Factors associated with hospital emergency readmission and mortality rates in patients with heart failure or chronic obstructive pulmonary disease: a national observational study

Modelling framework

The terms risk prediction and risk adjustment are closely related despite their differing aims, but a model for predicting mortality, for example, might not include the same set of variables as a risk adjustment model used to compare hospitals’ mortality rates. Risk prediction values parsimony and interpretability, whereas risk adjustment can focus more on confounder control. Risk prediction models could encompass factors such as staffing and bed numbers, or other factors that are (at least partly) under the hospital’s control, whereas this would be wrong for risk adjustment models for comparing providers. When exploring the associations between patient, GP and hospital factors and each outcome, we are in a risk prediction framework; when producing hospital-level measures for performance monitoring, we are in a risk adjustment one.

Prediction of binary outcomes, such as mortality or readmission, is usually done via logistic regression. With multiple health service units, such as GP practices, surgeons or hospitals, the modeller, in principle, needs to account for the ‘clustering’ of patients within these units. The common way of accounting for this is to use multilevel models, particularly with random intercepts for practices and hospitals, and fixed effects for covariates. For this, we used Statistical Analysis Software’s (SAS’s) procedure for binary outcomes, PROC GLIMMIX. From this we derived relative risks (RRs) for each hospital using predicted probabilities from only the fixed effects part of the model. 20,21 We used the noblup ilink options within PROC GLIMMIX to achieve this. These RRs are akin to standardised mortality ratios (SMRs), which represent the ratio of the hospital’s rate to the national average rate. However, as explained below, the hospital-level clustering was found to be minimal, and accounting for clustering by both general practice and hospital was found to be unfeasible, so we also calculated hospital-level RRs using PROC LOGISTIC, which does not account for any clustering.

The Centers for Medicare and Medicaid Services (CMS) in the USA uses, for its publicly reported outcome measures, empirical Bayes ‘shrunken’ estimates of the SMRs22 Confidence intervals (CIs) for the CMS measures are constructed using a complicated bootstrap procedure. Several studies have compared the results from fixed- and random-effects models with regard to provider profiling, in which a key aim is the identification of statistical outliers, especially units with higher than expected mortality. In general, these conclude, as Austin et al. 23 did, that ‘when the distribution of hospital-specific log-odds of death was normal, random-effects models had greater specificity and positive predictive value than fixed-effects models. However, fixed-effects models had greater sensitivity than random-effects models.’ For a full discussion of the evidence around the effect of different adjustments for clustering, including Bayesian methods for provider profiling reviewed by Austin,24 see Chapter 7 of our book, Statistical Methods for Healthcare Performance Monitoring. 25

Multilevel models allow for the estimation of the amount of variation between units at each level, for example, between practices or hospitals. The residual intraclass correlation coefficient (ICC), a measure of clustering used in hierarchical modelling, expresses the proportion of variability explained by the presence of clusters at, for instance, a hospital level. 26 It is computed for logit models as:

where τ_H is the hospital-level variance and π = 3.14159.

By building up the levels in a hierarchical model, one can assess, for example, how much of the variation in outcomes at hospital level is attributable to the variation between practices or as a result of differences in the distribution of patient factors.

Model fitting

The modelling in this project served two different purposes as explained above, and we therefore took different approaches accordingly. For logistic regression, all candidate covariates were initially retained: we did not use any stepwise methods because of their well-known drawbacks but instead removed non-significant variables (backwards elimination) after checking the impact on the coefficients for the retained covariates. For the most part we did not test for interactions.

Continuous variables were sometimes categorised, depending on their relation with the outcome. Age and LOS were categorised in line with our previous approaches.

In order to explore the relationship between hospital and general practice variables and outcomes, variables were divided into deciles and the percentages of patients who were readmitted/died for each decile were plotted against the mid-value for the explanatory variable deciles. Plots were inspected and, where there was evidence of a clear linear or non-linear pattern, a suitable categorical approach was identified, based on quartiles. Models in which continuous variables were categorised were compared with the main models in which they were treated as continuous. There was no impact on which explanatory variables were retained or on the interpretation of results.

Assessment of model performance

With any risk model comes the need to assess its performance. For binary outcomes, two standard measures for logistic regression are the area under the receiver operating characteristic (ROC) curve, also known as the c-statistic, and the Hosmer–Lemeshow (HL) test output. The former measures discrimination, the ability of the model to predict a higher probability of death for those who died than for those who survived. It is generally considered that values of c above 0.70 represent good and values above 0.80 represent excellent discrimination. The maximum value obtainable is often quoted as 1 but, in fact, varies with the distribution of risk in the population (see Cook27 for a full discussion on this statistic). The HL test describes the model’s calibration and divides the data set into risk deciles. The observed and predicted number of events are compared in each decile, which often shows poor calibration at the extremes, and summarised in a chi-squared statistic. It has been criticised for having high type I and II error rates. 28 Although a simple plot of observed versus predicted rates may be more useful, we will nonetheless report HL test results to be concise. For the HL chi-squared values, we give right-sided p-values. We used 10 bins and 8 degrees of freedom for the HL test, as is standard.

