Introduction: The George B. Moody PhysioNet Challenge 2023 aims to improve the automated analysis of continuous EEG data to predict neurological outcomes in comatose cardiac arrest survivors. Accurately predicting patient outcomes can help in designing targeted and effective treatment plans.
Methodology: In this study, we propose a model that combines a Convolutional Neural Network (CNN) based on a customized VGG architecture and a Random Forest (RF) method in a pipeline to predict patient outcomes. The model takes into account patient demographics and clinical data and EEG signals from 18 bipolar channel pairs, distributed in 5-minute epochs. The epochs are further partitioned into non-overlapping segments of 30 seconds, which serve as inputs to our CNN to classify them into five Cerebral Performance Categories (CPC) as stacked epoch logits. These epoch logits, patient demographics, and clinical data are the inputs of the RF method, which performs the predictions of the CPC and outcome for each epoch. The final patient CPC prediction is determined by the time-weighted average of each EEG epoch prediction.
Results: Our submission in the hidden test set achieved challenge metric scores of 0.18, 0.67, 0.63, and 0.70 for 12h, 24h, 48h, and 72h, respectively, placing the AIMED team in 4th place at the unofficial phase. These results indicate the potential of our proposed model to predict neurological outcomes in comatose cardiac arrest survivors.
Conclusion: Our study demonstrates the potential of a machine learning pipeline to accurately predict neurological outcomes in comatose cardiac arrest survivors using continuous EEG data. The proposed model could help address the challenge of analyzing and interpreting EEG signals in a cost-effective and efficient manner, potentially improving the quality of care for these patients.