Aim: Approximately six million people suffer cardiac arrests worldwide per year with very low survival rates (1-10%). Thus, the aim of this study is to estimate the probability of a poor outcome 72 hours after cardiac arrest. Accurate outcome predictions avoid removing care too soon for patients with potentially good outcomes or continuing care for patients with likely poor outcomes.
Materials: The study data consists of the cleanest five-minute samples per hour from 18 EEG channels. This data includes up to 72 hours after cardiac arrest for 607 patients.
Methods: The method is based on dynamical systems embedding theorems that show that a reconstructed phase space (RPS) topologically equivalent to an underlying system can be constructed from measured signals. Here the underlying system is the human brain after a cardiac arrest, and the signals are the EEG channels. We characterize the RPS with a Gaussian mixture model (GMM). The RPS parameters, time lag (12) and dimension (4), are estimated using the automutual information function and false nearest neighbor method, respectively. A 16 mixture GMM is estimated for each outcome, for each hour, and for two EEG channels. An unseen patient's outcome is predicted by summing the log probabilities for each outcome across all hours and channels. The outcome with the largest sum is selected. The method is robust against missing data as the log probabilities that correspond to missing channels and hours are not included in the sum.
Results: Our metric is the true positive rate when the false positive rate is 0.05. We achieved a score of 0.38 based on a five-fold cross validation of the 607 patients. Additionally, we have an AUC of 0.77 and an accuracy of 0.70. No official score was received.
Conclusions: This study provides a promising approach to predicting patient outcomes after cardiac arrest.