The effect of the semi-competing risk of death on readmission-type measures

As death precludes subsequent readmission, using logistic regression – which ignores any effect of death either during or following the index discharge – may be potentially misleading. In our previous project11,14 we therefore also applied cause-specific proportional hazards modelling and subdistribution proportional hazards modelling. These two survival analysis methods make different assumptions regarding post-discharge deaths. 29,30 Other methods exist, but these are the two most widely used. The PSHREG macro in SAS was run for subdistribution hazards. 31 Our prior work found high agreement between the odds ratios (ORs) and both sets of hazard ratios, so we can be fairly confident that the effect of post-discharge deaths is minor. For this reason, we did not consider the impact of post-discharge death on readmission within 30 days. As is standard, our analyses of readmissions were restricted to patients discharged alive from their index admission. As well as logistic regression we ran standard Cox models.

Methods specific to objective 1

After obtaining the non-hospital data, our first task was to decide what information to take from the patient and staff surveys. Although the inpatient survey is divided into domains, we used the approach of Bos et al. 32 and considered the questions in terms of the pathway a patient takes through hospital. Our two patient representatives highlighted the importance of ‘reassurance when arriving and leaving hospital’ regarding their confidence to manage their disease. We hypothesised that the experience of arrival at hospital as an emergency and the discharge experience, including information about medication and side effects, might affect a patient’s decision to return to the ED to seek help.

We included two indicators from the patient survey: (1) patients’ experience of arrival at the hospital – a single question that asks patients about the waiting time for a bed after arrival at the hospital; and (2) patients’ experience of discharge based on four questions covering discharge delay, information about danger signals and the purpose and side effects of medication.

From the 2010 staff survey, we selected three questions a priori that, we hypothesised, may reflect organisational culture and the quality of care patients received: (1) staff rating of effective teamwork based on five questions; (2) staff agreement with the statement: ‘if a friend or relative needed treatment, I would be happy with the standard of care provided by this trust’; and (3) the percentage of staff who are satisfied with the quality of care they give to patients. The responses to each question were aggregated at trust level, and a mean response ranging from one to five or a mean percentage was attributed to each trust.

Questions were selected that ascertain patients’ recall of their experiences of access to primary care, including their knowledge of how to access OOH services. In addition to access, their perceptions of overall care, including whether or not clinical staff took their problems seriously, whether or not they had been told that they had a care plan and whether or not they would recommend the practice, were included.

Descriptive analyses included summarising the outcome rates for each patient-level predictor by chi-squared tests and scatterplots. We investigated the need for hierarchical modelling when assessing the relations between variables at the patient, primary care and hospital level by inspecting the covariance parameter estimates for the random effects. As noted earlier, the data are clearly nested: patients are nested within hospital trusts and separately within general practices. General practices are not generally wholly nested within trusts, that is, their patients do not attend only one trust, although there are exceptions. This is known as a cross-classified design. The large number of general practices and trusts makes it difficult to summarise the level of cross-classification. In addition, it also means that the covariance matrix underlying the multilevel model needed to run a cross-classified model is of size n, in this case over 78,000, but does not have a block diagonal structure that allows efficient algorithms, which is the case in a simple hierarchical model. This resulted in computational challenges when attempting to run a cross-classified analysis and we needed to take an alternative approach. In order to determine the impact of clustering, models were considered that took into account clustering of patients within trusts and, separately, of patients within general practices by including a random effect for trusts and general practices, respectively. The covariance parameters were then compared. This process was repeated for empty models, that is, those with no explanatory models, and models that included all patient-level variables. The covariance parameters were less than 0.05 in all models. The ICC was less than 0.01 in all cases, providing evidence that the amount of variance explained by the trust or practice level was < 1%. As the evidence of clustering affecting model results and interpretation was very weak, the model selected was a simple logistic model, which does not take into account the hierarchical nature of the data.

For readmission as the dependent variable, we first combined all causes, as is generally done, before splitting into two (HF/COPD vs. any other primary diagnosis) and ran separate analyses for each. For mortality as the outcome variable, we ran time-to-event models as outlined above, testing for time-varying effects by including interaction terms with the log of time.

We used logistic regression models to predict having an ED attendance and then to predict admission at that attendance within (1) 7 days and (2) 30 days of index discharge. These models included the inpatient and outpatient survey results and the other predictors for objective 1. We added time (shift) and day of the week of attendance. An interaction term between shift and weekday was tried. As a check to see if any important information was lost by considering these fixed windows of 7 and 30 days, survival analysis was also run, checking for non-proportionality of hazards.

To further explore the ED attendance patterns, we calculated overall outcome rates by time and weekday of attendance, subsequently stratifying by model-predicted risk of death. This showed whether, for example, the sickest patients attended more, attended at different times or were more likely to be admitted than less sick patients. The time of attendance was categorised into three shifts: 00.00–08.00, 08.00–18.00 and 18.00–00.00 hours.

Methods specific to objective 2

The variance of the random effects from the hierarchical models and the quasi-likelihood model dispersion parameter was calculated for the outcomes of 7-day attendance without admission, 30-day attendance without admission, 90-day attendance without admission, 30-day readmission, 7-day attendance without admission or 30-day readmission and 1-year mortality.

The relative contribution of patient and non-patient factors to variation in each outcome was assessed using the omega statistic, which is a ratio. With patient variables in the numerator and community and hospital variables in the denominator, ω = 0 would mean that all the variation in the candidate indicator is predicted by factors other than patient characteristics. Low values of ω are therefore desirable. This statistic cannot be used to judge an individual indicator (administrative data will lack some important patient factors), but it is useful for comparing them. We followed the example of Brown et al. ,33 who compared measures for ICU performance.

