for every 800  Medicare patients to serve as the hub for population-based activities. Second, the model explicitly calls for the PCPs to develop systems of care for their patients when they are seen by other physicians or in other settings. Third, additional financial support was provided by GHP to pay for new services in the PCP office. An example is dedicated phone lines to allow high-risk patients to contact their case managers directly. Fourth, performance reports documenting the quality, utilization, and overall cost-of-care results were provided to the  practice. Finally, we added a shared savings incentive model to the GHP reimbursement arrangement. Qualityoutcomes were aligned with preexisting preventive and chronic disease care quality initiatives. Shared savings incentive payments then were based on improvement in bundled metrics for these services and other agreed-upon metrics.

Implementation was focused on the GHP Medicare Advantage population because the high prevalence of chronic illnesses and the resource use of this population provide the best opportunity to demonstrate and evaluate the impact of the interventions. In October 2006 and January 2007, the PHN model was introduced into 2 pilot GHS practice sites selected because of their large GHP Medicare Advantage population and because their locations made them easily accessible for our PHN management team. During 2007 and January of 2008, the PHN model was expanded to include the Medicare Advantage members in 9 additional practices (

). Prior to implementation at each site, all practice staff were  trained on the core components of the model.

Initially, the PHN teams focused on improving the management of the highest risk patients. The GHP-embedded case managers were integrated as part of the practice care team. They were provided with utilization and predictive modeling reports derived from GHP claims data. For the first time, these reports gave the practice teams a systematic way to identify relative risk for their GHP patients. The case managers then met with the highest risk patients to design patient-specific care plans. They also provided close follow-up for patients transitioning from hospital to home.This activity focused on reaching out to the patient within 48 hours of discharge, medication reconciliation, appropriate resources and social supports in the home, and timely follow-up with the patient’s PCP. Monthly team meetings that included PCPs, office staff, case managers, and GHP staff were held to evaluate results, discuss practice workflow and care access, and review hospital admissions for missed opportunities.

The case managers also formed partnerships with preferred home health agencies and nursing homes. Outreach and education regarding the PHN strategy were provided to  these agencies. Pharmacy management initiatives were developed to improve generic utilization, assist members approaching the Medicare Part D coverage gap, and provide  members with acute care protocols for treating exacerbations of chronic conditions.

As progress was made, expanded strategies focused on members at moderate and low risk. Patients with gaps in preventive or chronic care were identified by EHR registries and  health plan claims tools. Health plan nurses with training in disease management targeted moderate-risk members with hypertension, coronary artery disease, and diabetes for  selfmanagement education; worked with providers to ensure appropriate screenings; and assisted in optimizing medications. Site-based practice staff reached out to low-risk  members to coordinate preventive care screenings such as mammograms, colorectal screening, and influenza vaccinations.

Because we hypothesized that opportunities to reduce total healthcare spending through this model would relate to the ability of the practice to prevent hospitalizations and readmissions, we constructed monthly series of these events for each patient. Readmissions were defined as all medical—surgical patients admitted to acute care within 30 days from time of discharge for primary admission. Total healthcare spending (plan payment plus copayment) was computed for each member for each month by summing the allowed amount on medical claims. Pharmacy claims were not included in total spending because of variability in prescription drug coverage among members and over time because of the introduction of Medicare Part D in 2006. To protect the confidentiality of GHP payment information, we indexed spending so that the mean for patients in the nonintervention practices in January 2005 was set to $100.

We analyzed data at the patient-month level using multivariate linear mixed regression to measure the effect of thePHN intervention on hospital admissions per 1000 members, readmissions per 1000 members, and total nonprescription per member per month medical spending. Because we could access data from each clinic before and after PHN implementation along with concurrent data from non-PHN practices, our chosen approach was to model the expected incremental effect of PHN status on outcomes during the postintervention period, after adjusting for all covariates. The results reported in this article, therefore, represent the effect of PHN within a clinic and not simply a comparison of PHN clinics with non-PHN clinics. Following the approach of Berlin et al,

