Logistic versus additive EuroSCORE. A comparative assessment of the two models in an independent population sample

Objective: Validation of EuroSCORE outside the boundaries of the original database has been limited to the additive model and has occasionally shown inconsistencies. Therefore we sought to validate the logistic model and assess its predictive performance compared to the additive approach. Methods: Twenty-four hundred and twenty-six consecutive patients were prospectively assigned individual expected risks of dying calculated by the logistic and the additive EuroSCORE algorithms. Discriminating ability of the two models was tested by Receiver Operating Characteristic (ROC) curves. Calibration was assessed by the Hosmer-Lemeshow (H-T) test and further explored by additional cross-tabulations. A percent difference among the estimates was calculated and plotted across score groups. The series was then sorted by date of operation and split in halves to separately explore the potential effect of variation of performance. Results: Observed mortality (5.6%) was not significantly different from the additive (5.3%) and the logistic estimates (6.9%). Both models satisfactorily discriminated outcomes (ROC areas of 0.80 and 0.79 for the logistic and the additive model, respectively). The H-T test showed that calibration was good for the logistic model (P = 0.12) but turned out being inadequate for the additive model (P < 0.0001). Further cross-tabulations confirmed a good correlation among observed and predicted death rates by the logistic model across all groups. The additive model, on the other hand, revealed a propensity to over-predict in medium-risk categories and under-predict in the very high-risk cases. Direct comparison of additive vs logistic estimates showed a similar behaviour demonstrating it as an intrinsic property of the additive approach. The split-file analysis revealed a significantly improved outcome for patients treated in the second half of the series though the predictive performance of the two models was unaltered. Conclusions: Logistic EuroSCORE reliably predicted outcomes in our series despite the higher risk profile compared to the reference EuroSCORE sample and the observed variation in clinical performance during the study period. The additive model was less precise, exhibiting a predictive distortion which should be accounted for, particularly when employing it at the individual patient level. (C) 2004 Elsevier B.V. All rights reserved.