Methods specific to objective 3

We calculated RRs for these different measures for each hospital, adjusted for patient factors in logistic regression models. These RRs are akin to SMRs in epidemiology and are the ratio of the observed to the expected deaths or readmissions for each hospital, and so are the hospital’s outcome rate relative to the national average. We compared the sets of RRs using linear and non-linear correlation. As funnel plots are increasingly used to identify providers with unusually ‘good’ or ‘poor’ performance,34,35 we noted that the number and proportion of hospitals with outcome rates beyond 95% and 99.8% control limits. This was done for each outcome and patient group. We noted whether or not the same outliers were consistently identified across the sets.

To illustrate the statistical power to detect performance differences between hospitals, power calculations were carried out to determine the power to detect a change equivalent to 1.5 times the national rate in a small (25th percentile) hospital trust (n = 320 patients). This means that one would have greater power than this to detect larger differences than 1.5 for these hospitals, or to detect the same difference at larger hospitals. Our calculations are clearly not exhaustive but are fairly conservative. Power calculations were carried out using an online calculator provided by the Statistics Department of the University of British Columbia [www.stat.ubc.ca/∼rollin/stats/ssize/b1.html (accessed 23 October 2017)].

There were no methods specific to objective 4: results for HF and COPD were compared in a descriptive manner and are highlighted in each results subsection. Hospital-level RRs for HF and for COPD were compared using simple correlation.

Table 2 summarises the main analyses in this project by objective number.

Summary of public and patient involvement in this project

We asked our two patient representatives to go through the NHS Patient Survey questions to pick out those that they thought were particularly relevant either to their own experience of using the NHS or to other patients. They both responded with interesting viewpoints and ideas. They both identified readiness for discharge as problematic, and we specifically tested for the importance of questions related to this in the patient survey. Both representatives were asked to comment on our draft paper on PE score trends, and one of them did. We incorporated her remarks in the manuscript. Both representatives were asked to comment on the lay summary for this report; one of them did and suggested some edits to it.

Analysis of national NHS patient experience data 2005/6–2014/15

We aimed to determine if:

1. the PE in each setting has changed over time

2. hospital trusts have performed consistently over time

3. there is consistency between hospital settings (ED, inpatient department and OPD).

All 130 acute non-specialist hospital trusts that had inpatient survey results for the 10-year period from 2005/06 to 2014/15 were included. Initially, descriptive analysis of data from the NHS Patient Experience Tool36 was used to determine the patterns in PE scores over time for overall PE, domains and individual questions, inpatients, outpatients and ED scores. Scores in 2005/06 and 2014/15 were also compared. To determine if performance of the highest and lowest scoring trusts was consistent over time, the mean score for each trust for each domain in the first 3 years was calculated, and the 25% highest-scoring and 25% lowest-scoring trusts were identified. The mean scores for these groups of trusts were then calculated and plotted each year.

The consistency of trust performance over time, and trusts’ performances relative to one another, were analysed. The trusts were grouped into four groups using k-means cluster analysis on standardised PE scores. Ward’s minimum variance hierarchical clustering was used to determine the appropriate number of clusters, which ranged from four to nine in different years. Consistent performance was defined as being in the same ranked cluster for more than 5 years.

The results show that overall PE was good during the 10 years, with modest improvements over time across the three hospital settings. Individual questions with the biggest improvement across all three settings were cleanliness (inpatient department: +7.1 points, ED: +6.5, OPD: +4.7) and information about danger signals (inpatient department: +3.8, ED: +3.9, OPD: +4.0). The lowest-scoring questions, regarding information at discharge, were the same in all years and all settings.

The greatest improvement across all three settings was for cleanliness, which has seen national policies and targets. Information about danger signals and medication side effects showed the least consistency across settings and scores remained low over time, despite information about danger signals showing a big increase in score. PE of aspects of access and waiting declined, as has experience of discharge delay, likely reflecting known increases in pressure on England’s NHS.

Consistency over the 10 years was high. A total of 71.5% of trusts were in the same ranked cluster for more than 5 of the 10 years for overall scores. There was also high consistency for individual domains. The gap between the lowest- and highest-performing trusts in the initial period narrowed during the first 3 years, but there was little evidence of the lowest-performing trusts ‘catching up’ after this, except for the ‘Clean, Comfortable, Friendly Place to be’ domain.

Questions regarding waiting, information about medication side effects and danger signals have been consistently low scoring in all three settings since the survey inception. High-scoring questions also show consistency over time and across settings and include being treated with respect and dignity and being given sufficient privacy.

Before the analysis, we shared the patient survey questionnaires with our two public and patient involvement (PPI) representatives and asked them which specific questions or domains were of most importance to them, given their experience of the NHS. They both identified readiness for discharge as problematic. Our findings support this, and we therefore included scores for PE of waiting for a bed after arrival at hospital and PE of discharge in the subsequent regression models.

Other national data

Table 3 describes the variation by general practice or hospital for the non-patient-level data used in the regression modelling. This information came from 7756 general practices and 141 hospital trusts.

Sources of data for the above table:

• Number of GPs – Data as at 30 September 2010 37

• GP Patient Survey 2010/11 38

• Quality and Outcomes Framework 2010/11 39

• overall Patient Experience Scores 40

• Staff Survey 2010 41

• Monthly NHS Hospital and Community Health Service Workforce Statistics42,43

• Average Daily Number of Available and Occupied Beds by Sector, NHS Organisations in England, 2009–10. 44

Survey of pulmonary and cardiac rehabilitation programmes

The survey consisted of two parts: postal/e-mail questionnaire and inspection of programme websites. For the survey, despite the support and help from the national rehabilitation audits and two reminders, only 32 responses were received. This was felt to be too few to use, and these are not considered further. Websites were found for all but 2 of the 288 cardiac rehabilitation programmes and all but 6 of the 240 pulmonary rehabilitation programmes listed in the respective national audit programmes. Four were described as joint cardiac and pulmonary.

