Mortality prediction systems have been calculated and validated from large mixed ICU populations. However, in daily practice it is often more important to know how a model performs in a patient subgroup at a specific ICU. Thus, we assessed the performance of three mortality prediction models in four well-defined patient groups in one centre.

A total of 960 consecutive adult patients with either severe head injury (n = 299), multiple injuries (n = 208), abdominal aortic aneurysm (n = 267) or spontaneous subarachnoid haemorrhage (n = 186) were included. Calibration, discrimination and standardized mortality ratios were determined for Simplified Acute Physiology Score II, Mortality Probability Model II (at 0 and 24 h) and Injury Severity Score. Effective mortality was assessed at hospital discharge and after 1 yr.

Eight hundred and fifty-five (89%) patients survived until hospital discharge. Over all four patient groups, Mortality Probability Model II (24 h) had the best predictive accuracy (standardized mortality ratio 0.62) and discrimination (area under the receiver operating characteristic curve 0.9), but Simplified Acute Physiology Score II performed well for patients with subarachnoid haemorrhage. Overall calibration was poor for all models (Hosmer–Lemeshow Type C-values between 20 and 26). Injury Severity Score had the worst discrimination in trauma patients. All models over-estimated hospital mortality in all four patient groups, and these estimates were more like the mortality after 1 yr.

In our surgical ICU, Mortality Probability Model II (24 h) performed slightly better than Simplified Acute Physiology Score II in terms of overall mortality prediction and discrimination; Injury Severity Score was the worst model for mortality prediction in trauma patients.