5228Identification of reliable clinical variables for early prediction of outcome after out-of-hospital cardiac arrest using algorithm-based machine learning statistical methods

Background Several factors are known to influence both survival and long-term neurologic function after out-of-hospital cardiac arrest (OHCA). Previous studies have indicated that both pre-hospital circumstances as well as patients' history and clinical status on hospital admission are variables strongly associated with later outcome. This study aimed to identify and evaluate clinical variables for early prediction of outcome for unconscious survivors after OHCA using machine learning statistics analysis. Methods The Target Temperature Management (TTM) trial randomized 939 international patients with OHCA of presumed cardiac cause to TTM at 33°C or 36 °C for 24 h in intensive care units (ICUs). Patient outcome were survival and neurological function defined by the Cerebral Performance Category (CPC) scale. This multicentre cohort was used for a post hoc analysis using machine learning statistical analysis. A Conditional Interference decision forest algorithm was designed for training on the TTM-dataset to perform early prediction of outcome at 180 days. Results After ranking all available variables in the TTM-dataset based on their importance for the algorithm to make predictions, we could identify a slimmed list with eleven clinical predictors of a poor outcome including older age, low motor score on Glasgow Coma Scale (GCS), increasing doses of adrenaline, first monitored rhythm not shockable, longer duration of low flow, longer time from cardiac arrest to advanced life support, high BMI (Body Mass Index), low pH, bilateral absence of corneal and pupillary reflex, low initial body temperature and cardiac arrest location at home. Age was overall the most important variable for prediction. Our slimmed prediction model performed slightly worse with an AUC of 0.813 (0.741–0.916) compared to an extended model with all available variables included, AUC = 0.839 (0.778 – 0.886). When using all variables in a comparing logistic regression analysis the mean AUC was a corresponding 0.830 (0.792–0.882). Conclusion This algorithm with eleven clinical variables predicted outcome almost as good as a corresponding large model with cardiac arrest patients from the TTM-trial and could be a powerful clinical decision tool for early prediction of outcome after cardiac arrest.