We assessed each website on four criteria. Of the cardiac rehabilitation sites, 49% were specific to the rehabilitation programme (e.g. rather than to the parent trust), 58% provided contact details and/or the address, 40% gave the acceptance criteria for the programme and only 15% provided links to resources useful to patients, such as further information on heart disease or relevant charities. For the lung rehabilitation sites, these figures were similar at 42%, 50%, 31% and 16%, respectively. We were almost always able to obtain a postcode for the programme and were therefore able to estimate the as-the-crow-flies distance between the patient’s postcode and their nearest rehabilitation centres. Figure 1 shows the services on a map. These four pieces of information were considered as covariates in the mortality and readmission models.

FIGURE 1.

Location of cardiac and pulmonary rehabilitation programmes in England as of 2015.

Predictors of 1-year mortality

In total, 14.9% of HF patients died during their index admission and another 24.8% died within a year after discharge, resulting in an overall mortality rate within 1 year of index admission of 39.6%. With regard to COPD patients, 5.9% died during their index admission and another 18.2% died within a year after discharge, resulting in an overall 1-year mortality rate of 24.1%. Table 5 gives crude outcome rates for selected patient characteristics.

To include information on general practices and rehabilitation programmes, some records had to be excluded from both the mortality and the readmission models: patients with missing distance from nearest rehabilitation programme (seven for HF and five for COPD); patients who are registered at practices with missing QOF data (1368 for HF and 1901 for COPD); patients who are registered at practices with missing GP Patient Survey data (438 for HF and 145 for COPD); and patients who are registered at practices with missing GP supply data (143 for HF and 551 for COPD). These losses totalled 1956 for HF (2.4% of the full cohort) and 2602 for COPD (2.6% of the full cohort).

Table 6 gives the ORs for the final set of predictors for total 1-year mortality (in or out of hospital). For HF, we also included the National Audit hospital-level performance figures; the only measure significant at a p-value of < 0.01 was referral for echocardiography, but the effect size was tiny and it is therefore not shown.

The two mortality models fitted the data well in terms of residuals but showed some overprediction of low risk (miscalibration). Discrimination (c-statistic) was noticeably higher, at a c-statistic of > 0.7, compared with readmission (see Predictors of 30-day emergency readmission). Older age, male sex, non-white ethnicity and a number of comorbidities such as prior stroke, pneumonia, renal disease, cancer and cognitive impairment were associated with higher odds of mortality within 1 year of the index admission. LOS of ≥ 1 night and missed prior outpatient appointments were strong predictors for both conditions. Intensive care use (both conditions) and the severity proxies for COPD were all associated with higher odds. Hospital and GP factors that we considered were sometimes significant but small in size.

Predictors of 30-day emergency readmission

Approximately one in five (19.8%) HF and one in six (16.5%) COPD patients who were discharged alive from their index admission were readmitted for any cause within 30 days. Table 7 gives the ORs for the final set of predictors for readmission within 30 days of live index discharge. For HF, we also included the National Audit hospital-level performance figures, but none came close to statistical significance and these are not shown.

The two models fitted well, however, with low discrimination (c-statistic). Older age and a number of comorbidities such as ischaemic heart disease (IHD), renal disease, cognitive impairment, mental health conditions and pneumonia were associated with higher odds of readmission for both patient groups. LOS was significant for both, but the pattern differed, with same-day discharges for HF and 2-night stays for COPD having the highest odds of readmission. Missed outpatient appointments were a strong predictor for both conditions, with 9% higher odds of readmission per appointment missed in the previous year. Larger hospital size and fewer doctors per bed were both associated with higher odds, though teamworking rating by staff remained in the model for COPD only. No GP factors that we considered were retained.

As well as modelling the standard all-cause 30-day measure, we compared predictors for readmission, split into readmissions for the index condition and those for other causes. Of the 13,099 all-cause 30-day readmissions in the HF cohort, 28.6% had a primary diagnosis of HF, compared with a high of 32.8% at 7 days and only 22.7% at 1 year. For COPD, of the 15,074 all-cause 30-day readmissions, 39.1% had a primary diagnosis of COPD. This, again, was highest at 7 days (43.2%) and fell to 36.2% at 1 year, a much smaller difference than for HF.

Rather than present the two very large tables, we will now summarise the differences in the predictors for readmissions with the same primary diagnosis as the index admission (i.e. HF or COPD) and predictors for any other readmission diagnosis. For HF patients, the statistically significant (p < 0.01) effects of older age, comorbidities (such as IHD, peripheral vascular disease, pneumonia, COPD, diabetes mellitus, renal disease) and prior missed outpatient appointments were similar; no significant associations for any readmission diagnosis were seen for sex, trust factors, GP factors or other community factors. A few variables showed significant associations with readmissions for HF only: black ethnicity (OR 1.44, 95% CI 1.16 to 1.79; p = 0.001), valvular disease (OR 1.26, 95% CI 1.17 to 1.36; p < 0.00010), defibrillation (OR 1.61, 95% CI 1.18 to 2.20; p = 0.0024) and same-day index discharge. Compared with an index LOS of zero, an index LOS of 1 night had an OR of 0.77 (95% CI 0.65 to 0.90) and p-value of 0.0011, a 2-night index stay had an OR of 0.64 (95% CI 0.53 to 0.76) and a p-value of < 0.0001, and index stays of ≥ 3 nights had an OR of 0.64 (95% CI 0.56 to 0.72) and a p-value of < 0.0001. On the other hand, a few variables showed significant associations with readmissions for non-HF diagnoses only: deprivation (p = 0.009), cancer with metastases (OR 1.38, 95% CI 1.09 to 1.73; p = 0.0063), cognitive impairment [senility and dementia (OR 1.37, 95% CI 1.27 to 1.47; p < 0.0001) and mental health conditions excluding dementia (OR 1.21, 95% CI 1.12 to 1.30; p < 0.0001] and living alone (OR 1.11, 95% CI 1.04 to 1.19; p = 0.0017).

