Neurological Outcome Prediction after Cardiac Arrest: A Multi-Level Deep Learning Approach with Feature and Decision Fusion

Bill Chen, Jerry Yang, Will Wang, Perisa Ashar, Leeor Hershkovich, Hayoung Jeong, MD Mobashir Hasan Shandhi, Jessilyn Dunn
Duke University


Cardiac arrest is a prevalent and life-threatening condition that can result in severe brain injury. Accurate prognosis is critical in determining the appropriate treatment for post-cardiac arrest patients. However, the qualitative interpretation of continuous electroencephalogram (EEG) is a laborious and expensive process, leading to discrepancies in prognosis accuracy. In this study, we aim to address this challenge by proposing a set of machine learning models that use longitudinal EEG recordings to predict good or poor patient outcomes (i.e., Cerebral Performance Category scores) after cardiac arrest.

To achieve this, TheBIGbrain team utilized the International Cardiac Arrest Research Consortium (I-CARE) dataset, which contains over 50,000 hours of EEG recordings. We proposed an approach that combines automatically extracted features from the Random Convolutional Kernel Transform (ROCKET) with wavelet, spectral, and gradient features as inputs into our classifier. Our findings demonstrate that incorporating features extracted by the ROCKET algorithm, in combination with clinical and time-series features, results in better performance compared to using only manually extracted features in terms of accuracy scores, F1-scores, as well as True Positive Rates with maximum 0.05 False Positive Rates. Using 5-fold cross-validation on the training dataset, we were able to achieve 0.107 ± 0.004 TPR, 0.651 ± 0.002 Accuracy and 0.708 ± 0.002 F1 scores. Additionally, the ROCKET classifier exhibits significantly lower computational complexity in comparison to other deep learning-based methods, rendering it a promising choice for automating quantitative analysis of EEG signals.

We also aim to train and validate the following Time Series Classification methods in comparison to ROCKET: 1) Convolutional Neural Networks via transfer learning from larger EEG datasets, 2) Transformer to learn the best representation of each 5 minute time series EEG signals and then LSTM for predictive modeling, and 3) Mini-ROCKET and feature engineering accelerated using GPUs.