Systematic overestimation of mortality by SAPS II in contemporary ICU practice: a retrospective study analyzing ICU...

The present study re-evaluates the performance of the Simplified Acute Physiology Score (SAPS) II to predict in-hospital mortality of Intensive Care Unit (ICU) patients by analyzing the French nationwide hospital stay billing database between 2013 and 2024.

From 2,702,204 adult ICU admissions, 80% (n = 2,161,763; (mean SAPS II = 41 ± 21.3) were used to fit logistic regressions of the in-hospital mortality probability as polynomial functions of the SAPS II up to the 5^th degree. The remaining 20% (n = 540,441; (mean SAPS II = 41 ± 21.3)) of the analytical sample was used as the validation set on which we computed overall and age-stratified Brier scores for model selection and calibration comparison against the original SAPS II mortality prediction equation. Discrimination was checked using the receiver operating characteristic (ROC) curve.

The observed mortality rate in the validation set was 20.0%, while the original SAPS II mortality mapping predicted 32.0%. By minimizing the Brier score, we identify a quintic polynomial model that predicts an overall 20.0% mortality rate. This new calibration relatively improves by more than 19% the overall Brier score, and by more than 29% for intermediate SAPS II values (60-79). The discrimination remains similar (area under the ROC curve = 0.830 and 0.832 for the original SAPS II and quintic model, respectively).

Even if SAPS II maintains adequate discrimination, its systematic overprediction undermines its utility for contemporary critical care practice, suggesting a regular refitting of its ability to predict in-hospital mortality.