For COPD, the main similarities were for sex (females had lower odds of readmission), pneumonia, mental health conditions except dementia, referral for echocardiography (15% higher odds if recorded), prior missed OPD appointments (stronger effect seen for non-COPD readmissions) and, as with HF, the lack of any significant associations with any of the hospital, GP or community factors that we tried. Age relations differed considerably by readmission diagnosis. Compared with patients aged 60–69 years, those aged under 55 years had lower odds of COPD readmissions but similar odds of other readmissions. Those aged ≥ 70 years had only slightly higher odds of COPD readmission (and not statistically significantly higher for ages of ≥ 85 years), but much higher odds of non-COPD readmission, rising to a peak OR of 1.75 (95% CI 1.55 to 1.97) and a p-value of < 0.0001 for those aged ≥ 90 years. Just two variables showed significant associations with readmissions for COPD only: deprivation and non-invasive ventilation on admission (OR 1.29, 95% CI 1.15 to 1.45; p < 0.0001). A much larger number of variables showed significant associations with non-COPD readmissions only: almost all comorbidities, living alone (OR 1.21, 95% CI 1.13 to 1.29; p < 0.0001) and index LOS [the lowest odds were for 2-night stays (an OR compared with same-day discharges of 0.88, 95% CI 0.79 to 0.98; p = 0.0165) and the highest odds were for stays of ≥ 3 nights (OR 1.10, 95% CI 1.02 to 1.19; p = 0.0168]. The direction of association for hypertension differed by readmission diagnosis, with lower odds if readmitted for COPD (OR 0.91, 95% CI 0.86 to 0.96; p = 0.0013) but higher odds if readmitted for any other diagnosis (OR 1.07, 95% CI 1.02 to 1.12; p = 0.0032).

In summary, for both cohorts the set of predictors of readmission for the index condition showed several differences from the set for readmission for other conditions. There were fewer predictors, perhaps because of the smaller sample size. However, there were a few predictors of readmissions for the index condition that did not significantly predict readmissions for other conditions; for example, index LOS and defibrillation were significant predictors of readmission for HF only. There were very few predictors for non-COPD readmissions that differed from the list of predictors for all-cause readmissions for the COPD cohort.

Attendance at accident and emergency following the index stay and likelihood of admission during that attendance

Many of the patients with either of our two chronic diseases were regular visitors to each hospital setting. We began by counting the number of ED attendances, OPD appointments, elective admissions and emergency admissions per patient in the year following index discharge (survivors of the index only). For HF these are summarised in our British Medical Journal Open article (see Chapter 4, Dissemination activity), together with the proportions of patients for whom the National Institute for Health and Care Excellence guideline number 187,45 for cardiologist follow-up within a fortnight of inpatient discharge, was met (just 7% of patients overall, with large differences by age and comorbidity). To better understand the drivers of readmission, and as preparation for objective 2, we ran some further analyses on the use of ED after discharge from the index stay. There are two parts: (1) regression modelling to determine which factors are associated with (a) ED attendance and (b) admission to hospital at that ED attendance; and (2) an assessment of the statistical properties of different indicators covering ED attendance and readmission.

During our 2 index study years, 66,219 HF patients were discharged alive from their index HF admission, of whom 11,513 (17.4%) attended the ED within 30 days with no intervening elective or emergency admission (as described in Chapter 2). Of the 90,351 COPD patients discharged alive from their index COPD admission, 14,039 (15.5%) attended the ED within 30 days using the same definition. Of these ED attendances, 76.9% for HF and 74.0% for COPD resulted in admission. There were some variations in the proportion of attendances that resulted in admission by day of the week, time of day, the number of elderly patients waiting and the number of patients admitted during the hour of arrival. Table 8 gives the 7- and 30-day attendance rates and proportion who were subsequently admitted for each cohort.

In Table 8 the proportion admitted is slightly lower at the weekend, especially on Sunday, and lower during the day shift. In contrast, admission appeared more likely with an increasing number of elderly patients waiting when the HF or COPD patient arrived and also with an increasing number of patients admitted that hour. There was a large jump in the proportion of HF or COPD patients admitted, from no patients admitted (irrespective of presenting complaint) to one patient admitted. This was most noticeable for COPD patients, as if they arrived at the ED (within 30 days of their index discharge) in an hour when no one was admitted, their own chance of admission was 63%, and this rose to nearly 74% if one person was admitted. The proportion admitted was also higher if the patient arrived at a time when waits were longer, as measured by breach of the 4-hour target at either the old strict 98% or the less stringent 95% rate threshold. There were no clear associations with the number of boarders or with relative busyness of the department, although one might say that a few boarders – two or three – were associated with higher odds than having either no or many boarders.

