Objective Investigation of variations in provider performance and its determinants may help inform strategies for improving patient outcomes.
Methods We used the National Heart Failure Audit comprising 68 772 patients with heart failure with reduced left ventricular ejection fraction (HFREF), admitted to 185 hospitals in England and Wales (2007–2013). We investigated hospital adherence to three recommended key performance measures (KPMs) for inhospital care (ACE inhibitors (ACE-Is) or angiotensin receptor blockers (ARBs) on discharge, β-blockers on discharge and referral to specialist follow-up) individually and as a composite performance score. Hierarchical regression models were used to investigate hospital-level variation.
Results Hospital-level variation in adherence to composite KPM ranged from 50% to 97% (median 79%), but after adjustments for patient characteristics and year of admission, only 8% (95% CI 7% to 10%) of this variation was attributable to variations in hospital features. Similarly, hospital prescription rates for ACE-I/ARB and β-blocker showed low adjusted hospital-attributable variations (7% CI 6% to 9% and 6% CI 5% to 8%, for ACE-I/ARB and β-blocker, respectively). Referral to specialist follow-up, however, showed larger variations (median 81%; range; 20%, 100%) with 26% of this being attributable to hospital-level differences (CI 22% to 31%).
Conclusion Only a small proportion of hospital variation in medication prescription after discharge was attributable to hospital-level features. This suggests that differences in hospital practices are not a major determinant of observed variations in prescription of investigated medications and outcomes. Future healthcare delivery efforts should consider evaluation and improvement of more ambitious KPMs.
Statistics from Altmetric.com
Over the past few years, substantial effort has been made to ensure that patients with heart failure and reduced left ventricular ejection fraction (HFREF) receive guideline-recommended care that is known to improve outcomes. In many healthcare systems, professional societies and government agencies have formulated and endorsed quality standards for better implementation of recommended care.1–3 Despite these efforts, the quality of care and outcomes for individuals suffering from heart failure are unsatisfactory in many regions of the world.4 ,5
Investigation of variations in adherence to recommended care among providers can inform decision makers about the nature and drivers of deficiencies in quality of care and whether these might be amenable to organisational changes.6 Given that such changes often require substantial financial and human resource investment, it is prudent to explore the extent to which differences in care practices are responsible for variation in quality of care and to estimate the impact of any proposed changes to service delivery.7 ,8
We set out to measure the amount of variations among hospitals in their adherence to key performance measures (KPMs) for management of patients with heart failure having HFREF and the extent to which this can be explained by hospital-level factors.
This study is a part of the Understanding National Variation and Effects of Interventions at different Levels of Care for Heart Failure study, which aims to characterise variation in hospital care and outcomes for patients with heart failure.
We used the National Heart Failure Audit for our primary analyses. The audit enrols patients hospitalised with a primary diagnosis of heart failure in England and Wales.8 Initially, in 2007, participating hospitals were asked to provide data on at least the first 10 patients with a primary death or discharge diagnosis of heart failure in each month; this requirement has steadily increased and, from 2012, all hospitals in England and Wales were expected to report all unscheduled admissions due to heart failure to the audit. The dataset captures information about patient demographics, clinical characteristics and follow-up information. The audit was supplemented by a survey of 185 English and Welsh hospitals, included in the National Heart Failure Audit, that provide care for patients with acute heart failure, capturing information on hospital characteristics, including human resources (eg, number of cardiologists), referral pathways (eg, heart transplantation) and other organisational features.
Study population and outcomes
Only hospital admissions in which the patient survived to discharge were eligible for inclusion in the study because collection of treatment variables was not mandated for patients who died in hospital. We restricted our analysis to patients with HFREF, diagnosed using echocardiogram, because clearly defined and evidence-based treatment recommendations exist only for this subgroup. Contraindications to ACE inhibitors (ACE-Is)/angiotensin receptor blockers (ARBs) and β-blockers were recorded and patients with any contraindications were classified as missing for these variables. For patients with more than one reported hospital admission (10 280, 14.4%), we randomly selected one admission.
