Skip to main content
. Author manuscript; available in PMC: 2022 Dec 1.
Published in final edited form as: Resuscitation. 2021 Oct 24;169:86–94. doi: 10.1016/j.resuscitation.2021.10.034

Fig. 1. The proposed model architecture.

Fig. 1

The framework contains three components: CNNs, multiscale LSTMs and the demographics model. CNNs were used to extract features for every 10-s EEG segments. The outputs of CNN features were averaged as the inputs of Bi-LSTMs. The fine-grained Bi-LSTMs made predictions based on the most recent 6-h EEG features while the coarse-grained Bi-LSTMs aimed at the snapshots of EEG evolutions from the beginning to the current time. The modeling framework for EEG has two important novelty: EEG feature learning in a data-driven way and short-term and long-term time dynamic modeling. The demographics model made predictions based on the clinical variables age, sex and ventricular fibrillation. The outputs of all models were fused by averaging to obtain the final prediction probabilities of neurological outcomes.