Table 9 gives the significant predictors from the regression model of ED attendance within 30 days of index discharge. These included older age, particularly for COPD, male sex (COPD only), deprivation, living alone, prior missed OPD appointments, same-day index discharge (HF only), defibrillation implantation (HF only), referral for echocardiography (COPD only) and most of the comorbidities on our list. In contrast with 30-day readmissions in objective 1, a number of the hospital and practice factors that we included showed significant associations with ED attendance for COPD. The biggest effects were for staff survey – would recommend to friends and family (OR per point increase 0.66; p < 0.0001) and number of GPs (FTE) per 1000 patients in the practice (OR 0.87; p = 0.002) and, for HF, QOF-recorded prevalence in 2009 (OR per percentage point prevalence increase 0.84; p < 0.0001). These associations are in the expected direction: ED attendance was more likely in COPD patients when the hospital scored worse on the Friends and Family Test, when there was lower GP supply and, for HF, lower HF prevalence at the practice.

Once the patient had attended the ED, we assessed their odds of being hospitalised at that visit. Table 10 gives the ORs for variables with a p-value of < 0.01.

The main predictors were:

• older age (for HF, with weaker evidence for COPD) but also age < 60 years (COPD only)

• index LOS of ≥ 3 nights

• non-invasive ventilation during the index COPD admission

• evening or night attendance (both conditions)

• comorbidities of HF, pneumonia, obesity or cancer (all COPD only – having a coded mental health condition was associated with 13% lower odds of admission)

• two hospital-level variables:

  1. for each patient admitted for any condition from the ED during the hour of arrival, the odds of the HF or COPD patient being admitted rose by 5% for HF and 2% for COPD

  2. the odds of HF patients being admitted from the ED were 40% higher if the overall proportion of waiting patients seen within 4 hours was < 98% than if it was ≥ 98%.

In other words, admission from the ED became more likely if our HF patients arrived during the evening or at night, if they arrived when other patients were being admitted (both also true for our COPD patients) or if they arrived when other patients were waiting a long time. In contrast to the crude cross-tabulations in Table 8, the regression model found no association by day or time of arrival or with the number of elderly patients waiting.

Although the majority of 30-day readmissions were via the ED, around one in five was not. We might therefore expect some predictors of readmission to differ from the set of predictors for admission following an ED attendance, and this was true. Age patterns were similar, though the ORs for admission from the ED in the COPD cohort were much lower than the ORs for any readmission. Most of the comorbidities were statistically significant predictors of readmission, but only a few remained in the model for admission from the ED. The number of prior OPD appointments missed was a strong predictor of readmission but not associated with admission from the ED. Index LOS showed different patterns, with same-day discharges having higher odds of any readmission but much lower odds of admission from the ED. Hospital size and doctors per bed were significantly associated with readmission via any route but not with admission from the ED.

Objective 1: what are the main patient, primary care and hospital factors associated with variation in readmission and mortality rates?

In these two elderly, multimorbid cohorts, hospital use and 1-year mortality rates were high. A number of studies reported survival from diagnosis, but we were interested in survival since first admission. One-year mortality following HF admission has been estimated at between 20% and 40% in the review by Cowie et al. ;48 the most comparable figure in terms of population and methodology, although not time period, is 45% from Scotland in the decade to 1995. 49 Rates have fallen since, and our figure was 40%. In-hospital mortality for COPD varies from between < 10% and 60%, based on the severity level of the population studied50 and the 1-year rate is about one in four (28% for Canadian discharge data between 2001 and 2004;51 and 25% within 1 year of admission for the first acute exacerbation between 2005 and 2007 in Taiwan). 52 Our mortality rate within 1 year of an index COPD admission was 24%.

We found the main predictors for 1-year mortality for both conditions to be similar, and they included older age, male sex, non-white ethnicity, severity proxies, prior missed OPD appointments, LOS of ≥ 1 night and a number of comorbidities, such as prior stroke, pneumonia, renal disease, cancer and cognitive impairment. Hospital and GP factors that we considered were sometimes significant but small in size. Other studies have also found higher odds for older age,51,53,54 severity,51,55 comorbidities such as HF (in COPD patients), cancer, renal failure,55 depression56 and diabetes. 53 For COPD, Fidahussein et al. 54 also found an association for LOS. Not all studies reported sex differences, although some agreed with our finding of higher odds for males for both conditions. 51,55

We found that one in five HF and one in six COPD patients were readmitted for any cause within 30 days of live index discharge. These rates are slightly lower than those reported for the US Medicare population (25% for HF57 and 20% for COPD58) or 21% from the wider US population using the Nationwide Inpatient Sample,59 although much lower rates have been reported for a private insurance plan in the USA (8%). 60 Some studies also reported disease-specific readmission rates (e.g. 7% for acute lower respiratory infection or acute exacerbation for COPD in New Zealand61 and also 7% for COPD or bronchiectasis in the USA),59 showing that readmissions within 30 days for the index condition comprise about one-third of the total, as in our study.

