In this work, we describe the creation of our machine-learning based approach for coma prognostication after cardiac arrest using EEG recordings for the Physionet/Computing in Cardiology Challenge 2023. Our team, "ComaToast”, had its best submission ranked as 79th out of 179 submissions with a challenge score of 0.33 on the Unofficial leaderboard and obtained a local 5-fold cross-validation score of 0.34 on the provided training data.
For our current highest-scored submission, we use a combination of age and recording features from the frequency domain, specifically power spectral density from 4 different bands of frequencies (Delta, Theta, Alpha and Beta) from the last available recording for a given patient. After imputing missing values, these features are fed to an XGBoost classifier for final binary classification. Even though various deep learning architectures like residual networks and LSTMs were tried on the raw signal and features extracted from it like DWT, MFCC, and PSD coefficients, we observed that all these methods overfit the training data easily and also have a high false positive rate. We believe that Deep Learning approaches still hold great potential but are not able to perform well due to the relatively large size of the input data (72 hours) when compared to the number of patients (607).
In the future, we plan to try techniques like Supervised Contrastive Learning to effectively extract shorter embeddings from each recording, making learning easier for Machine/Deep Learning based models. We also want to try out other recurrent and attention based architectures which can effectively capture sequential information across recordings to help make appropriate predictions. As the challenge metric emphasises a low false-positive rate, techniques like class weights, under and oversampling can be tried for the different architectures.