Objective: Patients admitted to the intensive care unit greatly differ in severity and intensity of care. We devised a system for selecting high-risk patients that reduces bias by excluding low-risk patients and patients with an early death irrespective of the treatment.
Design: A posteriori analysis of a multiple-center prospective observational trial.
Setting: A total of 89 units from 12 European countries, with 12,615 patients.
Intervention: Demographic and clinical data: severity of illness at admission, daily score of nursing workload, length of stay, and hospital mortality.
Methods: We enrolled patients with intensive care unit length of stay of >24 hrs. Three groups of high-risk patients were created: a) Severity group, those with Simplified Acute Physiology Score (SAPS II) over the median; b) Intensity-of-care group, patients with >1 day of high level of care (assessed by logistic analysis); and c) MIX group, patients fulfilling both Severity and Intensity-of-care criteria. The groups were included in a logistic regression model (random split-sample design) to identify the characteristics associated with hospital mortality. We compared the outcome prediction of the SAPS II model (unsplit sample) against our model.
Main results: Out of 8,248 patients, the Severity method selected 3,838 patients, Intensity-of-care selected 4,244, and both methods combined selected 2,662 patients. There were 2,828 low-risk patients. Significant associations with hospital mortality were observed for: age, sites of admission, medical/unscheduled surgical admission, acute physiologic score of SAPS II, and the indicator variable "only Severity," "only Intensity-of-care," or MIX (developmental sample: calibration chi-square test, p = .205; area under the receiver operation characteristic curve, 0.814). Calibration and discrimination were better in our model than with the SAPS II model (unsplit sample).
Conclusion: All three indicator variables select high-risk patients, the Severity/Intensity-of-care MIX being the most robust. These stratification criteria can improve case-mix selection for clinical and organizational studies.