Key result
EuroSCORE II provides good discrimination for predicting mortality in general cardiac surgery patients, comparable to the STS model, though performance drops in highly selected subgroups.
Why the study?
Does EuroSCORE II accurately predict mortality in general and high-risk cardiac surgery patients compared to STS and older EuroSCORE models?
Does EuroSCORE II accurately predict mortality in general and high-risk cardiac surgery patients compared to STS and older EuroSCORE models?
Answers are grounded in this paper and cite their sources.
EuroSCORE II is a simple, easy-to-use model that performs comparably to the more complex STS model for predicting mortality in general cardiac surgery, though discrimination naturally decreases when applied to highly selected high-risk subgroups.
Two articles address the use of EuroSCORE II [1] in this issue of the journal. Kirmani et al. [2] compare the performance of the model against the Society of Thoracic Surgeons (STS) [3–5] models for predicting mortality in a single-institution cohort of some 15 000 cardiac surgery patients operated on between 2001 and 2010. They find that both models work very well, with an area under the receiver operating characteristic (ROC) curve of 0.818 (EuroSCORE) and 0.805 (STS). As EuroSCORE II is a relatively simple and easy-to-use model, they conclude that the choice between two equally discriminating models ‘falls in favour of the easier to use’. Howell et al. [6] compares EuroSCORE II against the older logistic [7] and additive [8] models in a bi-institutional cohort of 933 very high-risk patients (with a logistic EuroSCORE of >10). They find good calibration for EuroSCORE II (predicted mortality 9.3 and actual 9.7%) in this subset, but discrimination within this selected group has fallen to an area under the ROC curve of 0.66 (0.67 for the older models). The findings of Kirmani et al. are most encouraging for users of EuroSCORE II. They have shown that this simple parsimonious model, with fewer than 20 risk factors whereby risk can be calculated within a minute or so, is capable of matching the more complex and interaction-sensitive multi-tool models of the STS. They also found good news for the STS model: it can remain quite powerfully discriminatory even in the subsets of patients that it supposedly does not support. Finally, their comments on the limitations of the Hosmer–Lemeshow test are most apposite and should be taken seriously by those interested in the statistics of risk modelling. Howell's findings are superficially intriguing, but the explanation for them is statistically simple. If you stratify patients according to a risk score, and then consider just one of the strata, the regressors and their coefficients within the stratum will be different from those which allocated them to that risk group in the first place. Therefore, we should not be surprised if discrimination drops to a lower level within that stratum. In fact, we at the EuroSCORE project group were pleasantly surprised that it was as high as 0.66. It must be remembered that all the EuroSCORE models were created for stratifying risk in the overall cardiac surgical patient population (different risk groups and different procedures). Removing subgroups from this population for specific analysis will invariably result in the loss of discrimination. If you take this to extremes, and choose to apply the risk model to a substratum of exclusively 63-year-old females with good left ventricular function and no comorbidity having isolated first-time elective aortic valve replacement, do not be surprised if discrimination becomes altogether nonexistent. We have two excellent models for predicting risk in general cardiac surgery. This does not prejudice the fact that risk prediction in the subsets of cardiac surgery can be improved, and we invite and encourage others with an interest in this field to create and propose such models for use in these situations. In the meantime, the achievements of risk modelling in this specialty already far outstrip anything offered in any other field of medicine, and of that cardiac surgeons can be justifiably proud.
No takes yet. Share an insight, caveat, or question.
Nashef et al. (2013) conducted an editorial in Cardiac surgery. EuroSCORE II vs. STS model and older EuroSCORE models was evaluated on Mortality prediction (discrimination). EuroSCORE II provides good discrimination for predicting mortality in general cardiac surgery patients, comparable to the STS model, though performance drops in highly selected subgroups.
Synapse has enriched 10 related papers. Consider them for comparative context: