This abstract is part of the George B. Moody 2023 Physionet contest. The aim of the contest is to predict clinical outcomes in cardiac arrest patients using repeated EEG recordings. Our team's challenge score during the unofficial portion of the contest was 0.373. Our cross-validated challenge score is 0.361.
EEG data contains important structure at multiple resolutions. Traditional tools for analyzing multi-resolution structure, such as wavelets and Fourier transforms, show promise in levering this complex structure to enable prediction of clinical outcomes on the basis of EEG recordings. More recently, similar ideas have been incorporated into neural network architectures such as the WaveNet model and the Discrete Wavelet Transform Recurrent Neural Network. These architectures capitalize on the ability of neural networks to learn distinct filters relevant to the outcomes of interest, and in some cases have been shown to improve on traditional signal-processing methods.
In this work, we will use a model architecture based on the WaveNet model to predict patient outcomes from EEG data. Challenge data includes 5 minutes of EEG recordings per hour. It will be important to take full advantage of each of these recordings, and the sequence in which they occur. A modified WaveNet model will be applied to each model. In addition to being trained against the outcome of interest, the model will also contain an auto-encoder module. The auto-encoder will be trained against the traditional objective of signal recovery. The encoding of EEG recording will be passed along as input to the modified WaveNet model applied to the next EEG recording. In this way, the encoder piece of the auto-encoder will be trained against the outcome objective as well. The prediction rendered by the model will be a learned weighted average of the predictions on each signal.