The primary analyses of variation among hospitals in their adherence to KPMs used patient-level outcomes for three KPMs for HFREF: provision of an ACE-I/ARB, provision of a β-blocker and referral for follow-up with a heart failure specialist (either referral for follow-up with a cardiologist or with a heart failure specialist nurse) and their composite performance score (described below). We chose these KPMs as they were recommended by the American Heart Association's Task Force on Performance Measures in 2012 and the UK National Institute for Health and Care Excellence in 2014.3 ,9
To generate the composite performance score, a mean of the provision of an ACE-I/ARB, provision of a β-blocker and specialist follow-up was generated for each patient. This was averaged at the hospital level across all patients to generate a hospital-level composite performance score (continuous score, ranging from 0 to 1, or 0% to 100%).10 We chose to use process outcomes as our primary end points, rather than mortality, because the process outcomes selected represented established clinical standards and such outcomes are better suited than mortality or metrics with limited evidence base for comparing performances between hospitals.11
In a secondary analyses, we investigated variation among hospitals for risk-adjusted death at 30 days and at 1 year after discharge (adjusted for age, sex, New York Heart Association (NYHA) class (I, II, III or IV), peripheral oedema (none, mild, moderate or severe), history of diabetes, history of ischaemic heart disease, history of hypertension, history of valve disease, atrial fibrillation, left bundle block, previous myocardial infarction, concomitant diastolic dysfunction, left ventricular hypertrophy and valve disease).
We used multiple imputation using chained equations to impute missing covariates; five imputation sets were generated. No covariate was missing at a rate that exceeded 15%. Simulation studies have suggested that multiple imputation provides equivalent or better coverage and bias than complete case analysis, even when missingness is not at random.12 ,13 We therefore used multiple imputation rather than a complete case analysis.
Continuous data were summarised using the mean and SD or median and interquartile interval while categorical data were summarised using percentages. Hierarchical Poisson and normal regression models were used to examine time trends in the individual components of the composite score and the patient-level composite score, respectively, with yearly rates modelled through the inclusion of calendar year as a categorical covariate. As the patient-level composite score was not symmetric (2.3% of participants received a score of 0, while 58.1% received a score of 1), normal approximation condition for the patient-level composite score was verified by bootstrapping. We conducted bootstrapping using 10 000 replications and examined the distribution of the mean intraclass correlation coefficient (ICC—the primary outcome, described below). The mean ICC followed a normal distribution and the bootstrapped CIs were identical to those derived using a normal approximation.
Time trends in risk-adjusted 30-day mortality and 1-year mortality were derived through hierarchical Poisson models adjusted for patient case-mix (clinical characteristics listed below). Time trends of 30-day and 1-year risk-adjusted mortality for each year at the means of patient covariates were calculated. Statistical significance of changes in rates was then determined by including year as a continuous covariate.
To examine associations between hospital characteristics and their composite performance score, we divided hospitals into fifths based on their respective composite performance score (with the bottom fifth containing hospitals with the lowest composite score and the top fifth containing hospitals with the highest composite score). Hospital characteristics by fifth of composite performance score were summarised using proportions for categorical data and means for continuous variables. To examine whether hospital characteristics differed across fifths, continuous and categorical covariates at the hospital level were tested for trend using linear regression with hospital fifths as a continuous outcome. To further determine which hospital characteristics were associated with a better performance score, backward stepwise regression was performed with the inclusion of all hospital characteristics and a p value for exit of 0.01.
In our main analysis, to investigate the degree to which patient case-mix, year of investigation and hospital features accounted for the variation in hospital performances, we used hierarchical logistic regression models. We first fitted an unconditional (null) model with only a hospital random intercept to quantify the amount of between-hospital variations. In the second model, we added 24 patient characteristics: demographic characteristics (age, sex); clinical characteristics (NYHA class (I, II, III or IV), peripheral oedema (none, mild, moderate or severe), history of diabetes, history of ischaemic heart disease, history of hypertension, history of valve disease, atrial fibrillation, left bundle block, previous myocardial infarction, diastolic dysfunction, left ventricular hypertrophy and valve disease) and dummy variables for year of admission (2007, 2008, 2009, 2010, 2011, 2012, 2013) to investigate the extent to which differences in performance between hospitals was accounted for by patient case-mix and year of admission. In the third model, we added hospital characteristics independently associated with the composite performance score to determine whether the variation among hospitals could be accounted for by known hospital characteristics: tertiary hospital, number of full-time equivalent consultant cardiologists, a multidisciplinary team with a cardiologist, a multidisciplinary team with a consultant with an interest in heart failure, a multidisciplinary team with a pharmacist, a multidisciplinary team with a psychologist, a multidisciplinary team with a social workers, coronary angiography capabilities, cardiac MRI capabilities, an outpatient service with a cardiologist and an outpatient service with a heart failure specialist nurse. We used the ICC to estimate the proportion of variance in performance that was attributable to the between-hospital variations. We calculated the ICC from these hierarchical logistic models using the following equation: ICC=se2/(se2+(π2)/3), where se is the standard error of the random hospital intercept.14 This was supported by the median OR (MOR), which reports between-hospital variation that is not explained by patient characteristics. The MOR works on the principle that, if two patients with ‘identical patient-level characteristics’, from two randomly selected hospitals are compared, any OR >1 would represent differences in the hospital (as the patients are identical). The MOR is calculated using the formula: MOR=exp(0.95×se), where se is the standard error of the random hospital intercept.14 We also calculated ICC for KPMs by year and tested for trend across years using inverse variance weighted linear regression, to examine whether variation in outcomes has changed across years.
