The EuroSCORE: are we contributing to its overprediction of mortality in cardiac surgery nowadays?

We read with great interest the article by Akar et al. [1] regarding validation of the EuroSCORE risk models in the Turkish adult cardiac surgical population. This manuscript is in agreement with previous reports [2–4] that additive and logistic EuroSCORE models overpredict mortality in cardiac surgery nowadays. However, in our opinion there are a few issues that should be clarified. The first one is a definition of mortality. The original EuroSCORE model [5] defined operative mortality as death within 30 days from operation or even later than 30 days if still in hospital (not 30-day mortality as the authors stated in Discussion section of their manuscript [1]). On the other side, their calculation and statistical observations were based only on in-hospital mortality. It is possible that, with the inclusion of all patients who should be included as perioperative deaths (30-day mortality plus in-hospital mortality), the findings of the study [1] would no longer be statistically significant. When we validate a risk stratification model on a population out of its original database, we have to check its discriminatory power and calibration. Model discrimination was analysed by determining the area under the receiver operating characteristic (ROC) curve [1– 5]. However, several statistical procedures have been used to determine calibration. The Hosmer–Lemeshow (H-L) test has been used in >90% of all manuscripts to test the EuroSCORE calibration in different patient population (the H–L test confirms good calibration if the P-value is >0.05), including the EuroSCORE working group [5]. In a few manuscripts calibration plots, unpaired t-test [1], chisquare test [4] or observed to expected (O/E) ratios were used to prove that calibration of the EuroSCORE models was inadequate. Thus, we are curious as to what would be the results of the EuroSCORE calibration checking in these articles if the H–L test was used. On the contrary, Yap et al. [2] wrote an excellent manuscript with intention of validating the EuroSCORE model in Australia. Mortality definition was correct (in-hospital mortality plus 30-day mortality), areas under the ROC curves (0.82) confirmed an excellent discrimination for both additive and logistic EuroSCORE models, but calibration using the H–L test was poor (P 0.05) of the logistic EuroSCORE model. Finally, our question is—should we use the H–L test any more, or we are looking for a statistical procedure that will confirm poor calibration of the EuroSCORE models?