In our analysis, readmission was more likely for both index conditions with older age, missed outpatient appointments and a number of comorbidities such as IHD, renal disease, cognitive impairment, mental health conditions and pneumonia. LOS was significant for both, but the pattern differed, with same-day discharges for HF and 2-night stays for COPD having the highest odds of readmission. Males had significantly higher odds for COPD. Larger hospital size and fewer doctors per bed were both associated with higher odds, although teamworking rating by staff remained in the model for COPD only. No GP factors that we considered were retained. Other studies had also noted the association for older age and comorbidities,54,60,62 although the much-cited systematic review of risk adjustment and risk prediction models for readmission in HF patients by Ross et al. 62 found much inconsistency in the selection of variables across models, with the exceptions of age, sex and race/ethnicity. We did not find any statistically significant difference in odds by sex for HF or by ethnicity for either cohort, except lower odds for the ‘not known’ group, which we commonly find in our models,11,18 and may relate to this group consisting more of overseas patients and others with lower follow-up rates. For the COPD cohort, we found higher odds of readmission in deprived areas, and in US studies, race/ethnicity often captures the deprivation aspect. Sharif et al. 60 found an association for COPD patients with LOS (higher odds if < 2 or > 5 days, whereas for us it was 2 days) and several quality-of-care aspects, such as prescribing and follow-up.

Objective 2: should accident and emergency attendance and reattendance data be considered alongside readmission metrics when measuring hospital performance in terms of unplanned activity? If so, how?

Much less attention in the literature has been paid to ED attendances after inpatient discharge than to rehospitalisation. This gap was recognised by Brennan et al. ,63 who, like us, examined 30-day ED utilisation and all-cause readmissions following a hospital admission, albeit for all patients combined rather than for any given index condition. With their national US data, they found that nearly one in five patients presented to the ED within 30 days of an inpatient hospitalisation and over half of these patients were readmitted. They concluded that ‘readmission measures that incorporate ED visits following an inpatient stay might better inform interventions to reduce avoidable readmissions.’ Vashi et al.,64 in the USA, looked at ED visits within 30 days after discharge and also a combined measure of ED visits (not resulting in admission, or ‘treat and release’) or readmission within the same time frame, a rare example of combining the two as we have done. They reported figures overall and for some high-volume conditions. For HF and shock, their 2008–9 rates were as follows: ED treat and release, 9.6%; readmission, 27.7%; and combined (i.e. ED visit or readmission), 37.4% within 30 days. Our rates for HF were all lower: 4.0% for ED treat and release, 19.8% for readmission and 22.7% for combined within 30 days.

In our HF cohort, 6.2% visited the ED within 7 days and 17.4% with 30 days; for the COPD cohort these figures were 5.3% and 15.5%, respectively. Most figures in the literature are from the USA. A total of 22.5% of Medicare beneficiaries came to the ED within 30 days after hospitalisation for HF, with congestive HF the most common reason for attendance, although accounting for only 11% of these visits. 65 The same study noted wide hospital-level variation in post-discharge ED visit rates for each condition: acute myocardial infarction (AMI) (median 8.3%, 5th to 95th percentile 2.8% to 14.3%), HF (median 7.3%, 5th to 95th percentile 3.0% to 13.3%) and pneumonia (median 7.1%, 5th to 95th percentile 2.4% to 13.2%). The authors concluded that ‘policymakers and researchers should further study post-discharge ED visits as measures of health-care access and care transitions in the vulnerable Medicare population.’ Another US study found that 84% of patients who present to the ED with acute HF are admitted, with some regional variation but huge differences by insurance status. 66 With HES A&E data, we were unable to identify whether the ED visit was for HF or COPD.

Studies on COPD focused on ED presentations for acute exacerbations rather than all ED visits as with our study. In 16 EDs across Canada, 49% of 501 patients with acute exacerbations of COPD who were interviewed had been admitted. 67 In a Spanish cohort also covering 16 EDs, 62% of patients were admitted on arrival for a COPD exacerbation. 68 For our two cohorts, approximately three in four ED visits for any reason within 30 days resulted in admission. These conversion rates were higher than for all ED visits in England during the same period according to published figures from HES: for those aged < 65 years attending between 2009/10 and 2012/13, only 18% resulted in admission, whereas for people aged 65–84 years the figure was 50% and for people aged ≥ 85 years it was 63%. 69

Regarding the factors associated with admission at an ED visit, there have been a few studies covering all ED attendances. Using Healthcare Cost and Utilisation Project data, Pines et al. 70 found higher ED admission rates to be associated with a range of factors such as more inpatient beds, higher hospital occupancy rates and, at county level, more primary care physicians per capital.

An analysis of the 2010 National Hospital Ambulatory Care Survey ED data of hospitals with admission rates from the ED of between 5% and 50% did not report the proportion of ED visits resulting in admission, but focused on the variation in ED admission rates by hospital after accounting for patient and hospital factors: ‘even after accounting for hospital teaching status, ownership, urban/rural location, and geographical location, 7.0% of the variation in risk-standardized hospital admission rates from the ED was still attributable to an institution-specific effect.’71 Another US study, from three hospitals, assessed the variation in rates of admission from the ED by individual physician. Among the 89 attending emergency physicians, admission rates varied from 21% to 49%, similar to previous Canadian studies. 72 Taken together, these US studies show appreciable variation in ED admission rates by hospital, some of which can be explained by specific hospital factors but some that cannot, with individual physician decision-making being one likely contributing factor to the variation.