All analyses were performed using Stata IC (V.12). Study findings are reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology recommendations.15 No ethics approval was required for this analysis; the National Heart Failure Audit was conducted with the approval of the NHS Information Centre.
In total, 68 772 patients with HFREF were discharged with a primary diagnosis of heart failure from 185 hospitals, between 2007 and 2013. The median (interquartile interval) age was 75 years (68, 84) and 36.5% were women.
From 2007 to 2013, no significant improvement in rates of prescription of ACE-I or referral for specialist follow-up was observed (table 1). However, rates of prescription of β-blockers increased from 56% to 82% (p trend<0.001), which drove an increase in the composite score from 73% to 84% (p trend <0.001). Risk-adjusted 1 year mortality decreased from 27% in 2007 to 23% in 2011 (p trend <0.001), although no appreciable change was observed in risk-adjusted 30-day mortality (table 1).
The composite performance score at the hospital level showed modest variation (median 79%, range 50%–97%; figure 1A and online supplementary figure S1). Among the three individual components of the composite performance score, prescription of ACE-I/ARB at discharge had the highest adherence and lowest variation in its prescription among hospitals (median 84%, range 62%–99%; figure 1B). The prescription of β-blockers at discharge was lower but with similar variability among hospitals (median 75%, range 45%–96%; figure 1C); while follow-up with a specialist had the largest variation (median 81%; range; 20%–100%; figure 1D).
In stratified analyses of the composite performance score, hospitals in the top fifth of composite performance score (Q5), compared with those in the bottom fifth (Q1), had a higher proportion of tertiary hospitals and consultant cardiologists (table 2). The presence and composition of the multidisciplinary heart failure teams also differed between hospital fifth scores, with hospitals in the top fifth being more likely to have a multidisciplinary heart failure team. There were no statistically significant differences across fifths of hospitals with regard to referral pathways, including referral for heart transplantation or for cardiac rehabilitation.
With the composite performance score as a continuous outcome, we performed backward stepwise regression and included all hospital characteristics. We found two statistically significant predictors of higher adherence to KPMs and retained these in the final model: the number of consultant cardiologists working in the hospital and the presence of a multidisciplinary heart failure team with a pharmacist (table 3). However, their absolute effect on the composite performance score was small; on average, the presence of a multidisciplinary team with a pharmacist was associated with an increase of 3.9% (CI 1.1% to 6.8%; p=0.003) in performance score. The addition of 10 more full-time equivalent cardiologists was associated with an increase of 3.5% (CI 1.2% to 5.8%; p=0.007) in performance score.
In an unadjusted (null) linear multilevel model, the ICC for the composite performance score was 9% (CI 7% to 11%), suggesting that only 9% of the total national variation in adherence to the composite performance score was related to differences among hospitals (table 4). Thus, variability within hospitals (related to patient factors, random variation or other known or unknown factors within hospitals) greatly exceeded the variability among hospitals (due to known or unknown hospital features). When year of admission and patient features were added to the model, the ICC decreased from 9% to 8%, indicating that variation in the type of patient managed by different hospitals and year of admission explained only 1% of the total variability among hospitals (table 4). Addition of known organisational features reduced this variability by another 1%, suggesting that these features had a small, absolute and relative influence on the variability in the composite performance score (proportion of residual variability explained by known organisational features was 12.7%, table 4). No evidence of change was observed in the ICC for any of the five performance measures across years (see online supplementary table S1). Estimates were similar when a propensity score was adjusted for rather than adjusting for covariates (see online supplementary table S2).
