A continuously updated predictive analytics model for the timely detection of critically ill patients with a high risk of mortality

Abstract

ABSTRACT

Intensivists routinely encounter seemingly stabilized patients who expire before discharge. Current early warning systems have not proven effective for identifying these patients in sufficient time for clinical intervention. A novel algorithm is described here that is timely and accurate for recognizing patients in an intensive care unit (ICU) that have a substantial mortality risk.

Population: 59,400 admissions at 13 adult ICUs from 1/1/2012-9/30/2018.

Outcome: Mortality before discharge from the ICU. Overall rate was 6.9%.

Methods: All heart rate, respiratory rate and oxygen saturation values were obtained, as well as the start and stop times for those patients receiving mechanical ventilation. Data from the first two hours post-admission were used to find the cut points that maximized variability rates across ranges of vital signs values. These ranges were subsequenty mapped to a letter. A letter was then assigned to the median of each vital sign over consecutive 30-minute periods. Four consecutive letters were concatenated to form a pattern, and these were candidates for triggers (i.e. risk alerts). Using a genetic algorithm that weighted the outcome of mortality, we acquired a set of patterns that increased risk. Those patterns were then validated as triggers for increased risk.

Results: Patients with zero or one triggers had a mortality rate of 0%; patients with two to four triggers had a mortality rate of between 2.8% and 5.5%; five or more triggers were seen in patients with a 20.2% to 25.6% mortality rate.

Conclusion: Distinctive patterns in vital signs and whether a patient received mechanical ventilation can identify patients that have a high risk of mortality.  This methodology could be prospectively used in ICUs to identify high-risk patients in a timely enough manner to effect remedial treatment.