 we decomposed PHN status into 2 components: the percentage of the study period during which each practice implemented PHN (ie, clinic-exposure association) and whether PHN was in effect during each observation (ie, a time-varying component). This 2-variable approach was necessary because, although our primary interest was the impact of a clinic switching from non-PHN to PHN status, there could be confounding due to the selection of PHN clinics and variability of different clinics’ duration of exposure to PHN. The first variable was time invariant and thus could represent the incremental effect of the intervention itself. It could, however, absorb bias from clinic-exposure association as well as biases between clinics selected or not selected for PHN. By contrast, the second variable (ie, a binary indicator of current PHN status centered around the time-invariant component at each site) isolated the actual effect of the PHN intervention on a clinic population, which was the objective of the study. Other covariates in the model were Hierarchical Condition Category (HCC) risk score,

the year, and interactions between both PHN variables and year (to capture differences in trends within and among the PHN and non-PHN groups) and between PHN variables and HCC score. Age and sex were not included as separate regressors because they were captured in the HCC score. We also included a series of variables for calendar months to capture seasonal variation in utilization and spending.

Regression models were estimated using general estimating equations with exchangeable covariance. Consistent with the underlying nature of the data, we used a log-link function and quasi-Poisson distribution for monthly admissions and readmissions. To analyze monthly spending, we used a standard 2-part model modified to account for a small amount of capitation GHP pays uniformly for all members. The second part of the spending model was specified with a log-link function and normal distribution. These models allowed us to calcunlate the difference between the observed outcomes for active PHN participants and their expected outcomes if the PHN had not been implemented. To adjust for secular trends in the postintervention time period, we utilized data from a group of non-GHS practices that cared for GHP Medicare Advantage enrollees as a control cohort. We considered using either a GHS or non-GHS control cohort, but chose the non-GHS practices for 2 reasons: (1) the larger sample of non-GHS practices available for matching and (2) concerns about spillover effects of PHN on non-PHN practices within GHS. To ensure similarity between the intervention and nonintervention groups, we used propensity score matching to identify a subset of practices that were most similar to the intervention practices in terms of patient population and outcomes in 2005. Because the PHN model was implemented at the practice level, we sought to match the intervention and nonintervention cohorts at this level. We estimated propensity score models predating the adoption of the PHN model by using the following practicelevel variables, all measured for 2005: mean patient age, percentage of male patients, mean HCC score, total per member spending, inpatient admissions, and readmissions. Each practice that adopted the PHN model was matched to 10 practices that did not adopt the PHN model based on the proximity of their estimated propensity score. Only those practices with estimated log odds from the propensity model that were within 0.60 standard deviation of an intervention site were counted as matches. A sensitivity analysis (not shown) with a GHS control cohort yielded results that were within the statistical confidence intervals (CIs) of the results presented here.

Because of overlap between sites, the propensity score model yielded a total of 75 distinct non-Geisinger practices within the common support region that could be matchedwith 1 or more of the 11 intervention sites (data not shown). Data from each of these sites were weighted to appropriately reflect the number of intervention sites with which they were matched. At baseline (2005), there were no statistically significant differences in sex or HCC scores of patients treated in the 11 PHN sites compared with the propensity score—matched non-GHS practices (

). Patients treated in the intervention sites were approximately 6 months younger on average than those treated in comparison sites (

 <.001), but because all results were regression adjusted, any potential bias associated with this age difference should have been eliminated. Average monthly admissions and readmissions per 1000 patients also were similar between the 2 groups (

 =.74, respectively). Average spending per member per month was approximately 4% higher in the intervention cohort than in the comparison cohort (

present the cumulative percent differences in actual admissions, readmissions, and total spending for PHN members versus the expected outcomes if PHN had not been implemented. As described previously, these expected outcomes were calculated by setting the time-dependent PHN indicator variable to zero in each multivariable regression model while keeping all other covariates constant. Outcomes were expressed as estimated effects with bootstrapped 95% CIs and as P values. The PHN model wa associated with a total cumulative reduction of 56 admissions per 1000 members per year (18%; 95% CI, −30% to −5%; 

 <.01). The PHN model also was associated with a

cumulative effect of 21 fewer readmissions per 1000 members per year (−36%; 95% CI, −55% to −3%; 

 = .02). Finally, the regression model estimated that the PHN model reduced cumulative total spending by 7%, but this difference did not reach significance (95% CI, −18% to 5%; 

 = .21). Results were qualitatively similar if all non-GHS clinic sites, rather than the propensity-matched comparison group, were used as a control cohort (data not shown).

