This study showcases the contributions of our team, The BEEGees, to the George B. Moody PhysioNet Challenge 2023, which aims to predict neurological recovery from coma following cardiac arrest using 18-channel EEG signals and supplementary metadata.
We employed a two-step modelling approach for this challenge. Initially, we applied various transformations to the EEG signals, enabling us to utilise pre-trained deep learning models to extract hourly outcome probabilities from the EEG recordings. Subsequently, these predictions were combined with hand-selected features and input into a final classification model. During development, our models were evaluated using a 10\% held-out subset of the 607 patients provided in the training data.
Our final model achieved the highest challenge score of $0.73$ on the hidden validation set of the unofficial challenge phase for predictions made 72 hours after return of spontaneous circulation. Our most successful model incorporated a blend of hand-selected features and features extracted from EEG signals using a mix of data transformations and deep learning techniques. We employed a Random Forest as our final classification model. On our held-out subset, we achieved a challenge score of $0.829$ and an AUC of $0.957$.
In conclusion, our findings demonstrate the effectiveness of transfer learning for medical image classification in the context of EEG signal analysis.