Autoencoder Artefact Removal for Brain Signals and Impact on Classification Performance

As part of the George B. Moody Physio Net Challenge 2023, we developed a computational approach that uses 6 channels of electroencephalograms (EEGs) to predict neurological recovery outcomes of patients following cardiac arrest. Our team, UoM EEE, developed a2-Dimensional Convolutional Neural Network, using the Short-Time Fourier Transform to obtain an image representation of the EEG. It uses an optimised Binary Focal Cross-entropy loss function for balancing weights oft wo-outcome classes. As standard EEG analysis pipelines using Independent Component Analysis (ICA) to remove artefacts are not suitable due to the limited channel count, we hypothesized that an autoencoder machine learning approach may allow a channel count independent artefact removal, and potentially an improved true positive rate, while naturally complementing machine learning based classification used for the main Challenge problem. A5-run class-stratified nested holdout was performed, with Area under the Receiver Operating Characteristic Curve, AUC, as metric for model selection. Our model received a Challenge score of 0.39 (ranked 39 out of 73 teams) on the hidden validation set, and 0.67 averaged across 5-trialcross-validation on the public training data