The variability in prescribing ACE-I/ARB and β-blockers was concordant with the overall composite performance score. Of the total variability in prescription of ACE-I/ARB and β-blockers, only 8% and 7% were attributable to differences among hospitals (ICC 0.08, CI 0.06 to 0.10 and 0.07, CI 0.06 to 0.09, respectively). However, hospital-level variation in the rate of referral for specialist follow-up was 26%, even after adjustment for patient characteristics (ICC 0.26, CI 0.22 to 0.31). The MOR for referral for specialist follow-up was 2.94 (CI 2.63 to 3.33), suggesting that, on average, the odds of a patient being referred for specialist follow-up after discharge would differ approximately threefold from one randomly selected hospital to another hospital with higher odds (table 4).
In our secondary analyses of mortality following discharge, we found that at the hospital level, mortality at 30-days ranged from 2.1% to 14.3% and for 1 year from 10.4% to 43.6%. In an unadjusted multilevel model, the ICC for death at 30 days and 1 year after discharge were 2% (CI 1%, 3%) and 1% (CI 1% to 2%), respectively, suggesting that only 1% to 3% of the total national variation in death rates were related to variations in known or unknown hospital-level features (table 4).
This analysis of the National Heart Failure Audit shows substantial variation in hospital adherence to a composite of KPMs for management of patients hospitalised with HFREF. However, only a small fraction of this variation was attributable to between-hospital differences in care provision. This overall low hospital-attributable variation was mainly driven by high rates of ACE-I/ARB and β-blocker prescription with small degrees of hospital-attributable variations (7% and 6% of the total variability, respectively). However, variation in referral for specialist follow-up was substantial and 26% of it was due to hospital-level features.
Previous studies have reported wide variations in management of patients with heart failure across different healthcare systems4 ,16 ,17 and others have shown certain hospital-level features to be associated with better clinical outcomes.18 The present study goes beyond these earlier findings. By quantifying the extent to which variation in KPMs can be attributed to hospital-level features, this study raises important questions about the potential impact of further organisational changes that target individual hospitals. Broadly consistent with previous suggestions, we found several inhospital organisational features to be significantly associated with hospital-level performances. However, we further show that the absolute effect of these organisational features on explaining variability in adherence to KPMs, particularly ACE-I/ARB and β-blocker prescription, is small. Although the numbers of consultant cardiologists working in the hospital and the presence of a multidisciplinary heart failure team with a pharmacist were strongly associated with the hospital performances, they collectively accounted for only 1% of the total hospital variation.
These findings indicate that the majority of remaining variability in prescription of ACE-I/ARB and β-blocker is randomly distributed among hospital providers and is not determined by differences in organisational features among hospitals. While further investment into costly organisational changes for management of HFREF in hospitals in England and Wales may still be useful for changing other important healthcare outcomes across all hospitals, our study shows that such investments cannot be expected to lead to large reductions in variability in hospital adherence to these performance measures. Even if we assume that the observed associations between hospital preferences (eg, presence of multidisciplinary teams) and hospital performances (eg, prescription rates for β-blocker) are causally related, additional changes to such hospital preferences for inhospital care would only be expected to reduce the absolute between-hospital variability in the ACE-I/ARB or β-blocker prescription by <5%. The question whether such modest reductions in the process outcomes are worthwhile requires formal health economic evaluation before any recommendations can be made about additional changes to service delivery in the UK hospitals.19 These results also suggest that research should be undertaken to characterise hospital-level variation in KPMs for heart failure in other countries. The modest effects of quality improvement programmes on heart failure care in the USA20 may be due to limited variation in KPMs, as observed in this analysis.
The only key performance measure for which we observed large variation among hospitals was referral to specialist follow-up (either to a heart failure nurse or cardiologist). After case-mix adjustment, there was still an almost three times difference in the odds of a patient being referred for specialist follow-up between two randomly selected hospitals. About 20% of this variation was explained by known differences in hospital practices and overall the average adherence to specialist follow-up after discharge did not change from 2007 to 2013. If specialist follow-up influences outcomes in heart failure, as has been suggested by prior studies, such high levels of variation may partially explain differences in risk of death among hospitals. There are several potential explanations for this observation. First, it is likely that adherence to this key performance measure is more difficult and resource-intensive than making changes to prescription of evidence-based therapies. Second, it may be that there is less professional agreement about this performance measure because the level of evidence for it is less strong than pharmacological interventions.21 Randomised evidence on the effect of early specialist follow-up after discharge is limited22 ,23 and findings from non-randomised comparisons have been inconsistent.24–26
This study has several strengths. First, recommended performance measures were selected as the primary outcome,3 ,9 as opposed to clinical end points, thus avoiding the risk of uncontrolled residual confounding when mortality or hospitalisation are chosen.11 ,27 Second, we separated the evaluation of inhospital performance from post-discharge performance to take account of differences in recommended care during these very different phases of care. Third, we linked a survey of organisational features to hospital processes and used multilevel analysis to quantify the impact of such features on quality of care, thus extending previous studies which have been criticised for largely investigating associations between structural hospital features, such as number of beds or number of patients, which are not amenable to change.28 Finally, we included a large number of confirmed cases of HFREF with linked databases and hence were able to investigate the quality of care in a more accurate and detailed manner than previous reports.
