Discussion

Death during or within 30 days of a COPD admission is common and because it is measurable it is a potential quality indicator. There is wide residual variation between hospitals even after control for age, deprivation and comorbidity such that a quarter of hospitals are outliers on a 2-year standard funnel plot implying factors other than chance at work. However, while care quality might be a factor there are many other potential explanations that are either unmeasured (case severity) or uncontrolled for (admission thresholds) which limits its value as an indicator.

The 2003 NCROP audit suggested a lower death rate in larger units with more staff and facilities implying potentially avoidable deaths. Between 2006–2007 and 2007–2008 there had been no major initiatives expected to alter care and no change in HES coding so we might have expected some consistency in mortality rates. However, while national mean values for years 1 and 2 are similar and individual hospital values correlate statistically there is high variance of between-hospital values with as many rising as falling. For a given hospital to be considered statistically different from 1 year to the next a 3.8% change in mortality rate (1.67 SD) or a more than 40% shift from the mean is necessary. That alone makes year-on-year data challenging to interpret.

While many COPD deaths are inevitable the proportion that is preventable remains unknown and may vary with non-medical care factors. Thus, good practice such as admission prevention schemes could perversely increase hospital per cent mortality by not admitting those more likely to survive; although absolute death numbers for the population would remain similar. Conversely a hospital that succeeded in keeping patients with the most severe disease alive would look ‘good’ in that year, but the natural history cannot be radically changed and those ill patients will be added to the next few years' mortality data. Assessment of mortality over a 3–5 years period would have less variability and might identify a few hospitals with persistently high mortality but could still not correct for the balance between community and hospital while defeating the intent to have an annual metric. Additionally it is vital that attempts to reduce mortality do not contribute towards failure to provide adequate end-of-life care to patients where death is inevitable.

Strengths of the HES database are inclusion of all emergency admissions and our case selection algorithm was designed to include all inpatients that clinicians would recognise as having COPD. Thus COPD coded at positions 2 or 3 with a respiratory infection (bronchitis or pneumonia) at position 1 would have been entered by respiratory clinicians into the 2008 NCROP audit as a COPD exacerbation. Omission of cases where COPD was second or third excludes 19% of COPD admissions and 40% of deaths resulting in a mortality rate considerably lower than NCROP or death certification. While our coding definitions are new they have been presented, discussed and agreed with UK COPD specialists and match national audit data. Using linked 30-day mortality creates an outcome not dependent on length of stay, collected robustly by the Office for National Statistics, comparable for all hospitals and includes all-cause deaths including cardiovascular disease. However our analyses have been repeated using inhospital and 90-day mortality with similar results.

Limitations of HES are that we have no specific measures of COPD severity (spirometry or blood gases) and diagnosis depends on coding; an area we have tried to improve upon by using three rather than only the first diagnosis code. The univariate analysis (confirmed in logistic regression) showing the most deprived patients had lesser mortality was a surprise but there are pointers that the most deprived patients are younger and have more admissions per year suggesting that the case mix within the denominator may vary with deprivation. Deprivation is not a significant predictor if the logistic regression is repeated using only last admission for each patient but UK hospital indicators are based on analysis by admissions rather than by individuals.

In conclusion 30-day mortality after hospitalisation is too variable to be a reliable annual metric of hospital COPD care as there appears to be considerable year-on-year variation related to case mix rather than care delivery. Due to an absence of disease severity markers it is impossible to distinguish avoidable from inevitable deaths and the analyses provide no indication of what aspect of healthcare should be changed. Moreover even when a death is avoided disease progression continues unchanged so, unlike asthma or a revascularised heart, the benefits are not long lived. More worryingly if it were to be adopted we would need to understand the reasons for variability and what a hospital should do to improve or much effort might have no detectable benefit. It might also mean that obvious good practice, such as admission prevention schemes, could be unsupported. This is not to argue that metrics should not be applied rather that annual mortality is a poor COPD metric; something which would also be true for certain other chronic diseases. Use of specific therapy such as non-invasive ventilation might represent a useful hospital metric but it is not reliably coded (only 30% coding in our institution). A metric we have suggested previously is bed days per unit population controlled for prevalence. This reflects patients’ desire to be managed at home, the state's desire to reduce hospitalisation and requires a joint community hospital approach; however, there may be others.