Factors predictive of mortality and mortality prediction models in patients with ARF during the first 24 hours of intensive care unit admission

We examined outcomes for patients with ARF identified from a national comparative audit of adult general ICUs in England, Wales and Northern Ireland, the Case Mix Programme.

Over an 8-year period ARF was present during the first 24 hours of admission in 17,326 of 276,731 admissions to the ICU. In 14,118 (81.5%) sufficient data were present to predict mortality using the Stuivenberg Hospital ARF (SHARF), UK APACHE II, and Mehta mortality prediction models. Discrimination was assessed by the area under the receiver operating characteristic curve (ROC), calibration by the mortality ratio (observed versus predicted mortality), and overall fit by the R-statistic from Shapiro's Q representing the geometric mean of the probability assigned to the true outcome.

Factors predicting > 50% increased mortality risk were: history of chronic condition, CPR, IPPV, oliguria, prior hospital stay of 7+ days, MAP < 50 mmHg, acidosis, and reduction of GCS by 2. UK APACHE II scores showed the best discrimination and calibration, although the null hypothesis of perfect calibration was strongly rejected (P < 0.001) by both the Hosmer–Lemeshow test and Cox's calibration regression. UK APACHE II and Mehta underpredicted the number of deaths while SHARF T0 overpredicted (Table 1). SHARF T0 and Mehta's model showed poor overall fit by Shapiro's Q, with an R value < 0.5.

Table 1 Table 1

Although UK APACHE II scores performed best, none of the existing mortality prediction models reliably predict those who will survive to leave hospital following an episode of ARF in the ICU. Traditional risk factors such as oliguria, CPR, LOS > 7 days prior to admission to ICU and extremes of physiology continue to be associated with the greatest increased risks of mortality.