We agree with Challen and colleagues that mortality prediction tools and early warning scores should only be used to predict what they have been validated to predict. It is important to recognise, however, that early warning scores have been widely implemented in acute medicine in the UK, and it is therefore inevitable that junior and inexperienced physicians will use these as prognostic tools and to guide the intensity of intervention required in community acquired pneumonia (CAP). The need for higher level care, albeit important, is not the only clinical decision that needs to be taken when a patient with CAP is admitted to hospital. For example, the physician also needs to decide about the intensity of antibiotic therapy and other supportive measures, and prognostic assessment may also be useful in discussions with patients and their relatives. In order to decide on what tool to use, the key question for the physician at the frontline is: “What do I want to predict?” As the performance characteristics of all decision support tools are dependent on the context in which they are used, the answer to this question will partly depend on the environment in which the physician finds her/himself. This is likely to explain the different performance characteristics of CURB-65 and its predecessors in different patient cohorts. In a recent study by Capelastegui and colleagues, CURB-65 was as good as the pneumonia severity index in predicting mortality and also appeared usefully to stratify the need for mechanical ventilation (0.74% (11/1480) of non-severe patients vs 2.36% (7/296) of severe patients) and hospital admission in a mixed cohort of outpatients and inpatients with CAP.1

In the National Health Service in the UK, relatively few patients hospitalised with CAP are admitted to higher level care. Of 433 patients with full data to calculate a CURB-65 score in our own database, 14 (3%) were admitted to intensive care, although 38% had a CURB-65 score of ⩾3. In addition, those who were admitted were younger (mean age 54 years in those admitted to intensive care vs 70 years in the whole cohort) and had less co-morbidity (36% of those admitted to intensive care had a chronic illness vs 65% in the whole cohort). In this context, it is unsurprising that a tool that includes age as one of the criteria will perform less well than a tool that does not in predicting the need for admission to higher level care. Mortality, in contrast, is an outcome that applies to all patients regardless of age, co-morbidity or other patient characteristics. In a different environment such as in a country that admits a higher proportion of patients (eg, North America) or very few patients (eg, a developing country) with CAP to higher level care, or in the event of an influenza pandemic when higher level care is likely to be tightly rationed, the performance characteristics of both tools and what one wants to predict may change notably. Interestingly, in our own cohort, of the patients admitted to intensive care, CURB-65 still appeared to usefully stratify mortality (0/3 patients with a score of 0 or 1 died vs 2/5 (40%) with a score of 2 and 4/6 (66%) with a score of ⩾3). It is also worth noting that early warning scores have not been validated to predict mortality in patients with either sepsis or specific infections. In contrast, CURB-65 may predict mortality in a wider range of infections.2 As we state in our paper,3 there may be a case for using CURB-65 on admission to hospital to guide the initial intensity of management, and thereafter or at the same time, an early warning score to identify the small number of patients who will require higher level care. Research to derive and validate better and simple prognostic tools that predict a range of patient outcomes is clearly required.