Introduction of a medical home care delivery model was associated with a significant reduction in hospital admissions and readmissions for a population of Medicare Advantage enrollees. Our findings present a contrast to the recently published results of the Medicare Health Support demonstration, a set of parallel, randomized controlled trials of traditional disease management delivered by third parties to disease-specific populations.

 Despite targeting sicker individuals, participating programs had little effect on healthcare utilization or spending. We hypothesize that the comparative success of the PHN model was partly due to its ability to leverage existing physician—patient and interprovider relationships to fundamentally change the way care is delivered rather than work outside the system to improve care.

Because the PHN model is a complex intervention, it is difficult to ascertain which elements are responsible for specific improvements in care. Our findings, coupled with qualitative observations, however, highlight the importance of placing nurse case managers directly into the practices and arming them with data and analytical capabilities. With timely information on emergency department (ED) and inpatient use, case managers are able to manage transitions of care to ensure early postdischarge follow-up and medication reconciliation. In addition, proactive identification of at-risk individuals provides an opportunity to use patient-specific action plans to implement timely interventions for acute exacerbations of chronic illnesses.

Our lack of findings on per member per month spending likely was due at least in part to the sample size and duration of the study, coupled with the skewed distribution of the healthcare spending data. We note that based on its own actuarial analysis, GHP found that the PHN practices did generate savings and triggered incentive payments under the quality-based shared savings incentive system.

Based on our understanding of the initial focus of PHN practices and the policy importance of understanding the impact on resource use, we focused the analyses on hospital utilization and total costs. In future work, it will be important to examine a broader scope of process and outcome measures, including ED visits, clinical quality measures, physician and nurse professional satisfaction, and patient experience. Tracking this larger set of outcomes will not only allow us to better answer questions of sustainability but also permit refinement of the model.

These conclusions are tempered by several study limitations. First, our measure of medical spending excluded the cost of prescription drugs. There are concerns that improved care coordination may increase the cost of prescription drugs, thereby decreasing or eliminating medical services savings.

 However, a separate analysis of the changes in drug expense over time for both groups of practices demonstrated no differential impact or erosion of savings in the PHN sites. Second, the PHN model is situated in an integrated payer—provider environment (ie, the payer and provider are part of the same corporate entity) with long-standing use of an ambulatory EHR, in a Medicare population with high baseline spending and relatively little patient turnover. These factors almost surely contributed to PHN’s success and may therefore limit generalizability to other settings. The PHN model, however, has subsequently been introduced into non-GHS practices. Moreover, implementation experience to date suggests that the key components of the PHN model are on-site case management, the use of population data, and the shared savings incentives, all of which could be implemented outside of an integrated delivery model. Third, as described above, while the PHN model is aligned with the NCQA PPC-PCMH standards, it also includes other components. In particular, PHNimplementation likely differs from many other medical home efforts that may not include robust case management programs, attention to care delivered outside of the PCP office, shared savings reimbursement, or direct health plan support. Although advocates of the medical home will see these findings as general support for that model, attention should be paid to the specific elements of the intervention and the limits to generalizability noted above. Finally, although every effort was made to account for secular trends and confounding factors, our study was observational in nature and the usual caveats with regard to causal inference apply.

The strengths of the study design also are significant and include the fact that we were able to analyze healthcare spending and utilization for an entire population before and after the intervention alongside a well-matched comparison group. This approach allowed for relatively robust causal inference while minimizing the potential confounding effects of regression to the mean and survivorship biases.

A patient-centered medical home with intensive case management and a payer partner can significantly improve hospital utilization and may decrease total medical costs for a Medicare population.


Frederick J. Bloom Jr, MD, MMM

Articles

Unintended Consequences of a Quality Measure for Acute Bronchitis

Reducing Long-Term Cost by Transforming Primary Care: Evidence From Geisinger's Medical Home Model

Value and the Medical Home: Effects of Transformed Primary Care