A time-series approach to predict outcome from pediatric intensive care

Daily assessment of the physiologic status of intensive care unit (ICU) patients by a validated score is used to predict their discharge as either alive or dead. Daily scores were decorrelated by time-series analysis techniques to establish a predictor of the next day's score. Data from 3299 patients (290 deaths) in nine different pediatric ICUs were used to derive the predictor model. Using the highest predicted score observed in any survivor from this data base as a threshold value, this predictor correctly identified 21.9% (sensitivity) of the nonsurvivors with no errors in predicting a fatal outcome (100% specificity). This performance is significantly (P < 0.001) better than that with a static predictor based on the first- or second-day scores only (6.6 or 7.4% sensitivity, respectively), or a known empirical dynamic model using the scores of the 2 most recent days (10.3% sensitivity, P < 0.002).