However, this study also has several potential limitations. First, many other aspects of heart failure care, in addition to prescription of ACE-I/ARBs, β-blockers and specialist follow-up, may influence outcomes after hospitalisation for HFREF. However, other evidence-based interventions, such as mineralocorticoid receptor antagonists or cardiac resynchronisation therapy are currently not recommended as the minimum performance measures for inhospital care. These results therefore suggest that future research should examine whether these evidence-based interventions would be better-suited as performance measures, considering the relatively high prescription rates and modest hospital-level variation of ACE-I/ARBs and β-blockers observed. Second, we lacked information on the dosage of prescribed medications, which is an important determinant of outcomes in heart failure and may exhibit greater variation among hospitals than prescription rates.29 Third, the small between-hospital variation observed in this study may not be generalisable to health systems which are more diverse than the NHS in terms of organisation and delivery of care for patients with acute heart failure.4 Fourth, although hospitals were asked to provide information from the first unselected 10 or 20 patients admitted to their hospital each month, we cannot entirely rule out that patients included differ in some respects from those who have not been included in the reports. Finally, we focused our analysis on quantification of differences in natural variation among hospitals in order to estimate the impact of variation in hospital preferences on KPMs. In theory, it is, however, possible that interventions that equally target all hospitals could lead to an average increase (or decrease) in hospital performances across all hospitals even when there is not much between-hospital variability but average performance is uniformly low (or high).
In conclusion, we observed an improvement in hospital performances for management of patients with HFREF over time. Although substantial variation in adherence to KPMs was observed, only a small proportion of this could be attributed to between-hospital differences. These results suggest that further organisational changes that would specifically target low-performing hospitals will have limited impact on reducing variation in prescription rates of ACE-I/ARBs and β-blockers for patients with HFREF in the UK. Future hospital-level organisational changes should consider focusing on improving the rates of referral to specialist follow-up after discharge, for which there is substantial variation. Future research should also examine whether other evidence-based interventions, such as prescription of mineralocorticoid receptor antagonists or cardiac resynchronisation therapy or prescription of recommended dosages of medications, should be used as performance measures for inhospital care.
What is already known on this subject?
Quality of care and outcomes for individuals suffering from heart failure are unsatisfactory in many regions of the world. Previous studies have reported wide variations in management of patients with heart failure across different healthcare systems.
What might this study add?
Only 8% of the total variation in the composite performance measure was due to differences in known and unknown hospital features. However, variation in referral for specialist follow-up was substantial and 26% of it was due to hospital-level features.
How might this impact on clinical practice?
While further investment into costly organisational changes for management of heart failure with reduced left ventricular ejection fraction (HFREF) in hospitals in England and Wales may still be useful for changing other important healthcare outcomes across hospitals, our study shows that such investments cannot be expected to lead to large reductions in variability in hospital adherence to heart failure performance measures examined in this study. Future healthcare delivery efforts should consider evaluation and improvement of more ambitious key performance measures.
Contributors CAE and KR were involved in the design, implementation and analysis of the study and in the writing of the final manuscript. All authors were involved in revision of the manuscript for important intellectual content.
Funding KR is supported by the National Institute for Health Research (NIHR) Oxford Biomedical Research Centre and NIHR Career Development Fellowship. CAE is supported by the Rhodes Trust. MW is supported by a Principal Research Fellowship from the Australian Health and Medical Research Council and is a consultant for Amgen and Novartis. The work of the George Institute is supported by the Oxford Martin School. SGA is an Academic Clinical Lecturer in Cardiology and is funded by the NIHR. The study was funded by the UK NIHR.
Competing interests MW declares consultancy fees from Amgen and Novartis.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement Data and code are available from the lead author on request.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.