Team Polar Bear

Mechanical ventilation is a critical and often life-saving treatment required by patients with respiratory failure. However, during mechanical ventilation, a patient is at risk for developing ventilator associated complications (VAC), requiring further escalation of respiratory support. In fact, 5-10% of patients on a ventilator will develop a complication. This complication can significantly increase a patient’s mortality risk and increases the risk of other comorbidity complications. Predicting respiratory decompensation early would allow early intervention and potentially improve patient outcomes, but currently there is no reliable method of predicting respiratory decompensation. Our group is developing a predictive model using physiological time series and electronic health record data, utilizing the Physionet MIMIC-III database. The cohort consists of n=7734 patients receiving extended mechanical ventilation, of which n=1733 experience a VAC and n=6001 do not. We will build a generalized linear model to predict a time-varying risk score of a patient’s risk of declining respiratory status. Our model with EHR data alone resulted in an AUC of 0.69 and accuracy of 0.79. The results highlight the value of identifying potential predictive features and the computational need to explore further model types. We anticipate that this clinically significant early warning of decompensation could trigger interventions to aid in clinical decision-making for improved patient outcome.