External validation of a prognostic model predicting time of death after withdrawal of life support in neurocritical patients.

OBJECTIVE The ability to predict the time of death after withdrawal of life support is of specific interest for organ donation after cardiac death. We aimed to externally validate a previously developed model to predict the probability of death within the time constraint of 60 mins after withdrawal of life-sustaining measures. DESIGN The probability to die within 60 mins for each patient in this validation sample was calculated based on the model developed by Yee et al, which includes four variables (absent corneal reflex, absent cough reflex, extensor or absent motor response, and an oxygenation index >4.2). Analyses included logistic regression modeling with bootstrapping to adjust for overoptimism. Performance was assessed by calibration (agreement between observed and predicted outcomes) and discrimination (distinction of those patients who die within 60 mins from those who do not, expressed by the area under the receiver operating characteristic curve). SETTING Mixed intensive care unit in The Netherlands. PATIENTS We analyzed data from 152 patients who died as a result of a neurologic condition between 2007 and 2009. INTERVENTIONS None. MEASUREMENTS AND MAIN RESULTS A total of 82 patients had sufficient data. Fifty (61%) died within 60 mins. Univariable and multivariable odds ratios of the predictors were very similar between the development and validation sample. The prediction model showed good discrimination with an area under the receiver operating characteristic curve of 0.75 (95% confidence interval [CI] 0.63-0.87) but calibration was modest. The mean predicted probability was 80%, overestimating the 61% overall observed risk of death within 60 mins. Modeling oxygenation index as a linear term led to an improved version of the Mayo NICU model. (area under the receiver operating characteristic curve [95% CI] = 0.774 [0.69-0.90], bootstrap-validated area under the receiver operating characteristic curve [95% CI] = 0.74 [0.66-0.87]). CONCLUSIONS The model discriminated well between patients who died within 60 mins after withdrawal of life support and those who did not. Further prospective validation is needed.