Table 1 Consider T-MACS—a CPM developed for the prediction of acute myocardial infarction in patients presenting to the emergency department with chest pain [13]. Suppose our intended use is initially for hospitals within the Greater Manchester (UK) area, and then we are considering rolling out the CPM across the UK

…a particular population and setting of intended use is of interest

T-MACS was developed using data from a hospital in Manchester, and validated using data from other hospitals in Manchester [13]. Hence the original development and validation results match the intended use.

If, however, we wished to use T-MACS in London, UK, the validation above would be of limited value. We would need a new targeted validation to examine performance in London, UK.

…multiple populations and settings of intended use are of interest

Suppose we wished to implement T-MACS across all hospitals in the UK. Then, validation would be required across many hospitals in the UK, potentially using individual patient data meta-analysis of performance across hospitals to evaluate heterogeneity in performance [5].

Such a validation could also be useful to indicate expected performance of T-MACS in UK hospitals not included in the validation set. This is because, if heterogeneity in performance is low across the included hospitals, this gives some confidence that the CPM will perform well in all areas in the UK.

…an arbitrary dataset is used without consideration of the intended population or setting

A further validation was conducted in hospitals in Australia and New Zealand [14]. For our intended use, this validation offers little evidence. However, it would be very valuable if we were considering using the model in Australia and New Zealand.