There have also been some studies specific to HF and COPD patients. These have found higher odds of ED admission for age,66,67 various markers of COPD severity and respiratory distress,67,68 comorbidity and lower levels of social support, but not sex, although Vidal et al. 68 did not retain comorbidity or social support in their final model. We also found effects of older age (although Vidal et al. 68 did not, perhaps because of their detailed physiological variables) and comorbidity for the COPD group. Unlike Pines et al. ,70 we found no association between higher odds of ED admission and hospital size or quarterly occupancy levels, but instead found that the number of patients admitted during the hour of arrival – a likely proxy for bed availability and hence lower occupancy at that time – to be positively associated with the odds of admission. For HF patients, admission was more likely if they arrived during a period of long ED waiting time, as measured by the proportion seen within 4 hours. We are not sure why this would be the case, other than a chance finding. We have not seen other studies that have assessed this effect in this way.

In summary for this objective, previous studies have highlighted patient factors such as age and disease severity, and the role of the hospital and attending physicians in predicting the odds of admission from the ED. Our findings also show the importance of those patient factors plus the hour of attendance and the number of patients admitted at the same hospital during the hour of arrival. Some commentators have advocated the use of ED visits, perhaps in combination with readmissions, as a performance metric63,64 but without giving clear guidance on how this should be done. We took a statistical approach to comparing the different options and also found considerable variation in hospital-level rates, particularly for ED visits. Variations in readmission rates were consistent with random (binomial) variation as suggested by the dispersion parameter of near one, the small number of funnel plot outliers and low ICCs. In contrast, the ED visit rates showed a lot more than purely random variation – they were very overdispersed, generating many more funnel plot outliers. Before an indicator that shows such overdispersion can be used in practice, the reasons for the extra variation should be sought. Spiegelhalter35 warns against using indicators where the reasons are largely related to data quality flaws or inadequate risk adjustment. The ICC and omega statistics are useful here, and both showed bigger hospital ‘footprints’ for ED visits than for readmissions. Our a priori suggestion of combining visits within 7 days with readmission within 30 days performed no better than the standard readmission within 30 days, and the other combinations were also inferior to the straightforward ED visits. After including hospitals as fixed effects in the model, as per Brown et al. ,33 in order to capture remaining hospital factors for which we did not have data, only ED visits had omega statistics of < 1, with 95% CIs that excluded one for HF. This means that the influence of the hospital was greater than that of the patient in explaining hospital-level rates, a desirable property.

Objective 3: how consistently do hospitals perform across different readmission-type metrics?

We assessed this using simple correlation and by counting funnel plot outliers with 95% and 99.8% control limits. Correlations varied greatly. There was moderate correlation between 30-day ED visit rates and 30-day readmission rates (ρ just above 0.5); adding ED visits to readmissions made little difference to the relations between ED visits alone and readmissions alone. We have found before that even very high correlations, where ρ exceeds 0.9, the effect on which hospitals are labelled as funnel plot outliers is unpredictable: up to 10% of hospital trusts changed outlier status when different risk adjustment models were used in deriving hospital mortality rates. 73 With much lower correlations in this project, we found that many trusts had different outlier status depending on the measure used. ED visits generated more outliers than readmissions, although when hospitals were flagged as outliers on both measures, it was usually the same set of hospitals for both measures.

Objective 4: are the results for chronic obstructive pulmonary disease similar to those for heart failure?

The two cohorts had several similarities, such as having the clear majority aged ≥ 65 years with, often, a number of comorbidities – including the other condition – plus high use of the ED and inpatient wards after their index discharge and high mortality at 1 year. Many of the patient factors that we included, such as age, index LOS and comorbidity, were predictive of readmissions or mortality for both conditions, although the patterns for the index LOS differed. Likewise, the relations between more doctors per bed and lower odds of both 1-year mortality and 30-day readmission were significant and similar for the two groups, as was the relation between more hospital bed numbers and lower odds of readmission, but the other hospital, community and GP factors that we tried were usually not significant for either group or sometimes inconsistent.

Patterns of hospital-level rates of ED visits and/or readmissions were very similar for the two cohorts. Furthermore, there was significant but modest correlation between the rates for HF patients and the rates for COPD patients for readmissions and strong correlation between the rates for ED visits within 30 days. Correlation for 1-year mortality was moderate (ρ = 0.58). In contrast, we found no relation between readmission rates for the index condition (HF or COPD) and rates for all other conditions combined. Although the literature is large on the relations between performance on different measures or process versus outcome rates by health-care unit, there have been few such comparisons of rates for different index conditions, especially for readmissions. One example is the study by Horwitz et al. 74 The authors compared publicly reported 30-day risk-standardised mortality and readmission rates for Medicare patients admitted with AMI, HF and pneumonia across the USA. Every mortality measure was significantly, but modestly, correlated with every other mortality measure (range of correlation coefficients, 0.27–0.41; p < 0.0001). Every readmission measure was significantly but also modestly correlated with every other readmission measure (range of correlation coefficients, 0.32–0.47; p < 0.0001, similar to what we found with HF vs. COPD). Coefficients were higher among larger hospitals, presumably at least in part because large hospitals’ rates are estimated more precisely. The authors concluded that ‘that there may be common hospital-wide factors affecting hospital outcomes’. If this is the case, then improvement efforts might be better focused on hospital-wide, rather than simply disease-specific, activities. The authors also suggested that correlation could vary according to hospital characteristics; for example, smaller or non-teaching hospitals might be more homogeneous in their care than larger ones. Horwitz et al. 74 had over 4000 hospitals on which to run such subanalyses, but we would have much less statistical power with HES to do this in England.

Finally, for each cohort we considered whether or not a hospital with high (or low) readmission rates for the index condition also had high (or low) readmission rates for any other primary diagnosis in the same set of patients. We found no such correlation for either HF or COPD patients. We are not aware of other literature on this, apart from our earlier HF analysis. 15